Most of the time, the system catches the problem. Eventually. But detection has always depended on a convergence: the right capability, in the right place, at the right time. When any one of those is missing, problems that are entirely detectable go undetected.

The lab can identify what’s in the water, but it’s not at the water source. The reading specialist can catch a struggling student, but she can only sit with one child at a time. The cardiac imaging can track heart failure progression, but it requires a hospital equipped to run it. The capability exists…somewhere.

What connects this issue’s four stories is that AI didn’t create a new way to detect these problems. It moved an existing capability to where the problem actually is, when it’s actually happening. A portable device at the water source instead of a sample shipped to the lab. An AI tutor listening to every student in the room simultaneously. A single-lead ECG that works in a rural clinic instead of a cardiac imaging suite. In each case, the detection gap closed because the tool finally showed up where it was needed.

From Pilot to Infrastructure: How the CDC’s AI Tools Are Spreading Beyond the Lab

Legionnaires’ disease is a severe, sometimes fatal form of pneumonia caused by bacteria that thrive in warm water systems, particularly the cooling towers found on commercial and industrial buildings. Every dense urban area has hundreds of them. When an outbreak starts, investigators need to locate every tower in the affected area so they can test for contamination. The traditional method: manually scanning aerial and satellite imagery, building by building, block by block. That takes roughly four hours per area. During an active outbreak, four hours is a long time.

The CDC’s AI Accelerator Program built TowerScout, a computer vision tool that scans satellite imagery and flags cooling towers automatically. The review drops from four hours to five minutes. What happened next is more telling: Utah’s Geospatial Resource Center adopted the tool independently, scanning aerial imagery for the entire state. The effort processed 5,400 hours of computing time in nine days and identified 6,324 potential cooling towers. Utah now has a prebuilt registry investigators can reference the next time an outbreak develops, rather than starting the search from scratch.

TowerScout is one of several AI tools the CDC has deployed through the accelerator. Clinical Narratives uses natural language processing to extract diagnoses, medications, and symptoms from the free text buried in electronic health records, where critical details often sit unread in physician notes. The CDC reports the tool has improved the speed and accuracy of that extraction by 80% compared to baseline. NewsScape uses large language models to scan roughly 8,000 news articles and open-source reports daily for signals of emerging health threats. The agency is now expanding into “Deep Research” agents that compress three- to four-hour literature and policy reviews into minutes, with a human expert validating the final output.

In March 2026, the CDC formalized this trajectory with a five-year AI strategy built around moving these tools from isolated pilots into standard agency workflows, including integration with electronic health records and real-time surveillance systems. The strategy also addresses a problem familiar to any large organization adopting AI: the rise of unauthorized “shadow AI” use by staff, which the CDC plans to manage through controlled environments where researchers can work with approved datasets and models.

The efficiency numbers are striking, but they come with a caveat. The 98% time reduction for TowerScout and the 80% improvement for Clinical Narratives both originate from unpublished internal CDC studies. No independent validation has been reported for any of the tools. Utah’s statewide adoption suggests the tools work well enough for other agencies to invest in them. But the specific performance claims are still the agency grading its own homework.

Article
Updated CDC Strategy
Utah TowerScout Implementation

Key Stakeholders

CDC
UC Berkeley
Utah Geospatial Resource Center

From Days in the Lab to Minutes in the Field: A Portable AI for Water Safety Testing

Finding cooling towers is the first step in preventing a Legionnaires’ outbreak. The next is testing whether the water in those systems actually harbors dangerous microorganisms. Right now, that means collecting a sample and shipping it to a specialized lab. Results come back days or weeks later. For public health teams responding to a suspected outbreak, that turnaround means waiting for confirmation while the exposure window stays open.

The bottleneck isn’t the science. It’s where the science can happen. The equipment needed to identify and count microorganisms in a water sample costs tens of thousands of dollars, requires trained laboratory staff, and doesn’t leave the lab. Emerging technology from Georgia Tech could change that.

Researchers in Georgia Tech’s Pinto Lab built ARTiMiS, a portable imaging device that captures microscopy-level images of microorganisms in a water sample on site. PhycoSight, the companion AI software, identifies and counts those microorganisms from the images within minutes. Together, they’re designed to give field operators the kind of information that previously required a lab, fast enough to inform real-time decisions. A water utility operator could assess whether harmful microorganisms are present in a drinking water source, or engineers could monitor the biological processes in a wastewater treatment system, without waiting days for lab results.

The technology has been commercialized through a startup called Skopii, launched by two Georgia Tech doctoral students who won the Department of Energy’s national Algae Prize competition in 2023. Development was funded by the National Science Foundation, the Department of Energy, and the Water Research Foundation. The underlying research has been published in Environmental Science & Technology, and the ARTiMiS platform is open-access.

Skopii is early-stage. There are no confirmed customers and no published deployment data. The company is in early discussions with bioprocessing and algae reactor manufacturers about integrating the technology into commercial equipment, but no active commercial relationships exist yet. No independent accuracy benchmarks for PhycoSight have been reported: the speed advantage over lab testing is clear, but whether the AI’s identifications match lab-grade results under real-world conditions hasn’t been publicly validated.

What makes the story worth watching isn’t the startup. It’s the problem. Water utilities make safety decisions based on data that’s days old. Environmental managers tracking contamination are working from the last lab report, not from what’s happening right now. If portable AI-driven monitoring works as designed, it changes the timing of those decisions in the same way TowerScout changed the timing of cooling tower identification. Whether Skopii is the company that delivers on that promise remains to be seen. But the need is real, and the gap between field conditions and lab infrastructure isn’t getting smaller.

Article

Published Research

Key Stakeholders

Georgia Tech
Skopii
National Science Foundation
U.S. Department of Energy
Water Research Foundation

Reading Out Loud to an AI That Listens Back: Ferndale’s Answer to Fewer Reading Specialists

Teaching a child to read is one of the most labor-intensive things a school does. A reading specialist sits with a student, listens to them read, catches the mistakes, and builds a lesson around what they got wrong. It works. The problem is scale: one specialist can only sit with so many students in a day. When budget cuts hit, and they’ve been hitting school districts across Washington, that number shrinks.

Ferndale School District lost reading intervention staff as part of broader cuts it attributes to the state’s underfunding of public education. The gap left behind is the kind that shows up quietly: students who need targeted help simply getting less of it. The district turned to Amira AI, a reading tool developed using research from Carnegie Mellon University, to fill some of that space.

Amira works by listening. A student reads aloud to a cartoon character on screen, and the AI evaluates their fluency, accuracy, and vocabulary in real time. When a student misreads a word, a mini-lesson pops up immediately: the student might sound out the word several times, or watch a short video showing a mouth forming the correct pronunciation. A reading specialist at Central Elementary School said she reviews the results afterward to identify where she can support students through direct instruction. The AI handles the repetitive practice. The specialist focuses on what needs a human.

A classroom using Amira sounds like 15 lessons happening at once. One student reads with fluency while another across the room works on decoding words that start with “qu.” Each student gets a lesson calibrated to where they are, from individual letter sounds up through reading comprehension for older students.

Amira Learning holds contracts with 10 states and is approved for use in schools across most others, including Washington. Ferndale’s use is still limited: students identified as potentially benefiting from the tool get about 30 minutes a week, and the district paid under $5,000 for a limited number of seats this year. The district plans to evaluate effectiveness at the end of the school year before deciding whether to continue.

The tool isn’t perfect. Teachers have had to override errors that Amira flagged incorrectly. The AI struggles with multilingual learners and students with speech impediments, sometimes misinterpreting pronunciation differences as reading errors. One staff member whose son has a rare chromosome difference that affects his articulation said the program has gradually adjusted to his way of speaking, and that his reading and language have improved since he started using it regularly. She also noted that the tool handles corrections in a way that feels encouraging rather than punitive, which matters for a child who already knows reading is hard.

It’s early. The results are anecdotal, and the formal evaluation hasn’t happened yet. But the underlying logic is straightforward: when you can’t hire enough specialists, a tool that provides individualized practice at scale gives students something they’d otherwise go without.

Article

Key Stakeholders

Ferndale School District
Amira Learning
Carnegie Mellon University

A Year’s Warning from a Simple Heart Test: MIT’s AI Forecasts Heart Failure Decline

Heart failure doesn’t announce its next move. A patient can be stable for years, then deteriorate rapidly. The challenge for clinicians isn’t diagnosing the condition. It’s knowing which patients are about to get worse. Right now, there’s no clinical tool that forecasts that trajectory. Doctors monitor, adjust medications, and wait to see what happens.

Researchers at MIT, Mass General Brigham, and Harvard Medical School built PULSE-HF, a deep learning model that analyzes electrocardiogram data to predict whether a heart failure patient’s pumping ability will decline to dangerous levels within the next twelve months. A healthy heart pumps out 50 to 70 percent of the blood in its left ventricle with each beat. When that number drops below 40 percent, the patient has the most severe form of heart failure, with significantly higher risks of hospitalization and death. About half of all heart failure patients die within five years of diagnosis.

PULSE-HF was tested retrospectively across three patient populations: Massachusetts General Hospital, Brigham and Women’s Hospital, and a publicly available dataset called MIMIC-IV. The model’s accuracy scores ranged from 0.87 to 0.91 on a scale where 0.5 is random guessing and 1.0 is perfect. When the model predicted a patient would remain stable, it was right more than 97 percent of the time. That’s clinically useful: it could allow doctors to reduce the frequency of hospital visits and invasive monitoring for lower-risk patients while focusing resources on those the model flags.

The researchers also built a version of PULSE-HF that works with a single-lead ECG, requiring just one electrode on the body instead of the standard twelve. The single-lead version matched the performance of the full version. That means the tool could function in rural clinics, small practices, and settings that don’t have cardiac sonographers on staff to run full imaging workups. The researchers designed it that way on purpose: if the tool only worked with expensive equipment in major hospitals, it would help the patients who already have the most access.

The study was published in Lancet eClinicalMedicine, which gives it a level of peer-reviewed validation that most AI health stories lack at this stage. But it’s still retrospective: the model was tested on historical data where the outcomes were already known. The prospective study, testing on real patients whose future is unknown, hasn’t begun yet. Clinical deployment is likely years away. The data challenges were significant: the team spent years cleaning messy ECG records, dealing with signal artifacts from restless patients, and converting PDFs to usable formats. Real-world data will be messier than training data.

None of that diminishes the core contribution. No comparable tool exists for forecasting heart failure trajectory from an ECG. If PULSE-HF holds up in prospective testing, it changes how clinicians allocate attention: not just who has heart failure, but who is about to get worse.

Article
Published Study

Key Stakeholders

MIT
Mass General Brigham
Harvard Medical School

Takeaway

Every story in this issue has the same structural problem: a capability that exists but can’t reach the problem it’s designed to solve. The lab can test the water, but not where the water is. The specialist can catch reading errors, but not for every student at once. The cardiac imaging can track heart failure, but not in a clinic without a sonographer.

When AI moves those capabilities closer to the problem, detection arrives earlier. And earlier detection changes what’s possible on the other end. A microorganism identified in the field means a utility manager can act before exposure spreads. A reading error caught in real time becomes a 30-second mini-lesson, not a summer remediation program. A heart failure trajectory flagged a year out gives a clinician time to adjust treatment, not just react to a crisis.

The pattern worth paying attention to isn’t the AI. It’s the relationship between when you catch something and what you can do about it. Earlier detection doesn’t just save time. It changes the entire range of available responses.

Share the Knowledge

If you know someone working in public health, education, or healthcare who’s dealing with the gap between what they can detect and when they can detect it, forward this to them.

Between issues, I track AI tools across all of these sectors. If a story in this issue is relevant to your work and you want to dig deeper, reply and I’ll point you to what I’ve found.

Subscribe now

The people responsible for catching problems in healthcare, education, and infrastructure are almost always working on a rotation. They check in, respond to what they find, and move on to the next person or site. That’s not a flaw in the system. It’s how human attention works when one person is responsible for many. But the problems that develop between those visits don’t follow the same schedule. In ecology, the edge effect describes what happens at the boundary between two environments: that’s where the most activity and vulnerability concentrate. Institutions have their own version. The boundary, or period, between one check-in and the next is where patients fall, students get stuck, equipment faults, and wounds change. This issue looks at four stories where AI shows up in those gaps.

62% Fewer Falls in 90 Days: A Phoenix Senior Community Rewires Its Safety Net

Falls in senior living communities follow a grim math. The CDC estimates that 50 to 75% of long-term care residents fall each year, generating more than $50 billion annually in U.S. medical costs. The window between a resident’s first sign of trouble and the fall itself is often too short for anyone on a scheduled rotation to catch.

Bethesda Gardens, a faith-based nonprofit senior living community in Phoenix, deployed a radar-based AI monitoring system across all of its assisted living and memory care units, covering up to 147 residents. The device, built by Helpany, tracks movement patterns continuously without cameras, preserving privacy while flagging behavioral changes that precede falls or health deterioration.

Within 90 days, falls dropped 62% in assisted living and as much as 78% in memory care, compared to the previous 12-month averages. Injury rates fell 46% below industry benchmarks. Hospitalizations dropped 53% below national averages.

Bethesda Gardens isn’t the only site reporting these results. Two other Arizona senior living nonprofits using the same system, Fellowship Square Historic Mesa and Westminster Village, reported fall reductions of 80% (measured by fall-related 911 calls) and 72% respectively over similar timeframes. All three organizations operate independently of one another.

