Benefits
Benefits
Search could not load.
Nothing selected yet. Cleared.
- An Apple Watch called for help after a runner’s heart stopped
A case report in Resuscitation Plus, published on 12 August 2026, concerns a 69-year-old man who collapsed in ventricular fibrillation while running alone. He was hidden behind a hedge on a dead-end road.[1] On detecting the fall, his Apple Watch automatically called emergency services on 112 and alerted his emergency contacts.
- In a Taiwan trial, fewer patients died when AI picked out high-risk ECGs
A trial in Taiwan tested alerts to physicians from an AI analysis of electrocardiograms. Its primary endpoint was death from any cause within 90 days.
- Drug found with a knowledge graph lowered COVID-19 deaths in a trial
On 4