The system’s value went beyond fall prevention. Because the radar monitors movement patterns continuously, it can detect behavioral shifts that signal emerging health issues. A resident getting up far more frequently at night may be showing early signs of a urinary tract infection. Changed mobility patterns after a medication adjustment may indicate side effects. Over 90 days, the system flagged 486 such condition changes, an average of 5.4 per day, giving caregivers a chance to intervene before symptoms escalated. Roughly 23% of residents showed gaps between their documented care plans and their actual observed needs, prompting the team to realign how care was being delivered.

A caveat: all three sets of results come from the communities and the vendor, not independent research. The observation windows are 90 days, the sample sizes are small, and no control groups are described. The results are promising, and the consistency across three independent sites is harder to dismiss than a single case. But they haven’t been externally validated. What makes the story worth watching is the scope of what the system surfaced: not just fall prevention, but a continuous picture of resident health that staff couldn’t realistically assemble through periodic check-ins alone.

Article
Fellowship Square Historic Mesa Case Study
Westminster Village Results

Key Stakeholders

Bethesda Gardens Phoenix
Helpany

“Basically a Teacher Online”: Allentown’s AI Chatbot Gets Mixed Reviews from Sixth Graders

A teacher circulating among 30 students can only be in one place at a time. A kid gets stuck on a writing prompt, raises a hand, and waits. By the time the teacher arrives, the moment has passed or the student has moved on to something else. The student isn’t falling behind because the teacher isn’t good enough. The structure of a classroom with one instructor and 30 learners means help arrives on a rotation, and learning doesn’t wait for the rotation.

Allentown School District in Pennsylvania deployed CourseMojo, an AI-powered chatbot, across its sixth-grade language arts classrooms. What sets it apart from generic AI tutors: CourseMojo was trained on the district’s actual literacy curriculum, StudySync. When students respond to reading comprehension questions or writing prompts, the AI matches their answer against criteria that human educators wrote, then delivers feedback from a bank of human-produced responses. It’s not generating answers on the fly. It’s sorting student work against a rubric and responding with guidance that teachers designed.

The classroom results are unusually candid for an ed-tech story. Some students found the tool genuinely helpful, saying it helped them get unstuck while writing or describing it as basically having a teacher online. But students also reported frustration: feedback sometimes felt vague or repetitive, and classmates who entered similar answers occasionally received different responses, which confused group work.

Teachers are presented a different picture, and it’s one that keeps them firmly in the loop. The system flags students who may need a one-on-one conference and identifies class-wide misconceptions in real time. During one lesson, CourseMojo signaled that many students were confusing “being a trailblazer” with “being skilled,” and the teacher paused the room to address it on the spot. Teachers can track individual responses through a dashboard, see how many times each student attempted a question, and surface strong answers worth sharing with the class. The AI handles the first pass on student work so teachers can focus their attention where it matters most: the students who need direct help and the patterns that require instructional adjustment.

The $45,000 implementation cost (licensing, coaching, and setup) is covered by Digital Promise, a nonprofit supporting education technology. CourseMojo currently serves nearly 60,000 students across 75 districts nationally, and Allentown is considering expansion to seventh and eighth grade.

No quantitative outcome data exists yet. No test score improvements, no literacy gains. This is explicitly year one, and the article is honest about that. But the operational detail is rich: a tool that’s embedded in the actual curriculum rather than layered on top of it, iterated in response to real student and teacher feedback, and reported on with enough candor to include the parts that aren’t working yet. That combination is worth following.

Article

Key Stakeholders

Allentown School District
CourseMojo
Digital Promise

Seeing Faults Before the Fire: Why AI Is Moving to the Grid’s Edge

Power grids have always been monitored from the center. Sensors and meters collect data across the distribution network, route it to a utility’s operations center, and analysts decide what to do. That works when the grid is relatively simple and threats develop slowly enough for the data to make the trip.

The grid is getting more complex. A traditional grid moves electricity in one direction: from a power plant to your house. But homes with rooftop solar sometimes consume power and sometimes push it back onto the grid, reversing the flow depending on weather, time of day, and cloud cover. EVs add large, unpredictable load spikes when people plug in. Battery storage systems charge and discharge on their own schedules. None of this increases total demand the way a new factory would. It increases the number of points where conditions change quickly and unpredictably, and centralized monitoring systems weren’t designed for that. By the time sensor data travels to a utility’s operations center, gets analyzed, and triggers a response, the window for prevention may already be closing.

A recent demonstration by Itron and Nvidia shows one approach to the problem. Instead of sending raw data back to a central server for analysis, their system runs AI directly on the devices already installed across the grid: the meters, sensors, and endpoints that sit at the point where conditions actually develop. The AI processes electrical waveform data on-site, detects anomalies that could indicate equipment faults or emerging wildfire conditions, and learns continuously, recognizing new threat patterns as it encounters them. The insight stays at the edge rather than making a round trip to headquarters.

This is a demonstration, not a deployment. No utility has announced an operational rollout. There are no metrics: no fault detection improvement rates, no latency comparisons, no before-and-after data from a real grid. The source is a trade publication covering what is effectively Itron’s industry announcement. All of that warrants caution.

But the concept is worth paying attention to, because the underlying problem is real and growing. Grid complexity is increasing faster than utilities can hire analysts to monitor it. The May 2024 solar storm, the strongest to hit Earth in over two decades, disrupted GPS systems, degraded high-frequency radio communications, and stressed power grids worldwide. Precision agriculture equipment lost the centimeter-level GPS accuracy it depends on, forcing farmers to halt planting.

No major blackouts were reported, but the storm was a reminder of how quickly external threats can develop and how dependent grid infrastructure is on systems that a more powerful storm could disable. Wildfire risk in the western U.S. and Australia has made grid monitoring a public safety issue, not just an engineering one. If AI can reliably detect the early signatures of a fault or fire risk at the source, seconds or minutes before a centralized system would catch it, the safety implications are significant.

This is a space worth watching. As deployments emerge, the questions that matter will be straightforward: how much faster does edge detection identify threats compared to centralized monitoring? How many false positives does it generate? And does it actually reduce outages, equipment damage, or fire risk in practice?

Article

Key Stakeholders

Itron
NVIDIA

The Photo Your Surgeon Hasn’t Seen Yet: Mayo Clinic’s AI Wound Triage

Outpatient surgery is increasingly common. More than 50 million procedures are performed in ambulatory settings each year in the U.S., accounting for roughly 70% of all surgeries. As virtual follow-ups expand, a growing number of care teams are asking patients to submit photos of their surgical wounds through online portals so healing can be monitored remotely. The system depends on someone reviewing those photos quickly enough to catch a problem before it becomes serious.

Right now, reviewing those photos is a manual process. Clinicians look at each image individually to assess whether a wound is healing normally or showing signs of infection. The researchers behind this study describe the process as time-consuming and note that it can delay care, particularly as outpatient procedures and virtual follow-ups become more common.

Researchers at Mayo Clinic built an AI system designed to triage those photos automatically. Published in the Annals of Surgery, the study describes a two-stage model that first identifies whether the image contains a surgical incision (94% accuracy), then evaluates whether that incision shows signs of infection, correctly distinguishing between infected and non-infected wounds 81% of the time. It was trained on more than 20,000 images from over 6,000 patients across nine Mayo Clinic hospitals, and the researchers report consistent performance across diverse patient groups, meaning it didn’t perform better for some demographics than others.

The tool is still in the research phase. The researchers are clear that prospective validation studies are underway and further testing is needed before clinical use. An 81% accuracy rate for infection detection is a solid starting point, but it’s not yet at the level needed for a system that patients would depend on to catch a worsening wound. A missed infection carries real consequences, and the accuracy bar for deployment in this context is high.

What makes the research worth covering now: the problem it targets is growing alongside the trend toward outpatient surgery and virtual follow-ups. If the model’s accuracy improves through ongoing validation, it could become a frontline screening tool that routes clinical attention to the wounds that need it most, particularly for patients in rural areas or resource-limited settings where an in-person follow-up isn’t easy to get.

Article
Ambulatory Surgery Centers Trend (ExactRx)
Acute Hospitalizations Following Outpatient General Surgery (NCQA)

Key Stakeholders

Mayo Clinic

Takeaway

The fundamental constraint in every story this week is the same one: a person responsible for many can only focus on one at a time. A nurse attending to a patient in Room 12 isn’t watching Room 14. A teacher conferencing with one student isn’t hearing the three others who just got stuck. That’s not a staffing problem or a performance problem. It’s how human attention works.

What these AI tools share is the ability to monitor broadly while the humans monitor deeply. The nurse can give her full attention to the patient in front of her because the radar system is tracking movement patterns in every other room simultaneously. The teacher can pull one student aside for a meaningful conversation because the chatbot is providing feedback to the rest of the class in real time. The utility engineer can focus on the site that needs repair because the AI is processing data from every other endpoint on the grid. And if something does develop, the system alerts them immediately, which means they can focus fully on the task at hand without worrying about what they might be missing somewhere else.

Neither piece works well alone. Continuous AI monitoring without human judgment produces alerts that nobody acts on. Human expertise without continuous coverage produces gaps where things get missed. The value is in the pairing: AI handles the breadth so the human can go deep where it counts.

The same pattern extends well beyond the four sectors in this issue. A child welfare caseworker visits families monthly, but school attendance records, emergency room visits, utility disconnections, and police calls to a household address are all generating data between those visits. A probation officer meets a client every two weeks, but GPS monitor data, check-in compliance patterns, and employment verification records are sitting in systems nobody has the capacity to review across 200 cases simultaneously. An environmental inspector audits a facility quarterly, but continuous discharge monitors and water quality sensors are already tracking conditions between visits. In each case, the data between check-ins already exists. What’s missing is the ability to watch it all at once and flag what matters.

As AI agents continue to evolve, the number of data points that can be monitored between visits will only grow, producing a more complete and continuous picture of what’s actually happening on the ground. The gap between check-ins doesn’t have to be a blind spot.

The evidence behind these tools is at different stages, from self-reported results across multiple sites to peer-reviewed research not yet in clinical use to a demonstration with no deployment data at all. That range is worth acknowledging. But the underlying pattern is already clear enough to pay attention to: problems don’t have to get lost in the gaps anymore. With these tools, the practitioners own the edge.

Share the Knowledge

If you know someone in senior care, education, healthcare, or infrastructure who’s trying to figure out where AI fits into their work, forward this issue to them. These stories are a good starting point for that conversation.

And if your organization has gaps between check-ins where things are getting missed, I’d like to hear about it. Reply and tell me what’s falling through the cracks. I’m always looking for the next story.

The initial result is often where a story gets told. A pilot works. A study finds something promising. An early deployment shows measurable improvement. If a headline can be made, we’ll hear about it. What we often won’t hear about is what happens later.

That’s where this issue picks up. In each of these four stories, someone looked at early results and decided to go further: expand the scope, test it in harder conditions, build policy around it. That second decision can often be quieter, but it’s where the real commitment shows up.

From Clinical Trial to Campus Platform: Dartmouth’s AI Therapy Chatbot Keeps Going

Getting a first appointment with a mental health professional often takes weeks or months. Depending on insurance and cost, it can be out of reach entirely. Community services have even longer wait times. In the U.S., there is roughly one mental health provider for every 1,600 people diagnosed with depression or anxiety. The help exists. Getting to it is another story.

In March 2025, Dartmouth researchers published the results of the first clinical trial of a generative AI therapy chatbot. The study appeared in the New England Journal of Medicine AI, one of the field’s most respected journals. The trial enrolled 106 people diagnosed with major depressive disorder, generalized anxiety disorder, or an eating disorder. Participants interacted with the chatbot, called Therabot, through a smartphone app. They could respond to prompts about how they were feeling or start conversations when they needed to talk.

The results caught the attention of the research community. Participants with depression experienced a 51% average reduction in symptoms. Those with generalized anxiety saw a 31% reduction, with many dropping below the clinical threshold for diagnosis. Participants at risk for eating disorders showed a 19% reduction in body image and weight concerns, significantly outperforming a control group that received no treatment. The researchers reported that these improvements were comparable to what patients typically experience in traditional outpatient therapy.

One finding surprised even the research team. Participants reported levels of trust and communication with Therabot that were comparable to what patients report for human providers. People weren’t just answering prompts. They were initiating conversations, particularly during nighttime hours, turning to the app during moments of acute distress that an office-bound therapist would never reach. Senior author Nicholas Jacobson described it as people treating the software almost like a friend.

Therabot has been in development since 2019. This is not a ChatGPT wrapper rushed to market. The research team monitored all conversations for safety, and the app prompted users to contact crisis services when it detected high-risk content like suicidal ideation. The researchers were explicit: no generative AI agent is ready to operate fully on its own in mental health care. Rigorous safety benchmarks and clinician oversight remain essential.

That’s where the story stood a year ago. Here’s what came next.

Dartmouth launched a new clinical trial testing an adapted version of Therabot for people dealing with cannabis use disorder alongside anxiety or depression. The study is focused on whether AI-assisted support can help manage substance use and emotional distress together, a combination that traditional treatment often addresses separately.

The bigger move is Evergreen. Announced in October 2025, Evergreen is a campus-wide AI wellness platform designed not as therapy, but as what the team calls a “smart front door to care.” Where Therabot was built to treat diagnosed conditions, Evergreen focuses on prevention: helping students manage stress, build study habits, connect with campus resources, and notice when things are starting to slip before they reach a clinical threshold.

The platform integrates passive sensor data from phones and wearables, with user permission, to help students recognize rising stress levels before they feel the symptoms. More than 200 Dartmouth students are involved in the development, helping to write the chatbot’s dialogue so it sounds like a peer rather than a clinical script. The first randomized controlled trial with 200 students is expected to be underway this year, with a larger trial of up to 600 students planned for 2027.

Jacobson framed the approach directly: “Moving fast and breaking things doesn’t work with student wellness. We need scientifically developed, rigorously tested interventions.”

That progression matters. A clinical trial produced real evidence. The team didn’t stop there. They expanded into harder conditions, built a broader platform, and kept the same commitment to testing before deployment. In a space crowded with apps making big claims on thin evidence, the Dartmouth team is doing the slower, harder work of proving it at each step.

Article
Evergreen Announcement
Cannabis Use Disorder Trial (ClinicalTrials.gov)

Key Stakeholders

Geisel School of Medicine at Dartmouth
Center for Technology and Behavioral Health (CTBH)
Dartmouth Health
Dartmouth Hitchcock Medical Center

A Red Pixel in the Snow: How AI Is Changing Mountain Rescue

In September 2024, an experienced climber and orthopaedic surgeon named Nicola Ivaldo set out alone into the Cottian Alps. He didn’t tell anyone which route he was taking. When he failed to show up at work on Monday, rescuers found his car parked at the village of Castello di Pontechianale, in the Valle Varaita. From there, he could have gone toward either of the area’s two most prominent peaks: the 3,841-meter Monviso or its neighbor Visolotto. His last phone signal placed him somewhere in the area, but that still left an enormous stretch of mountain terrain to search.

More than fifty rescuers searched on foot for nearly a week. A helicopter flew multiple sorties. No one on the popular routes had seen him, which meant he had likely taken one of the more remote paths. By the time early snow arrived in late September, any hope of finding him alive had faded and the search was called off.

In July 2025, the Piemonte Mountain and Speleological Rescue Service went back. This time, they brought drones and AI.

Drones launched from high on the mountain slopes covered 183 hectares of mountainside over five hours, capturing more than 2,600 high-resolution photos. AI software scanned the images pixel by pixel, flagging dozens of potential anomalies in a matter of hours. Work that would have previously required a human analyst to review each photo individually was compressed into an afternoon.

But the AI’s output still needed human judgment. The software flagged pieces of plastic garbage, unusually colored rocks, and what one operator called “hallucinations.” Rescue teams narrowed the list using their knowledge of Ivaldo’s skill level and the routes he would have been likely to take. They ended up with three possible locations, including one that contained a red object.

The next morning, drones went to check the sites. The red object turned out to be Ivaldo’s helmet, lying in a shaded gully on the north wall of Monviso at roughly 3,150 meters. His body was found nearby, partially covered in snow. The software had detected the red color even though the helmet was in shade when the image was taken. Without the AI flagging that dot, he may never have been found.

The technology came too late for Ivaldo. But the Italian search was the most detailed deployment so far of a capability that has been building across countries for several years.

In 2021, software called SARUAV, developed by a spinoff from the University of Wrocław in Poland, helped locate a missing 65-year-old man with Alzheimer’s disease who had suffered a stroke the day before he disappeared. The software analyzed 782 aerial images and detected him in just over four hours, providing his coordinates to the Bieszczady Mountain Rescue Service. He was found alive, believed to be the first time an automated human detection system was directly involved in a rescue. Two years later, the same algorithm located a body in the Austrian Alps. In Scotland, a different system developed by the Lake District Search and Mountain Rescue Association found a missing hillwalker in 2023.

The technology still has real limitations. Drones are nearly useless in forests or dense vegetation. The AI produces false positives that require experienced human filtering. Croatian mountain rescue teams, among the first in Europe to adopt drones back in 2013, have reported that the mix of vegetation and complex karst terrain in their mountains generates too many misleading results for current algorithms to be reliable.

The application extends well beyond mountain rescue. Missing persons cases in populated areas face the same fundamental problem: large areas to search, limited personnel, and a shrinking window of time. Stories of a child or an adult with dementia gone missing from home, with the body found days or weeks later in a nearby wooded area or field, are more common than anyone would like. Drone-based AI search could give law enforcement and volunteer teams the ability to cover that ground in hours rather than days. The terrain is often easier than alpine rock faces, and the stakes are the same.

The next step researchers are working toward is real-time analysis during active searches, processing images on board the drones as they fly rather than after they land. That shift could be the difference between recovering a body and rescuing a person who is still alive.

Article

Key Stakeholders

Mountain and Speleological Rescue Service of Piemonte (CNSAS)
University of Wrocław
SARUAV
Bieszczady Mountain Rescue Service (GOPR Bieszczady)
Lake District Search and Mountain Rescue Association (LDSAMRA)
Italian Research Institute for Geo-Hydrological Protection (IRPI)
Politecnico di Torino University
University of Glasgow
Croatian Mountain Rescue Service (HGSS)

Share

From Pilot to Policy: Utah Goes All In on Government AI

It’s not uncommon to hear about a government pilot project, get the results, and then move on to something else. Utah started with a pilot and kept going.

In June 2024, the state’s Division of Technology Services gave 257 employees Gemini licenses to see whether AI could improve productivity across state agencies. The results were clear enough to act on: among the heaviest users, 79% reported saving one to five hours per week. Eighteen percent, mostly in coding and technical roles, reported saving more than six hours per week.

Then Google changed the math. In September 2024, the company announced that its standalone Gemini chatbot would be available to all Google Workspace users at no additional cost. What had been a paid pilot for 257 employees could now expand to the full workforce. Utah moved quickly. By February 2025, the state had rolled out Gemini for Workspace across its agencies. Roughly 5,500 of Utah’s 22,000 state employees are now at least weekly users.

The state picked a single metric to track progress: time saved. A survey last July showed employees collectively saving about 40,000 hours per month. A more recent survey pushed that number to 60,000 hours per month. The state’s goal is to reach one million self-reported hours saved per year by the end of 2026.

Utah’s AI Director Christian Napier has been candid about the limitations of this metric. It’s self-reported, not independently audited, and it doesn’t capture quality improvements or other intangible benefits. But he argues that settling on one clear number and sticking with it is what prevents the “analysis paralysis” that stalls government AI programs. When you’re asking a legislature for continued funding, simplicity matters.

One deliberate decision stands out. Napier’s office has been careful not to connect time savings to arguments for reducing headcount. That reframing shifted the internal conversation from “AI will take my job” to “AI helps me do my job.” It’s a governance lesson as much as a technology lesson: how you talk about AI adoption inside an organization determines whether employees resist it or use it.

Utah is also piloting Claude Code, Anthropic’s AI coding tool. Early survey results show 55% of participants reporting at least 20% efficiency gains after just one week, though that data is very preliminary.

All of this would make Utah a notable story on its own. But what sets the state apart is where it went next.

Through the Office of Artificial Intelligence Policy, Utah has created a regulatory sandbox that allows AI companies to test applications in regulated industries under state oversight. The sandbox doesn’t waive regulations permanently. It provides temporary, structured relief so the state can evaluate whether these tools work safely before deciding on broader policy.

The most significant agreement so far is with Doctronic, an AI health platform authorized to handle prescription medication renewals for chronic conditions. This is the first state-approved program in the country that allows an AI system to legally participate in medical decision-making for prescription renewals. The scope is deliberately narrow: a formulary of 192 low-risk, non-controlled medications like those for high blood pressure or cholesterol. A physician network and pharmacists remain part of the process. The pilot is tracking refill timeliness, patient satisfaction, safety outcomes, and cost impacts, with findings to be shared publicly.

Two other sandbox agreements are active. ElizaChat, a Utah-based company, is piloting an AI mental health support app in participating school districts. The app doesn’t provide therapy. It engages students through guided conversations, encourages coping strategies, and escalates to licensed professionals when it detects risk. Dentacor, a dental technology company, is testing AI-assisted diagnostics under similar oversight.

Utah isn’t operating in isolation. Pennsylvania has reported that its ChatGPT pilot saved individual employees about eight hours per week. New Jersey’s in-house AI assistant is cutting time on manual processes across agencies. Arizona, Texas, and Wyoming are building their own AI sandboxes. But Utah’s combination of scaled productivity tools, a clear measurement framework, and a regulatory sandbox for healthcare AI puts it further along the path than any other state.

The progression tells the story: pilot, scale, measure, then push into harder territory. Each step built on the evidence from the one before it.

Article
Gemini Pilot (InnovateUS)
Doctronic Partnership (Utah Dept. of Commerce)
ElizaChat Agreement (Utah Dept. of Commerce)

Key Stakeholders

Utah Division of Technology Services
Google
GovRAMP
Anthropic

The Six-Week Gift: How AI Reshaped Special Education Paperwork in a Year

Every public school student with an educational disability in the U.S. is legally required to have an Individualized Education Program. These IEPs can run up to 50 pages. They cover academic goals, behavioral plans, accommodations, and specially designed instruction, all of it grounded in federal law. And they have to be rewritten every year for each student.

That work falls primarily on special education teachers, a workforce already under severe pressure. In the 2023-24 school year, 70% of public schools reported special education teacher vacancies. The paperwork burden is one of the main reasons teachers cite for leaving the field. When caseloads are unmanageable, the quality of IEPs suffers. Teachers copy and paste between students. One parent received her son’s IEP with another student’s name scattered throughout the document.

In January 2025, a preliminary study published in the Journal of Autism and Developmental Disorders offered an early signal that AI could help. Fifteen special education teachers who used ChatGPT to write IEP goals for preschool students with autism produced higher-quality goals than a control group. On a standard clinical rating scale from zero to ten, the AI-assisted group scored between 9.1 and 10. The control group ranged from 5.5 to 9.2. The study was small, but the finding was specific: AI didn’t just save time. It produced better individualized goals.

That was early 2025. The field moved fast.

By late 2025, a report from the nonprofit Center for Democracy and Technology found that more than half of licensed special education teachers were using AI in some form to help develop IEPs and 504 plans. Fifteen percent reported using AI to write a plan in full, nearly double the rate from the previous year. Teachers using AI tools weekly reported saving up to six weeks of administrative work over the course of a school year. For a profession losing people to burnout and paperwork, six weeks back is not a minor benefit.

The tools themselves changed too. Early adopters had been using generic platforms like ChatGPT. By early 2026, specialized tools designed specifically for special education had largely replaced them. Platforms like Monsha, Playground IEP, and PowerSchool’s PowerBuddy for Special Programs are built for compliance with IDEA (the Individuals with Disabilities Education Act, the federal law that governs special education) and student data security, problems that general-purpose AI was never designed to handle. These tools generate compliant IEP goals, draft transition documents, and manage the procedural requirements that consume so much of a teacher’s time, all within systems that protect personally identifiable student information.

The research kept advancing. Andrea Harkins-Brown at Johns Hopkins University, whose team ran the original ChatGPT study, expanded her work in two directions. She began replicating the study with elementary school students and a different disability profile to test whether the quality improvements hold across settings. Her team also published research on using AI to translate complex IEP terminology into plain language for families, addressing a longstanding problem: parents who receive a 50-page document full of legal and medical jargon they can’t easily parse.

Harkins-Brown’s team has also been working with districts on an important distinction: the difference between procedural compliance and substantive compliance. AI can handle the procedural side efficiently, making sure the right sections are filled in and the legal requirements are met. But substantive compliance, making sure the plan genuinely reflects what a specific child needs, still requires a teacher who knows the student. The risk with AI-generated IEPs is that they look complete but aren’t truly individualized. Her team is training districts to use AI for the procedural work while protecting the substantive judgment.

States have started responding with formal oversight. Oklahoma and South Carolina enacted laws requiring school districts to maintain formal AI policies. Some jurisdictions now require written parental consent if AI is used to draft any portion of a legally binding education plan. The Center for Democracy and Technology flagged concerns about bias in AI-generated recommendations, warning that models trained on limited data can produce stereotypical or inequitable suggestions for students with disabilities.

The transformation has been remarkably fast. In roughly a year, AI in special education went from a 15-teacher experiment to a tool used by more than half the profession, accompanied by purpose-built platforms, ongoing research, and new state legislation. The speed of adoption reflects the severity of the problem it’s solving. But it also means the guardrails are still being built while the tools are already in widespread use. That gap between adoption and oversight is the space to watch.

Article
CDT Report (GovTech, Nov 2025)
State AI Legislative Tracker (FutureEd)

Key Stakeholders

Consortium for School Networking (CoSN)
Center for Democracy and Technology
Center for Technology in Education at Johns Hopkins University
National Center for Education Statistics (NCES)
PowerSchool
Monsha AI
Playground IEP
Teach Plus

Takeaway

The surface pattern is obvious enough: four organizations took early results and expanded. But the more interesting observation is that none of them expanded by just doing more of the same thing.

Dartmouth didn’t run a bigger therapy trial. They pivoted from treating diagnosed conditions to building a prevention platform for an entire campus. The mountain rescue teams didn’t just repeat the same drone searches. They refined methods across countries over several years, and researchers are now pushing toward real-time analysis that could change the technology from a recovery tool into a rescue tool. Utah didn’t just give more employees access to Gemini. They built a regulatory sandbox and let AI participate in medical decision-making. Special education didn’t just keep using ChatGPT. The field developed purpose-built tools, researchers drew a line between procedural and substantive compliance, and states started legislating.

In each case, the second step was harder than the first. It involved higher stakes, more institutional commitment, and new questions the pilot was never designed to answer. A clinical trial can show that a chatbot reduces symptoms. Deploying a wellness platform for thousands of students is a fundamentally different undertaking. A drone search can find a missing person. Processing images in real time during an active search is a different engineering problem. Proving that employees save time with AI is straightforward. Allowing AI to legally renew prescriptions requires building an entire regulatory framework.

That’s the distinction worth paying attention to. The first result proves a concept. The second decision reveals what an organization actually believes about it. And the distance between those two moments is where the most important work happens.

Share the Knowledge

This issue covers mental health, search and rescue, government operations, and education. If you know someone in any of those spaces who has early AI results and is trying to figure out what to do with them, forward this to them.

If your organization ran a pilot that worked and then stalled, or if you’re in the middle of making that second decision right now, reply and tell me about it. That’s exactly the kind of story I want to find.

Share Smarter Society AI

Most organizations don’t have a data problem. They have more of it than they know what to do with. Hospital records go back years. Police departments log everything from overtime hours to file access. Research repositories hold datasets from thousands of participants. The collection part is solved.

What’s harder is knowing what to look for. Data collected for one purpose can contain signals that matter for a completely different reason. A radiology report filed to document an injury may also be documenting a pattern of abuse. Attendance records maintained for payroll may also reveal early signs of professional deterioration. The information is already there, doing one job. The question is whether anyone is reading it closely enough to see what else it’s telling them.

Reading the Records: AI Spots Intimate Partner Violence Risk in Routine Medical Data

Intimate partner violence is one of the hardest problems for healthcare to catch. Victims often don’t disclose, for reasons that make complete sense: fear of retribution, safety concerns, stigma. The system has relied on patients to volunteer information that many are not in a position to share. When they don’t, the opportunity for early intervention passes.

A research team at Harvard Medical School decided to test whether the answer was already in the data. They trained AI models on several years of hospital records from nearly 850 affected female patients and 5,200 matched controls, comparing three approaches: one that reads structured data like lab results, visit frequency, and demographics in tabular form; one that reads unstructured data from written medical notes and radiology reports; and a combined model that fuses both.

Radiology reports are particularly useful here. Patterns of physical trauma, including the type, frequency, and location of injuries, can show up across imaging records in ways that are consistent with abuse, even when the patient attributes each individual visit to something else.

The combined model identified IPV risk with 88% accuracy. That’s a strong starting point, but in a context where both false positives and false negatives carry real consequences, there’s clear room for improvement as the models develop. Ideally, the tool would also provide clinicians with a probability score for each case rather than a simple flag, giving them a better sense of how to approach the conversation. The tool doesn’t diagnose or force disclosure. It flags risk so clinicians can start a supportive conversation earlier, backed by guidance the research team developed for approaching those conversations carefully.

The tool isn’t deployed yet. The team plans to embed it in electronic medical record systems for real-time risk evaluation. But the underlying argument is hard to dismiss: the patterns were already in the records. The system just wasn’t reading them.

Article

Key Stakeholders

National Institutes of Health (NIH)
National Institute of Biomedical Imaging and Bioengineering (NIBIB)
Harvard Medical School
Mass General Brigham

Watching the Watchers: London’s Met Police Pilots AI for Officer Misconduct

Police work, like a lot of professions, tends to go unnoticed when officers are doing things right. The spotlight finds them when things go wrong. I’ve done field research with multiple police departments over the years. I’ve worked with officers who were great, some who weren’t so great. I’ve seen the good, bad, and ugly. Anything that helps identify problematic behavior before it becomes a misconduct case is worth paying attention to.

London’s Metropolitan Police is piloting an AI tool built by Palantir that cross-references internal data the force already collects: sickness levels, absence patterns, overtime, and file access. Rather than relying on periodic vetting snapshots, the system generates a continuous risk score for each of the force’s 46,000 employees. Officers who cross a set threshold are flagged to the Directorate of Professional Standards for human review.

The Met says internal evidence shows a correlation between significant levels of sickness, increased absences, and unusually high overtime on one hand, and failings in standards, culture, and behavior on the other. The Police Federation pushed back hard, calling the approach “automated suspicion” and warning that the tool risks misinterpreting unsustainable workload pressures or legitimate illness as indicators of wrongdoing.

Both sides raise fair points. The Met has faced sustained public pressure to improve its culture following high-profile misconduct cases, and a tool that surfaces behavioral patterns early could help supervisors intervene before problems escalate. But monitoring tools are only as fair as the assumptions built into them. What counts as “unusual” overtime? At what point does a sickness pattern become a red flag rather than a sign that someone is burning out?

The approach isn’t limited to policing. Any profession with high stress, heavy caseloads, and burnout risk has known risk factors that correlate with larger problems down the line. Healthcare, social work, teaching, emergency services. Monitoring those risk factors proactively, rather than waiting for a crisis to surface, could help organizations intervene early with support rather than discipline. The behavioral signals may differ across professions, but the underlying logic is the same: the data is already being collected. Using it to prevent problems is a better outcome than using it to document them after the fact. What matters is that the algorithm identifies possible concerns, not makes decisions. Human judgment is still essential to determine whether a flag reflects a real problem or just a difficult month.

The pilot is time-limited, and no outcome data has been published. Whether the tool actually improves standards or just adds a new layer of institutional anxiety is an open question. But the premise is worth watching: the data was already being collected. The question is whether connecting it reveals something useful or just something uncomfortable.

Article
MPS confirms the use of AI to spot potential misconduct

Key Stakeholders

Metropolitan Police
Palantir Technologies
Police Federation of England and Wales
NHS England
Ministry of Defence
Liberal Democrats
Labour Party
Global Counsel

From Data to Diagnosis: AI Finds Preterm Birth Signals Researchers Missed for Years

My youngest daughter was born six weeks premature. She spent 11 days in the NICU because her body still needed to learn to properly regulate its temperature. It was stressful but not worrisome, because doctors said this was normal and there were no other signs of issues. Other families often aren’t that lucky. Premature births can lead to lasting motor and cognitive impairment, and about 1,000 babies are born too soon in the U.S. every day. It remains the leading cause of newborn death.

Researchers at UC San Francisco and Wayne State University wanted to find predictive signals in pregnancy data that scientists had been collecting for over a decade. They had microbiome data from about 1,200 pregnant women tracked across nine studies, along with blood and placental tissue samples from the March of Dimes Preterm Birth Data Repository. The problem wasn’t the data. It was the time and expertise needed to analyze it.

The team tested eight generative AI tools on the same task that had previously taken expert teams months of work through an international research competition. Four of the eight produced prediction models that matched or outperformed those built by experienced data scientists. The AI generated working analytical code in minutes rather than the hours or days it would take experienced programmers. A UCSF master’s student and a high school student built viable models with AI assistance, work that would normally require a much larger team of specialists.

The significance goes beyond speed. If AI can handle the labor-intensive work of building analytical pipelines from a short specialized prompt, the bottleneck in medical research shifts from “can you code the analysis?” to “can you ask the right question?” For a condition as poorly understood as preterm birth, that shift matters. The study was published in Cell Reports Medicine, and the researchers stress that AI is not a replacement for scientific expertise. Only half the tools tested produced usable results, and human oversight remains essential to catch errors. But the path from raw data to actionable prediction just got considerably shorter.

Article

Key Stakeholders

UC San Francisco (UCSF)
Wayne State University
Bakar Computational Health Sciences Institute (BCHSI)
March of Dimes Preterm Birth Data Repository
DREAM (Dialogue on Reverse Engineering Assessment and Methods)
New York University (NYU)
National Institute of Child Health and Human Development (NICHD)
ImmPort
Huron High School

Seeing the Storm Before It Strikes: NASA’s AI Reads the Sun

Solar storms don’t announce themselves. They build. Magnetic stress accumulates beneath the Sun’s surface, energy reorganizes, and then an eruption sends charged particles toward Earth at high speed. The consequences are real: damaged power grid transformers, overloaded satellite electronics, disrupted GPS and navigation systems, and increased radiation risk for astronauts and polar flight crews. The May 2024 X-class flare caused widespread disruption and was visible as auroras at unusually low latitudes.

The problem with current forecasting is timing. Traditional systems react after solar activity intensifies. By then, the window for protective action is already shrinking. Satellite operators, power grid managers, and launch directors need lead time.

NASA researchers at Marshall Space Flight Center built Surya, a large AI model trained on nine years of imagery from the Solar Dynamics Observatory, the spacecraft that has continuously photographed the Sun since 2010. The Sun is captured in multiple types of light and magnetic field measurements, and Surya analyzes all of them at once to build a more complete picture of what’s happening on the surface. Rather than waiting for an eruption to begin, the model learned to identify the subtle patterns that consistently precede one, enabling multi-hour flare forecasts and four-day solar wind predictions that outperformed traditional image classifiers in testing.

The model has limitations. It sometimes misses subtle but important features because it prioritizes overall accuracy over fine detail. Future versions may incorporate uncertainty estimates to better reflect the Sun’s unpredictable nature. The research has been published as a preprint but has not yet been peer-reviewed, though the NASA institutional backing carries weight. No agency is using Surya for live forecasting decisions yet.

NASA open-sourced the model and code on GitHub, a meaningful step given that proprietary solar forecasting tools already exist commercially. If the approach holds up under further testing, earlier warnings could give operators the time they need to reorient satellites, protect grid infrastructure, or delay launches before a storm arrives.

Article

Key Stakeholders

NASA’s Marshall Space Flight Center

Takeaway

The pattern across these four stories is straightforward: AI reading signals in data that was already being collected. But the more interesting shift is what kind of reading is happening.

In each case, the data was originally gathered for one purpose and turned out to contain information useful for a completely different one. Radiology reports were filed to document injuries, not to identify patterns of abuse. Attendance records were kept for payroll and scheduling, not to flag early signs of professional deterioration. Pregnancy microbiome data was collected to study reproductive health broadly, not to predict premature delivery from AI-generated code. Solar imagery was captured to observe the Sun, not to train a model that forecasts eruptions hours before they happen.

That distinction matters. The previous generation of AI-for-good stories was largely about doing existing tasks faster: processing applications, writing reports, summarizing documents. These stories are about something different. The AI isn’t accelerating work people were already doing. It’s finding meaning that people weren’t looking for, in data they were already collecting for other reasons.

That raises a practical question for any organization sitting on years of records, logs, or datasets: what else might your data be telling you that you haven’t thought to ask? The answer may not require new collection. It may just require a different set of questions.

Share the Knowledge

This issue touches healthcare, law enforcement, maternal health, and infrastructure protection. If you know someone working in any of those spaces who’s thinking about what AI can do with the data their organization already collects, forward this to them.

This issue’s stories all started with the same realization: the data was already there, just being used for something else. That’s true in most organizations. If you’re sitting on records, logs, or datasets that you suspect contain more insight than anyone has had time to look for, reply and tell me about it. I’m always interested in where the next signal is hiding.

Share Smarter Society AI

A city truck drives the same streets every morning, passing potholes and dumped mattresses without noticing them. A benefits application sits in a queue, about to get rejected because the photo was too blurry. A patient in an encampment three blocks from a clinic is not likely to walk through the doors.

None of it is hidden. The systems in place just weren’t designed to spot them.

This issue: three stories where AI gave institutions the ability to see what was already in front of them. Plus a look at how AI is helping researchers find patterns in online conversations about addiction that have been happening in the open for years.

Seeing Its Own City: San Jose’s AI Pilots Deliver Results

San Jose has been running two AI pilots since 2023, each targeting a different piece of city infrastructure. Both just posted Phase 2 results.

Street hazard detection. Most cities learn about potholes and illegal dumping the same way: a resident files a 311 report. That system works well enough for people who know it exists, have time to use it, and speak English. For everyone else, the problem sits. San Jose mounted AI-powered cameras on city vehicles that already drive nearly every street: garbage trucks, utility vehicles, fleet cars. As they run their regular routes, the cameras flag hazards automatically. In Phase 2, the system identified 70% of street issues before any resident reported them and detected 50% more non-pothole hazards, like mattresses and garbage piles, than 311 calls captured at all.

Transit signal prioritization. The second pilot puts AI on the city’s traffic signal network. The system analyzes real-time traffic and adjusts signals so buses encounter fewer red lights. After a 2023 test on two routes cut red-light wait times by 50%, San Jose rolled the system out citywide. Bus speeds have increased 20%.

San Jose funded both projects by combining five separate grants, with 90% of the money coming from state and federal sources, paired with private partnerships including the Toyota Mobility Foundation. For cities watching their budgets, that’s a replicable funding model.

Article

Key Stakeholders

City of San José
Toyota Mobility Foundation
LYT
Valley Transportation Authority (VTA)

The Blurry Document: How a Bad Scan Derails a Benefits Case

Applying for public benefits like SNAP often means uploading documents from a phone: pay stubs, ID cards, utility bills. A blurry photo, a wrong file type, a missing signature. Small mistakes, but any one of them can stall or sink an application. The applicant may not find out until weeks later, after a caseworker opens the file and sends it back. For families waiting on food assistance, that delay can be the difference between having meals or missing them.

Pennsylvania’s Department of Human Services launched an AI tool that scans uploaded documents the moment they come in. If something is illegible, blurry, or missing, the system flags it immediately and prompts the applicant to fix and resubmit before the case ever reaches a caseworker. During a two-month pilot covering roughly 12,000 documents, the tool flagged quality concerns in 25% of submissions. Illegible or incorrect documents dropped 80%, saving caseworkers more than 700 hours.

A companion tool targets the other end of the process. A SNAP Case Checker reviews applications for common errors, like miscalculated wages or missing signatures, before they go to final quality control. During the pilot, it flagged more than 1,000 cases for additional review, directly reducing the state’s error rate.

Neither tool makes eligibility decisions. Both catch problems that would have silently delayed or denied benefits for people who qualified.

Article

Key Stakeholders

Pennsylvania Department of Human Services
CODE PA
Pennsylvania Office of Administration

The Patients No One Sees: AI-Powered Street Medicine in California

Nearly half of people experiencing homelessness in the U.S. report poor or fair health. Only 39% have a primary care provider. The clinics are there, but getting people through the door is another story.

Akido Labs built a platform called Scope AI that flips the model. Instead of waiting for patients to show up, non-medical outreach workers carry tablets into encampments and shelters and conduct structured interviews on-site. The AI listens in real time, suggests follow-up questions, and generates a recommended diagnosis and care plan. A remote physician reviews and approves the plan, typically the same day. Since launching in LA County encampments in 2023, the system has served more than 5,000 patients. Each physician’s active caseload grew from roughly 200 patients to 350, a 75% increase in capacity without adding doctors.

The model has expanded to the Bay Area, with partnerships now active in San Francisco, Oakland, Richmond, and San Jose. In Los Angeles, two-thirds of patients now see a provider on their first day of enrollment, contributing to a 40% reduction in emergency department visits.

The approach has drawn criticism. A homeless patient’s living situation affects what treatments are practical in ways AI may not account for. There are also concerns about algorithmic bias and deploying undertested systems on populations least equipped to push back if something goes wrong. And the 99% diagnostic accuracy figure comes from Akido itself, not independent validation.

The counterargument is straightforward: these patients currently receive no care at all. A system that gets most of it right, with a licensed physician still signing off on every plan, changes the math for people who were never part of the equation.

Article

Key Stakeholders

Akido Labs
Five Keys
Reimagine Freedom
Young Women’s Freedom Center
CalMatters

Notable Mention: Using AI to Hear From the 95% Who Never Seek Treatment

Only about 5% of people with substance use disorder ever seek formal treatment. The other 95% never walk into a clinic. But many of them are talking openly about their experiences in online communities on Reddit, TikTok, and YouTube.

A researcher at Drexel University is using large language models to analyze these forums at scale, identifying patterns in how people talk about drug use, stigma, and harm reduction that clinical surveys can’t capture. Machine learning analysis of Reddit posts found that most users in substance-related forums were seeking practical safety information, not glorifying use. On TikTok, 33.9% of substance-related content analyzed was recovery advocacy; only 6.5% showed active drug use. Monitoring Reddit discussions could also help predict opioid-related overdose rates faster than CDC data, which typically lags six months or more.

This isn’t a deployed tool, which is why it’s here rather than in the main stories. But the method matters. Any organization sitting on large volumes of unstructured text, whether it’s client feedback, community forum posts, or open-ended survey responses, can apply the same approach to find patterns they’d never have the capacity to identify manually.

Article

Key Stakeholders

Drexel University

Takeaways

A city relies on residents to flag problems its own crews can miss, given the number of streets they have to keep track of. A benefits system lets a blurry photo stall an application without telling the applicant. A healthcare system waits for patients who are never going to show up.

Those aren’t failures. But they place an obligation on the people the system exists to serve. And the people least equipped to meet that obligation are usually the ones who need the service most. A resident working two jobs isn’t filing a 311 report. A parent applying for SNAP on a cracked phone isn’t producing a perfect scan. A person living in an encampment isn’t scheduling a primary care appointment.

When they can’t meet the obligation, the system doesn’t break. It just quietly stops working for them. No error message. No alert. The pothole stays. The application stalls. The patient goes unseen.

What AI did in each of these stories wasn’t just make data more visible. It eased the obligation. San Jose stopped requiring residents to be the city’s eyes. Pennsylvania started catching document problems before they became denials. Akido brought the exam to the patient instead of waiting for the patient to find the exam.

The question worth asking isn’t whether your system works. It’s who it stops working for when the process doesn’t go smoothly.

What’s worth noting: this isn’t sector-specific. Every organization that depends on the people it serves to trigger the right process, submit the right document, or show up at the right time has the same vulnerability. And AI that catches what people miss, or meets them where they are, doesn’t ask anyone to change their behavior. The system just starts working better for the people it was already supposed to serve.

One more thing…

Between issues, I’ve been using AI to build small tools that make my own workflows easier. No coding background required. If this newsletter covers how organizations are using AI to solve big problems, I’m also finding it can solve the small ones. If you’ve been thinking about how AI could help in your own work, bring it to your manager. The conversation has to start somewhere.

Every system has friction. Forms that can take weeks to process. Databases that require several queries to get all the information needed. Or spending the better part of a day waiting for a review approval that should take minutes.

The people running those systems feel it. But the people waiting on the other side pay for it.

The Relatives Were There. Maryland Just Needed Help Finding Them.

A caseworker trying to find relatives for a child entering foster care is often working from an incomplete list of names: whatever a biological family member can provide in the middle of a crisis, which often may not be much. The relatives who would say yes to taking that child in may be out there, but manually tracking them down takes time that caseworkers often don’t have. And with only 57 licensed foster homes for every 100 children in need across the U.S., the stakes for finding those connections fast are high.

Maryland’s Department of Human Services is now using Binti, an AI tool that pulls from public and proprietary databases to surface potential relatives and close family friends, then helps caseworkers send outreach directly.

Since launching in September 2025, caseworkers have run more than 4,500 searches and identified more than 4,300 potential kin connections. As a result, the share of foster children being placed with family has jumped 33%. Maryland also paired the technology with a 2024 law prioritizing kin placement and a streamlined licensing pathway for family caregivers, which pushed licensed kin caregivers from 25% to 86% in a single year. The software made the search feasible. The policy made the placement possible.

(The full article includes additional context on the nationwide foster care shortage.)

Article

Key Stakeholders

Maryland Department of Human Services (DHS)
Binti

Eyes on Every Hallway: Midland ISD’s AI Gun Detection System

When a suspected weapon appears in a school hallway, seconds matter. But automated systems that alert law enforcement directly carry their own risk: a false alarm triggers a police response, and the consequences of getting it wrong cut both ways.

Midland Independent School District in Texas is deploying ZeroEyes across all 40 of its campuses, covering roughly 29,000 students and 3,200 staff. The system layers onto existing security cameras and flags suspected firearms automatically. But it doesn’t call law enforcement directly. Every alert routes first to a monitoring center staffed by military and law enforcement veterans, who verify the threat and dispatch actionable intelligence: the suspect’s description, weapon type, and last known location. The whole process takes 3 to 5 seconds.

For a district of this size, consistent human monitoring across 40 campuses would be either prohibitively expensive or simply impossible. ZeroEyes makes the coverage feasible while keeping a human in the loop before any response is triggered.

Article

Key Stakeholders

Midland (TX) Independent School District
ZeroEyes

Connecting the Dots: Virginia Builds a Single View of the Overdose Crisis

When an uptick in overdoses starts developing in a Virginia community, the data that can reveal it likely already exists: in hospital records, 911 logs, drug seizure reports, social service files. The problem is that it lives in four or five separate agencies. Harm reduction workers and policy officials had to submit requests to each one and wait for answers. By the time the picture came together, the window for early intervention had often already closed.

Virginia’s new Substance Use Disorder Analytics platform pulls that data into a single view, with heat maps that pinpoint emerging overdose hot spots in real time. But the more useful feature isn’t just seeing where the crisis is developing. It’s being able to compare those trends against where funding and services are currently directed, and quickly identify whether the response infrastructure is actually in the right place.

The platform also tracks what’s in the drugs being found, illicit market activity, and regional demographic trends, giving officials enough detail to tailor interventions rather than react broadly. Virginia built SUDA with input from agencies across the commonwealth rather than copying another state’s model, and plans to expand access to local governments, researchers, tribal organizations, and emergency responders.

Article

Key Stakeholders

Virginia Office of Data Governance and Analytics
Virginia Information Technologies Agency
Virginia Opioid Abatement Authority
Deloitte

Finding the Hours: How Idaho Is Using AI Across State Government

Idaho has been quietly deploying AI across multiple areas of state government, targeting the same thing in each case: time that staff spend on process instead of people. The results are starting to show.

The most striking example is in child welfare. Caseworkers carry enormous caseloads, and every hour spent reviewing documentation to make a recommendation is an hour not spent with the children they serve. In Idaho, that review process was taking 4.5 to 6 hours per case.

The state built a language model, with specific privacy protections for the sensitive data involved, that now handles that review in under a minute. Caseworkers still make the final call, but the time burden that came before it is largely gone.

Idaho has applied the same logic elsewhere. At the DMV, AI now automates data entry and transcribes handwritten documents, improving accuracy and saving staff time. A new onboarding tool draws on agency procedures and rules to bring new employees up to speed faster: what used to take months now takes weeks.

Article

Key Stakeholders

Idaho Office of Information Technology Services (ITS)
Idaho Transportation Department (ITD) - Division of Motor Vehicles (DMV)
Idaho Department of Health and Welfare
Government Technology
Idaho State Government

Notable Mention: Avocado Health Puts Pediatric Guidance in Any Parent’s Pocket

Anyone who has raised a child knows how many questions come up during the day and night. They show up at 2 a.m. when a baby spikes a fever, or mid-grocery-run when a toddler melts down. Most parents turn to Google or ask their own parents. The answers are abundant and often contradictory.

Avocado Health is a 24/7 SMS and WhatsApp-based parenting support platform that limits its AI to a curated knowledge base built by pediatricians, therapists, and child development experts. No app download required. When a question falls outside that scope, it escalates to a human Child Life Specialist rather than guessing. The platform is HIPAA-compliant and designed to work on any phone, with Spanish support planned.

The platform is still in early pilots with no published outcome data yet, which is why it’s here rather than in the main stories. But if it delivers on its promise, it could give parents an evidence-based resource around the clock, and take at least a little of the stress out of one of the hardest jobs there is.

Article

Key Stakeholders

Avocado Health

Takeaways

None of these stories are about missing information. The relatives existed. The patterns were in the data. The case details were in the file. The cameras and security infrastructure were already in place. What was missing was the time and capacity to get to it fast enough.

That distinction matters. These tools didn’t create new knowledge. They removed the cost of accessing knowledge that was already there. And in each case, that cost wasn’t being paid equally. A child welfare caseworker buried in six-hour reviews has fewer hours for the children on the caseload. Officials managing overdose response couldn’t quickly tell whether their resources were positioned where the need was actually developing.

As these tools scale, the potential compounds. Faster case reviews mean more children placed with family sooner. Better resource positioning means fewer communities where the response arrives after the damage is done. The friction was always a tax on outcomes. Reducing it doesn’t just save time. It changes what becomes possible.

Share the Knowledge

If you know anyone who might find these stories interesting, please forward it along.

And if friction is slowing down your work in ways that feel ridiculous, reply and tell me about it. I’m always looking for the next story.

Every call matters. The clinical judgment that catches a patient declining before it becomes a crisis. The coaching conversation that helps a struggling teacher improve. The police report that accurately captures what actually happened. The screening that identifies someone who needs help before they’re ready to ask for it.

Getting the call right isn’t just about having good instincts. It’s about having the right information at the right moment. This issue looks at five cases where AI is helping make that possible.

Catching Sepsis Before It Wins: Tampa General’s AI Command Center

Sepsis is one of the hardest conditions to treat, not because the treatment is complicated, but because it moves quickly. An infection turns systemic before symptoms are obvious. By the time the pattern is clear, the window for effective intervention is already narrowing.

Tampa General Hospital was already running a care coordination center across its seven hospitals and 150+ care locations. Working with a data analytics platform, they rebuilt it into a fully integrated system, giving care teams real-time visibility across the entire network. The goal was to catch problems earlier in their progression, giving staff the window they needed to intervene before patients got sicker. The first major application was a Sepsis Hub that flags high-risk patients before the situation becomes critical, pushing alerts to care teams while intervention is still possible.

The outcomes have been significant. As of November 2025 (reported in an updated story linked below), the Sepsis Hub has been credited with saving more than 700 lives, with sepsis patients leaving the hospital 30% sooner. The broader system improvements followed. Bed assignment time dropped 83%, meaning patients get to the right care faster. MRI turnaround improved 30%. And time spent in post-anesthesia recovery dropped 28%, reducing one of the most common bottlenecks between surgery and an available hospital bed.

Article
Story Update

Key Stakeholders

Tampa General Hospital
Palantir

Building the Evidence Base: Stanford’s AI Research Hub for K-12 Education

Stanford’s Generative AI for Education Hub is trying to solve a real problem in K-12 education. The hub, part of the university’s SCALE (Systems Change Advancing Learning and Equity) initiative, was built in response to a growing concern: schools are adopting AI tools faster than anyone can evaluate them. Teachers are using an average of 49 different ed-tech products, with little evidence the accumulation is improving outcomes. The achievement gap hasn’t closed. Equity goals remain unmet. The tools keep coming anyway.

The hub collects, conducts, and presents research on AI use in K-12 schools, giving district leaders a searchable resource to inform their decisions. Studies are searchable and filterable, so administrators can quickly find research relevant to their specific needs without wading through academic papers. The repository recently crossed 1,000 studies.

The TutorCoPilot study is a good example of what the hub surfaces. Instead of replacing human tutors with chatbots, researchers embedded an AI assistant to support 900 tutors in real time. Students working with AI-assisted tutors were 4% more likely to achieve topic mastery than the control group. For students paired with less experienced tutors, the gains were even larger, up to 9%. The takeaway: AI may be most valuable not as a replacement for human instruction, but as a tool that raises the floor for less experienced educators.

Article
Article Update
Tutor CoPilot

Key Stakeholders

Scale Hub
Stanford University

A Simple Picture Test That Could Change How We Diagnose Addiction

Getting an accurate picture of substance use disorder is harder than it sounds. Patients often underreport or deny symptoms because of the stigma attached to addiction. That can make the initial screening process slower and less reliable than it needs to be.

Researchers at the University of Cincinnati developed an AI tool that takes a different approach to the problem. Instead of relying solely on self-reporting, the system asks patients to rate how much they like or dislike 48 pictures with mildly emotional content, and complete a short questionnaire about their substance use behaviors. Combined with basic demographic information, the AI uses those judgment patterns to assess addiction.

The study examined 3,476 participants. The system identified the type of substance being used, stimulants, opioids, or cannabis, with 82% accuracy, and predicted the severity of the addiction with 84% accuracy. Because the assessment can be completed on a smartphone or computer, it offers a low-cost, noninvasive first step for clinical screening that doesn’t require patients to directly confront or disclose their addiction.

Article
Original Study

Key Stakeholders

UC College of Engineering and Applied Science
University of Cincinnati
Northwestern University
Massachusetts General Hospital
Harvard Medical School
Gold Research Inc.
Office of Naval Research
Chestnut Health Systems

From One Hour to Fifteen Minutes: Motorola’s AI Suite for Public Safety

Police officers spend roughly 40% of their shift behind a keyboard. Call handlers spend nearly half their time on 911 calls just verifying information. The paperwork isn’t a minor inconvenience. It’s time taken directly away from the community.

Motorola’s Responder Assist Suite addresses this by pulling together data from 911 audio, body cameras, in-car cameras, and radio transcripts into a single unified record. The flagship tool, Narrative Assist, is designed to help officers write more accurate reports by pulling together everything that was captured during an incident — the 911 call, body camera audio and footage, radio traffic, and in-car video — and cross-referencing that against the officer’s own account, flagging any discrepancies for review while keeping a clear audit log of every AI suggestion.

The White Bear Lake Police Department in Minnesota has been the most documented early adopter. Narrative Assist cut report writing from one hour to 15 minutes. Video redaction dropped from 35 hours to one. The department estimates it saves 40 hours a week across the force. Because the suite is built into software already used by approximately 3,600 of the 6,000 emergency dispatch centers in the US, many agencies can access these tools through a simple software update rather than a new system installation.

Article

Key Stakeholders:

Motorola
White Bear Lake Police Department

Raising the Bar: AI Steps Into the Teacher Evaluation Process

Teacher evaluations are supposed to drive improvement. In practice, they often don’t. Feedback can take days or weeks to reach a teacher after an observation, by which point the specific moments that needed discussion are hard to reconstruct. In districts with high turnover, the problem compounds. There simply aren’t enough experienced administrators to keep up.

Minnesota’s Southwest Intermediate District 288 is piloting Evaln, an AI-powered classroom observation tool that records lessons and generates rubric-aligned feedback before the coaching conversation even happens. Once a video is uploaded, Evelyn analyzes it against the Danielson Framework for Teaching, a widely used research-based evaluation model, and produces timestamped notes and direct quotes for administrators to review.

In beta testing, Evaln’s evaluations aligned with those of trained human administrators to within 0.07 points on average. The tool can also track multiple student conversations simultaneously, something no single observer can do in real time. The pilot hasn’t been without complications. Early testing showed Evelyn scoring special education classrooms too critically because it didn’t account for instructional accommodations. But that’s exactly what a pilot is supposed to surface. The district identified the problem, made adjustments to the system, and is continuing to evaluate results. That kind of iterative refinement is how responsible AI adoption is supposed to work.

Article

Key Stakeholders

Owl Labs
Evelyn Learning
The Danielson Group
MN SW Intermediate District 288

Suggested Takeaway

The pattern across these five stories isn’t hard to spot: AI delivering better information, faster, to the people who need it most. But the deeper shift is worth noting.

In every case here, a human is still making the final call. The clinician decides on treatment. The officer signs the report. The administrator conducts the evaluation. The sepsis alert draws from thousands of patient data points. The police report is cross-referenced against every piece of captured evidence. The classroom evaluation tracks conversations no single observer could follow simultaneously. AI isn’t replacing that judgment. It’s changing what that judgment is based on, drawing from a richer, more complete evidence base than any individual could assemble alone.

That’s a meaningful distinction. The question worth asking as these tools expand isn’t just whether they work, but whether the humans relying on them understand what they’re being given. Better information leads to better decisions only when the person receiving it knows how to act on it.

Share the Knowledge

Know someone in healthcare, public safety, education, or social services who’s navigating AI adoption in their work? Forward this issue to them.

The tools in this issue all keep a human in the loop. But that raises a real question: how do you make sure the people using these tools actually know how to interpret what they’re being given? If you’ve seen this done well, or poorly, reply and tell me about it.

The traffic engineer is standing at an intersection with a clipboard. The climate planner is staring at a spreadsheet that covers twelve provinces and twelve years of data. The teacher has thirty essays due back by Friday, and it’s already Wednesday. In Nigeria, a patient needs to talk to someone at 2am.

In each case, the professional exists. The expertise is real. There just isn’t enough of it to go around.

This issue looks at what happens when AI extends that expertise further than any of these people could reach alone.

Sitting at a Red Light That Doesn’t Know You’re There: Raleigh’s AI Traffic Pilot

Every driver knows the feeling. Sitting at a red light, wondering how long the cross traffic gets to go before it’s your turn, waiting for a signal that has no idea what’s actually happening around it. For a city like Raleigh, that frustration multiplies across hundreds of intersections, and until recently, the only way to understand traffic patterns was to send staff out with clipboards to count cars by hand.

Raleigh is piloting an AI system that converts real-time video from about a dozen of the city’s 200+ traffic cameras into continuous traffic counts, giving timing engineers data they can act on immediately rather than waiting for periodic manual surveys. The city doesn’t store the footage or collect identifying information. On a corridor like Capital Boulevard, which carries 60,000 vehicles a day, even a 1% improvement in signal timing could save tens of thousands of dollars every day in time and fuel. The goal: if the pilot works, scale it citywide.

Full Article

Key Stakeholders:

City of Raleigh (NC)

Drowning in Climate Reports While the Clock Keeps Ticking: IWMI’s AI Helps Planners Act Faster

Climate planners in developing nations aren’t short on data. They have vulnerability assessments, national statistics, climate scenarios, infrastructure models, and years of project investment records. What they often lack is the capacity to pull it all together into a budget decision before the next funding cycle closes.

The International Water Management Institute unveiled its Climate Risk Intelligence Assistant at the United Nations’ annual global climate conference (COP30). The tool is built into IWMI’s existing Climate Smart Governance dashboard and lets national planners ask targeted questions rather than manually parsing datasets. A planner can ask how much has been invested in water resilience over the past five years, or simulate what a 20% rise in drought-prone areas would mean for infrastructure costs.

The system scans climate scenarios from the World Bank, UN Sustainable Development Goal indicators, and national project data, then surfaces a consolidated answer. The dashboard is already in use across Kenya, Senegal, Sri Lanka, and Zambia, with the AI assistant now layered on top to make it faster and more accessible.

Full Article

Key Stakeholders:

International Water Management Institute
CGIAR
Coalition for Disaster Resilient Infrastructure (CDRI)
World Bank
United Nations (UN)
Artificial Intelligence Climate Institute

From Weeks to Minutes: Magna Education’s AI Grading System

Every parent has seen it. A child works hard on an assignment, hands it in, and then waits. And waits. By the time feedback arrives, the class has moved on.

Magna Education’s AI grading system targets that gap directly, providing rubric-aligned feedback on essays, data analysis, and written work in minutes rather than weeks, while surfacing patterns across an entire class that teachers would normally spend weeks uncovering. Teachers get time back to focus on instruction and deeper engagement with students; students get feedback while the material is still fresh. In one Louisiana school using Magna for Advanced Placement courses, overall student scores improved significantly in a single semester, with students earning top marks tripling from 12% to 35%. Gains weren’t limited to top performers. Score distributions shifted at every level.

Full Article

Key Stakeholders:

Studygenie
Magna Education

It’s 2 a.m. and There’s No One to Call: AI Chatbots Fill Nigeria’s Mental Health Gap

Nigeria has 262 psychiatrists for a population of 240 million. The healthcare budget consistently falls below 5% of national spending. For most Nigerians dealing with depression, anxiety, or emotional crisis, professional mental health support isn’t delayed. It simply isn’t available.

AI-powered chatbots are stepping into that gap. Platforms like HerSafeSpace’s Chat Kemi and FriendnPal, developed with scripts from licensed Nigerian psychologists, offer mood tracking, cognitive behavioral therapy techniques, and immediate first-line support, including referrals to licensed therapists when users are in crisis.

The tools have limitations: psychologists caution they can’t replace professional judgment for serious conditions, and data privacy protections remain weak. But for a country where the alternative is often nothing, the impact is straightforward. At 2 a.m., when something is wrong and there’s no one to call, something answers.

Full Article

Key Stakeholders:

FriendnPal
Blueroomcare
HerSafeSpace

Notable Mention: The Agenda Is Public. Getting Through It Is Another Story: SBX Builds an AI Civic Navigator

Nobody wants to spend their evening sifting through city council agendas, budget allocations, and policy documents. But those decisions directly affect everyday life, and for low-income families, foster youth, and English learners, the complexity is enough to overwhelm even the most motivated resident into giving up.

SBX Youth & Family Services, a Moreno Valley-based nonprofit, received a grant from the OpenAI Foundation’s People First AI Fund to build an AI Civic Engagement Navigator. The tool will transform public records and government documents into clear, digestible information with practical steps for residents who want to take action. SBX plans to engage youth, parents, and local partners as co-designers throughout development, with the goal of building something replicable across other regions. The underlying argument is simple: when communities can understand what their government is doing, they can actually participate in it.

Full Article

Key Stakeholders:

SBX Youth & Family Services (Sigma Beta Xi, Inc.)
OpenAI Foundation

Suggested Takeaway

The pattern across these five stories isn’t subtle. A dozen cameras doing what dozens of engineers couldn’t. An AI assistant turning months of climate reports into a budget answer in minutes. A grading system that gives students feedback while the lesson is still fresh. Chatbots fielding mental health calls at 2 a.m. in a country with 262 psychiatrists for 240 million people. A tool that turns impenetrable government documents into something a family can actually act on.

None of these tools created new expertise. They acted as a translation layer, sitting between overwhelming complexity and the humans who need to act on it, converting what was too slow, too dense, or too scattered into something decision-ready.

The real question isn’t about replacement. It’s about reach. Where are people already overwhelmed? What becomes possible when AI extends their capabilities? These five stories suggest the answer lies in exactly the places where the need has always outpaced the capacity, but never had enough resources to close the gap.

Share the Knowledge

Know someone working in a field where demand has always outpaced resources? Forward this to them.

Where is capacity the real bottleneck in your work? Reply and tell me where your team is already stretched thin. I’ll keep an eye out for relevant stories.

But raw data isn’t insight. A genetic test returns thousands of results. Which ones actually explain the symptoms? A camera captures the block. Is that pile illegal dumping or scheduled pickup? Food gets thrown out. Was it avoidable? A blind person’s reflection exists. But what does it actually show?

This newsletter usually focuses on what’s working now. But understanding where AI is heading matters as much as where it’s landed. This issue looks ahead: four tools still in development, each designed to decode reality into something specific enough to act on.

From Thousands to Ten: AI Prioritizes Genetic Test Results

Genetic tests return overwhelming results. A single patient can have thousands of mutations flagged, most with unclear significance. Doctors sift through the noise trying to find the one that actually explains the symptoms.

Scientists at Mount Sinai’s Icahn School of Medicine have built an AI system that predicts not just whether a mutation is harmful, but which disease it likely causes. In testing with real patient data, the system ranked all mutations by likelihood of causing the disease. The actual culprit frequently appeared in the top ten—turning thousands of possibilities into a manageable shortlist.

The tool could accelerate diagnosis for rare conditions and help drug developers identify targets for new treatments. It’s still in the research phase - no clinical deployment yet - but represents a step toward medicine tailored to an individual’s specific genetic profile rather than one-size-fits-all approaches.

Full Article

Key Stakeholders

The Mount Sinai Hospital
Icahn School of Medicine at Mount Sinai
The Charles Bronfman Institute for Personalized Medicine
The Mindich Child Health and Development Institute
National Institutes of Health (NIH)
Fondation Leducq
Leona M. and Harry B. Helmsley Charitable Trust

The Mirror That Speaks: AI Describes Appearance to Blind Users

Most people check their reflection without thinking. For blind people, that information has been inaccessible. Until now.

Apps like Be My Eyes and Envision use AI-powered image recognition to analyze photos and generate detailed verbal descriptions of users’ faces and bodies. Some users have requested appearance ratings. For many, this is the first time they’ve received specific feedback about how they look.

The technology comes with tradeoffs. Researchers warn that access to appearance feedback exposes blind users to the same unrealistic beauty standards sighted people face. AI models trained on narrow datasets can produce biased or inaccurate descriptions that users can’t verify. Long-term research on the psychological effects of these tools on blind users doesn’t exist yet.

Still, users largely view the technology positively. It provides information they couldn’t access before. The question is whether the benefits of self-knowledge outweigh the risks of external judgment.

Full Article

Key Stakeholders

Be My Eyes
Envision
University of Bristol
Northeastern University
Aira

Every Block, Every Route: Dallas Puts AI Eyes on Garbage Trucks

Illegal dumping is hard to catch in real time. Code enforcement officers can’t be everywhere, and complaints come after the fact.

Dallas is mounting AI-powered cameras on its yard waste and trash trucks starting early next year. As trucks drive their regular routes, cameras will photograph curbsides and AI software will flag problems: illegal dumping, trash piles, bulk waste violations.

The system gives sanitation staff a city-wide picture without adding vehicles or headcount since the trucks are already driving every block. Privacy advocates have raised concerns about cameras on municipal vehicles. Dallas says the system will blur faces and license plates automatically.

Other cities have deployed similar technology. San Jose uses AI cameras on city vehicles to detect potholes and hazards. Tacoma monitors recycling contamination. Dallas is betting the same approach can reduce dumping.

Full Article

Key Stakeholders

Dallas City Council
Dallas Sanitation Department

Bake Less, Feed More: Whole Foods’ AI Turns Waste into Eggs

Grocery stores throw away food, then order more. The waste is visible in aggregate numbers but invisible at the source. No one tracks which items hit the bin.

Whole Foods plans to install AI-powered bins from startup Mill across its stores by 2027. The bins dehydrate and grind food scraps into chicken feed. But the bigger value is the data: AI analyzes what gets thrown away in real time.

If the system detects a pattern, like croissants consistently landing in the bin, it signals the kitchen to bake fewer. The goal is cutting waste at the source, not just processing it after the fact.

Mill says the bins reduce waste volume by up to 80%. The scraps become chicken feed, the feed goes to hens, and the eggs end up back on Whole Foods shelves: a closed loop from wasted produce to protein.

Full Article

Key Stakeholders

Amazon
Whole Foods Market
Mill

Takeaway

The pattern here: AI as translator. Each of these tools takes data that already existed and converts it into a form specific enough to act on. Genetic tests already flagged thousands of mutations. Garbage trucks already drove every street. Stores already threw away food. The information was there. The interpretation wasn’t.

That shift matters. For decades, the bottleneck was collection: how do we get more data? Now the bottleneck is meaning: what does this data tell us to do? These tools sit at that translation point, turning raw signal into ranked priorities, flagged incidents, inventory decisions.

But translation isn’t neutral. The AI Mirrors story makes that clear. When AI describes a blind person’s appearance, it’s not just reporting facts, it’s also applying standards, making comparisons, rendering judgment. Users can’t verify what the AI says. They have to trust it.

That’s the open question as these tools develop: when AI decodes reality, whose interpretation are we getting? Organizations adopting these tools should be asking: what was this trained on, what does it prioritize, and what does it miss? The translation is only as good as the lens.

Share the Knowledge

Know someone drowning in data they can’t make sense of? Whether it’s a researcher, city planner, or operations manager sitting on information they can’t act on—forward this to them.

What raw data is your organization sitting on? Reply and tell me what information you’re collecting but struggling to interpret. I’ll keep an eye out for tools that might help translate it.

The pattern is there. The question is whether you catch it in time.

This issue: four stories where AI spots the warning signs and buys time to act. Thirty seconds before a patient falls. Six weeks before a reef bleaches. Milliseconds before a near-miss becomes a collision. And training that lets officers practice de-escalation before the real crisis happens.

Kentucky’s AI-Powered De-Escalation Lab

Stephen Ware’s lab at the University of Kentucky is building intelligent training designed to calm tense encounters, improving officer decision-making and preparedness. The system works like a “choose your own adventure” story. If an officer tries a new way to talk someone down, the digital character reacts naturally, forcing the officer to think on their feet.

The goal is to give officers tools to bring everyone home safely by resolving conflicts with words rather than force. If it works, AI- and VR-powered training could become a standard tool in public safety globally.

Full Article
Video

Key Stakeholders

Narrative Intelligence Lab (University of Kentucky)
National Science Foundation

How Cities Are Using AI for Road Safety

Cities are getting smarter about where crashes happen before they happen. Austin is using computer vision to watch intersections for near-misses, then recommending cheap fixes like reflective tape and better crosswalk paint. Hawaii’s DOT built a statewide model that predicts where crashes are most likely and suggests what to do about it. Early estimates: 6.9 lives saved over three years if they follow the recommendations.

New York City equipped 3,000 vehicles to communicate with traffic infrastructure. Drivers reported feeling safer (83%), and speed compliance improved 16%.

Full Article

Key Stakeholders

City of Austin (TX)
Hawaii Department of Transportation
NYC Department of Transportation
MicroTraffic
Velodyne
TAPCO
Miovision
Draper

30 Seconds Before the Fall

Hospital falls follow a pattern. A patient shifts, starts climbing out of bed, reaches the tipping point. By the time a nurse rotating between rooms notices, it’s already happened.

Emory University Hospital Midtown is using AI-powered sensors that watch for the movements that come before a fall. When the system spots trouble, it alerts staff up to 30 seconds before impact and plays a voice telling the patient to stay put. Meanwhile, virtual nurses working remotely handle the paperwork (admissions, medications, discharge planning), so bedside nurses can focus on patients instead of screens. Emory is expanding to 1,000 beds this year.

Full Article

Key Stakeholders

Emory Healthcare
VirtuSense Technologies

Predicting Coral Bleaching Before It Starts

Florida’s Coral Reef is bleaching more frequently and more severely. The 2023 marine heatwave was a record-breaker. By the time you see bleaching, the damage is already done.

Scientists at the University of Miami built an AI model that forecasts heat stress at specific reef sites up to six weeks in advance, accurate to within a week. That’s enough time for managers to initiate emergency cooling or shading measures, prioritize monitoring, and adjust plans before the coral starts dying. The model complements NOAA’s existing Coral Reef Watch system by adding the localized timing signal it was missing.

Full Article

Key Stakeholders

University of Miami Rosenstiel School of Marine, Atmospheric, and Earth Science
Cooperative Institute for Marine and Atmospheric Studies (CIMAS)
University of Colorado - Department of Atmospheric Science
NOAA Coral Reef Watch

Takeaway

The pattern here: AI buying time before harm occurs. Thirty seconds for a hospital patient. Six weeks for a coral reef. Milliseconds for a driver approaching a hidden pedestrian.

But the deeper shift isn’t about time. It’s about localization.

We already knew these harms happen. Patients fall. Intersections are dangerous. Reefs bleach during heatwaves. Encounters escalate. None of this is new information.

What’s new is precision. The AI contribution isn’t discovering risk — it’s making risk specific enough to act on. Not “patients fall” but “this patient, in the next 30 seconds.” Not “reefs bleach” but “this reef site, starting this week.” Not “intersections are dangerous” but “this intersection, these near-miss patterns, these cheap fixes.”

We’ve always had the aggregate data. What we lacked was the ability to intervene before a specific harm, at a specific place, at a specific time. These systems convert diffuse risk into precise warnings that someone can actually do something about.

Share the Knowledge

Know someone working on prevention — in public safety, healthcare, or environmental management? Forward this email to them.

Where could earlier warning help in your work? Reply and tell me what risks you’re currently reacting to that you wish you could see coming. I’ll keep an eye out for relevant stories.

The paperwork gets in the way. Applications take weeks. Permits stall. Doctors spend evenings on documentation instead of patients. Case managers dig through policy manuals while clients wait.

This issue: four stories where AI handles the administrative friction. The people doing the real work can actually do it.

Cutting Red Tape for Families: Maryland’s Benefits Overhaul

Maryland is integrating Claude across state agencies to help residents get government assistance faster. Often, families who qualify for housing or food help get stuck in a maze of confusing applications. By implementing Anthropic’s Claude to help expedite state services to those who need them, the goal is to reduce child poverty and improve access to housing. A Claude-powered virtual assistant helps residents apply for benefits, update their information, and track applications. The tool will also identify additional programs that they may qualify for.

Percepta will be working with Maryland agencies to help with the implementation. The state is also receiving support from the Rockefeller Foundation to complete this pilot project.

Full Article
Maryland Governor's Press Release

Key Stakeholders

Anthropic
Maryland Governor's Office
Maryland Department of Information Technology
Percepta
Rockefeller Foundation

Faster Permits, More Homes: Louisville’s AI Pilot

Louisville (KY) is testing an AI system to solve its housing shortage by speeding up the construction permit process. Right now, the permitting process is described as “unpredictable” and “prone to delays.” The new system will be used to see how much time the process can be reduced. Developers and builders will be providing feedback to further help smooth out the process. The goal is to help builders start projects sooner and lower costs, making it easier for the city to build the affordable housing its residents need.

Full Article

Key Stakeholders

City of Louisville
Govstream.ai

Eyes on the Patient: How AI is Ending Doctor Burnout

Virginia Mason Franciscan Health (VMFH) is showing how AI can be a serious time-saver. By bringing AI into their daily workflows, they’ve started catching care gaps (like missed screenings) that people might overlook, which has already led to a 58% jump in colonoscopy rates.

The tools also handle the heavy lifting of note dictation and complex discharge summaries, saving clinicians about 20 to 30 minutes per patient. For doctors, this means drastically less “pajama time” (documentation done at home after hours) and more energy to actually engage with patients during visits. VMFH keeps a “human in the loop.” Physicians review every AI-generated note to ensure everything is accurate, equitable, and personal.

Full Article

Key Stakeholders

Virginia Mason Franciscan Health
AMA Center for Digital Health and AI

The Digital Assistant: Helping LA Social Workers Do More

Imagine LA (now Amplifi), a nonprofit focused on reducing homelessness and poverty, has partnered with Nava to launch an AI-enabled chatbot for benefits navigators. The tool gives case managers instant access to information on public benefits programs and tax credits, helping them determine client eligibility faster. The underlying “Benefits Navigator” system has already helped households secure an average of $10,869 in additional benefits by streamlining eligibility determinations. To ensure accuracy, the chatbot provides citations for its answers, allowing human caseworkers to verify information against official government sources before advising clients. Cornell and Georgetown’s Better Government Lab are studying the chatbot’s accuracy and impact on administrative burden.

Full Article

Key Stakeholders

Imagine LA (now Amplifi)
Nava PBC
Cornell University
Better Government Lab

Takeaway

The common thread here is that “cutting paperwork” is no longer just a side benefit. It is the primary way modern organizations are getting back to their core missions. Whether it is a doctor, a social worker, or a city planner, these tools are removing the administrative barriers that have historically stood in the way of effective service.

What’s worth noting: every story here keeps humans in the loop. Maryland’s virtual assistant identifies programs, but residents still decide whether to apply. VMFH’s AI drafts notes, but physicians still review and sign off. Imagine LA’s chatbot provides citations so caseworkers can verify answers against government sources. The pattern isn’t AI replacing judgment. It’s AI doing the lookup so humans can focus on the judgment.

That distinction matters. Bureaucracy wasn’t arbitrary: it existed to ensure accuracy, accountability, and fairness. These tools don’t eliminate that burden; they shift who carries it. For now, the shift is working. Whether it holds as these systems scale remains the open question.

Share The Knowledge

Know someone in public service or healthcare? Forward this email to them so they can see how these tools are solving real problems in other cities.

Got a bottleneck? Reply to this email and tell me: what is the biggest “paperwork headache” in your organization that AI could help fix? I’ll see if I have any articles or case studies that may be of interest to you.

The help exists. Parents’ rights are in the IEP document. Domestic abuse hotlines operate worldwide. Pediatric specialists can catch delays early. Dispatchers are trained to save lives.

Reaching it is the hard part. The barriers: language, literacy, geography, safety.

This issue: four stories where AI builds the bridge—translating legal jargon into plain language, erasing digital traces for abuse survivors, putting specialist coaching in rural parents’ hands, and identifying foreign languages in microseconds when someone dials 911.

The right help exists. Now it’s finally getting through.

San Francisco Schools IEP Translation Software

San Francisco parents of special education students now have access to a tool to help understand individualized education plans (IEPs). Developed by Innovate Public Schools, in collaboration with Northeastern University, “AiEP” translates the long (about 50 pages) and dense educational, legal, and medical information found in IEPs into plain language. The tool is free and open-source. It can also translate this information into different languages, such as Spanish, Vietnamese, Cantonese, and Mandarin.

About 200 local families have tested the tool, which addresses a real problem: SFUSD has approximately 2,300 English language learners in special education, but district translation services can take 10+ days. The tool helps parents prepare for IEP meetings with specific questions and catch errors—like one parent who discovered her son was receiving 100 minutes of language arts support instead of the 227 minutes promised.

https://thefrisc.com/this-ai-software-translates-special-education-plans-for-sf-parents/

Key Stakeholders & Developers

Innovate Public Schools
Chan Zuckerberg Foundation
Charles and Helen Schwab Foundation
San Francisco Unified School District (SFUSD)
Northeastern University — Burnes Center for Social Change

Sophia The Domestic Violence Chatbot

Meet Sophia, an AI chatbot designed to help victims of domestic abuse across the world.

Sophia enables abuse victims to anonymously learn about warning signs, understand their legal rights, gather evidence, and seek help—all without leaving a digital trace. The tool is available 24/7 via web browser at sophia.chat, or through WhatsApp, Telegram, and Viber. A video chat option serves users with low literacy. There have been over 42,000 conversations across 172 countries, breaking down language and geographic barriers for people who need it most.

Developed by Spring ACT, Sophia is designed for safety: when users close the window, the tool rewrites browser history and replaces the page with a decoy website. An encrypted digital safe lets users store evidence like photos or screenshots, protected by steganography that hides the password inside an image. The tool doesn’t replace human services; it fills the gap when those services are unreachable or unsafe to contact—particularly for victims living with their abuser.

https://news.microsoft.com/de-ch/2025/07/10/ai-chatbot-sophia-supporting-victim-survivors-of-domestic-violence-wins-the-united-nations-global-ai-for-good-impact-award-2025/

The following article from 2024 provides more detailed information on the tool’s development and features:

https://www.eib.org/en/stories/chatbot-against-domestic-violence-springact

Key Stakeholders & Developers

SpringACT
AI For Good
Microsoft

3 Moves Ahead NC Developmental Delays Platform

Dr. Gavin Colquitt, at Appalachian State University, received a $1.1 million NCInnovation grant to develop “3 Moves Ahead NC,” an AI-powered platform that helps parents identify and address movement-related developmental delays in children from birth to age 3.

Western North Carolina families dealing with developmental delays in young children have limited access to specialists and often have long drives to appointments. Early intervention only works when you can actually access help. An estimated 50% of parents with children facing these delays are forced to reduce work hours or leave the workforce entirely, losing an average of $18,000 in annual income. Appalachian State’s 3 Moves Ahead NC platform tackles this with AI-powered remote coaching that puts personalized support directly in parents’ hands, including customized video content, progress tracking, peer networks, and best practices for children birth to age 3.

The platform launches in Western NC with plans to expand statewide and nationally, with partnerships including military health systems, community colleges, and the Cerebral Palsy Foundation.

https://today.appstate.edu/2025/12/18/ncinnovation

Key Stakeholders & Developers

Appalachian State University
Beaver College of Health Sciences - App State U
Appalachian Institute for Health and Wellness
NCInnovation

Texas World Cup 911 Language Detection

North Texas is preparing for what officials are calling a “39-day rush hour”—the 2026 FIFA World Cup—with an AI dispatcher designed to solve a problem most emergency systems never face: callers speaking 20 to 30 different languages. The North Central Texas Council of Governments approved $200,000 for AI software that can identify a foreign language within microseconds and route the caller to the right translator, bypassing the time-consuming process of conferencing in human interpreters.

AT&T Stadium in Arlington will host nine World Cup matches including a semifinal, drawing hundreds of thousands of international visitors to an area already struggling with traffic and infrastructure. Emergency dispatchers will face callers who might speak Portuguese, Japanese, Arabic, or any of dozens of other languages, all needing help in high-pressure situations. The AI acts as a triage layer: identify the language, connect to translation, get help moving.

What makes this notable: instead of hiring multilingual dispatchers for a six-week event, North Texas is using AI to handle temporary capacity spikes. Whether it works under pressure remains to be seen.

https://www.chron.com/sports/article/world-cup-artificial-intelligence-911-calls-21279120.php

Key Stakeholders & Developers

North Central TX Council of Governments

Takeaway

Four different sectors, same story: the service existed, but people couldn’t reach it. These tools aren’t replacing caseworkers, dispatchers, or specialists—they’re clearing the path so people can actually reach them.

What’s worth noting: none of these tools fix the underlying problem. District translation is still slow. Rural specialists are still far away. Emergency interpreters still take time to conference in. AI built a parallel path. For now, that’s a workaround. But as these tools continue to evolve, the parallel path may become the solution.

What if an organization could see next week’s crisis today? That’s the shift happening across sectors right now. Weather agencies are forecasting hurricanes in minutes instead of hours. Hospitals are filling empty appointment slots before patients even cancel. Cities are modeling evacuations before disasters strike. This week, we look at four stories of organizations using AI to help turn prediction into prevention.

Outpacing the Storm: NOAA’s 40-Minute Forecast Revolution

For decades, predicting the weather meant running massive supercomputers for hours to crunch complex physics equations. NOAA has officially changed the game with “Project EAGLE,” a new AI-powered system that generates a global 16-day forecast in just 40 minutes. Because it uses 99.7% less computing power than the old method, forecasters can now run updates much more frequently. This speed isn’t just convenient; it gives officials significantly more lead time to prepare for rapidly changing storms.

While the system is showing impressive results, especially for tracking hurricane paths, NOAA isn’t throwing out the old maps just yet. Instead, they deploy this AI alongside traditional models as a safety check. This “hybrid” approach ensures that while they take advantage of the AI’s incredible speed, they still have the trusted, physics-based data to back it up. It’ll be interesting to see how well the AI works in predicting rapid intensification hurricanes, as they have become more concerning over the past several years.

This approach represents a shift where AI is used to validate and accelerate human understanding, rather than replacing it entirely.

Links:

NOAA Launches Project EAGLE: The AI Revolution in Global Weather Forecasting

NOAA

Microsoft

NVIDIA

ECMWF

Google DeepMind

The End of the Empty Waiting Room: Recovering £300M in Lost Time

Missed appointments cost the UK’s National Health Service nearly £1 billion a year and waste valuable doctor time. To fix this, the NHS has deployed “DrDoctor,” an AI tool that analyzed 4 billion past records to learn the patterns of why people skip visits. By looking at factors like age, previous history, and even the time of day, the system can predict with high accuracy who is likely to miss their appointment.

The result? If the system predicts a high chance of no-shows for a particular slot, hospitals can strategically overbook, similar to how airlines manage flights, to ensure doctors aren’t left sitting idle. In early tests, one hospital was able to see an extra 9,000 patients in just three months without hiring more staff. If rolled out across the entire country, this simple efficiency gain could save the NHS £300 million annually.

Links:

AI appointments tool could save the NHS £300m a year by reducing no‑shows

DrDoctor

NHS

Mapping the Exit Before the Emergency

Planning an evacuation for a major city usually takes experts hours of calculations, but San Jose is speeding that up ahead of the World Cup and Super Bowl. The city has adopted “Ladris,” software that acts like a super-powered traffic planner for disasters. It digests vast amounts of data—from population density and car ownership rates to live flood maps—and models the best evacuation routes in minutes.

During early tests, the system offered a glimpse of its potential when it predicted where homeless encampments would move during a flood, something human planners hadn’t fully anticipated. This level of “spatial awareness” helps emergency managers position trucks, supplies, and rescue teams in exactly the right spots before a disaster strikes. It turns emergency response into a precise, coordinated operation rather than a chaotic reaction.

Links:

San Jose adds AI tools to its disaster tactics

City of San Jose

Ladris Technologies

Predicting the Surge: How AI is Staffing the Frontlines

Winter is a crisis point for hospitals, with flu season causing sudden overcrowding in Emergency Rooms. The UK government is tackling this with a new AI forecasting tool now deployed across 50 NHS organizations. The system analyzes seasonal health trends, flu data, and upcoming weather patterns to predict exactly when patient numbers will spike—potentially weeks before the surge happens.

This foresight gives hospital administrators a crucial advantage. Instead of being overwhelmed when the waiting room fills up, they can adjust staffing schedules and clear bed space in advance. It transforms hospital management from a reactive scramble into a proactive strategy, reducing wait times for patients and easing the pressure on frontline staff during the busiest times of the year.

Links:

AI Eases A&E Delays, Boosts Health Worker Support

NHS

The Takeaway: AI Gives You Time—But Only If You Spend It

Every story in this issue comes down to the same resource: time. NOAA now updates forecasts more frequently because AI made it affordable. Hospitals book more patients because they know which slots will go unused. San Jose compresses hours of evacuation calculations into minutes. NHS administrators adjust staffing weeks before flu season peaks.

But here’s what separates winners from watchers: the organizations succeeding don’t treat AI-generated time as a bonus. Instead, they treat it as an opportunity to do something different. NOAA runs a hybrid ensemble with both AI and physics models. The NHS embeds engineers alongside booking staff. San Jose asks their software to predict where homeless encampments will relocate during floods, something human planners hadn’t anticipated.

What makes these stories different isn’t the prediction - it’s what became possible because of it. Strategic overbooking. Anticipating movements no one had asked about. Staffing for a surge that hasn’t arrived. The real advantage isn’t knowing what’s coming; it’s inventing the response that only makes sense when you do.

Subscribe now

Subscribe now

Smarter Society AI - The Inaugural Issue

This is Smarter Society AI - a weekly look at how AI is accelerating human progress to solve real problems.

Each issue covers 4-5 stories of AI being deployed in social impact sectors: healthcare, education, government, social services, and beyond. The focus is on what’s actually working - named organizations, specific tools, measurable outcomes.

Let’s get to it.

Riverside County Tests AI to Cut Paperwork, Keep Human Connection

The Riverside County Children and Families Commission (CA) is working with Nava (a public benefit corporation) and Amplifi (nonprofit tech) to build an AI tool to help fill in forms. The project received a $1.5 million grant from Google’s Generative AI Accelerator program.

To the average person, it may seem crazy to spend that sort of money to use AI to fill out forms. However, the time it takes for case workers to assist with forms is extensive. And the information needed is critical to assist with benefits and services. If AI can automate much of this process, it frees up time for case workers to actually interact and better understand clients’ needs.

Links:

How an AI agent can help caseworkers maintain the ‘human relationship’ with clients

Riverside County Children and Families Commission

Nava

Amplifi

Radiology AI Leads to 625 Early Cancer Detections

The Mid-South region has lung cancer mortality rates nearly double the national average. Baptist Memorial Health Care made a public commitment to cut deaths by 25% by 2030 - what they call the “Mid-South Miracle.” The problem: their teams were manually pulling information from multiple systems, validating and re-validating patient lists, working through them repeatedly to make sure no one slipped through the cracks. It wasn’t sustainable at scale.

Baptist implemented Eon, an AI tool that scans radiology reports as soon as they’re signed. When the system spots a potential lung nodule, it flags the case and routes it to the right provider for follow-up. The results: 625 lung cancers detected that might otherwise have been missed or caught later. More importantly, over 70% of those cancers were caught at Stage I or II - the most treatable stages. The system paid for itself in seven months.

Links:

Baptist Memorial bets on AI-enabled care management and wins

Baptist

Eon

76% Report Measurable Gains After AI Accelerator

Charles and Lynn Schusterman Family Philanthropies and the Overdeck Family Foundation launched an AI Accelerator to assist 22 nonprofits in building AI tools to help improve organizational outcomes. Over a six-month period involving coaching and design sprints, every participating organization successfully developed a functional AI prototype. As a result of these efforts, 76% of participants cited measurable improvements in their work processes, while 57% reported significant time savings.

Links:

What We Learned from Supporting 22 Organizations to Build AI Capacity—and Why It Matters for Philanthropy

Charles and Lynn Schusterman Family Philanthropies

Leading Educators

Playlab

Meet Ava: The AI Answering Parking Questions So Dispatchers Don’t Have To

Because of increasing call volumes and staffing shortages, police departments in Wisconsin are turning to artificial intelligence. By using AI to handle non-emergency administrative calls - such as questions about parking or local events - this frees human dispatchers to focus on 911 emergencies. La Crosse County implemented an AI agent named Ava that screens calls, answers routine questions, directs callers to appropriate county departments, and transfers callers to a person when emergency keywords are detected. The system handled more than 40,000 calls between May and December 2025. Waukesha County is in the process of implementing this system as well.

Links:

Wisconsin counties turn to AI to manage non-emergency police calls

Aurelian

Take-aways

A pattern runs through this week’s stories: each organization identified a specific bottleneck - a task consuming disproportionate time relative to its value - and deployed AI there.

Caseworkers in Riverside County spent half their time navigating benefit applications. Dispatchers in Wisconsin fielded thousands of calls about parking tickets. Teams at Baptist Memorial were manually validating and re-validating patient lists to make sure no one was missed.

The AI didn’t replace the work that matters. It cleared the path to it.

That’s the practical lesson here: before asking “how can we use AI?”, ask “what’s eating our staff’s time that shouldn’t be?”

About Heath

Heath Dingwell, PhD, has spent two decades inside several of the sectors this newsletter covers—university professor, field research with police departments and drug courts, content development for healthcare institutions (Mayo Clinic among them) and educational companies. He’s the author of eight books translating academic and policy research for general audiences.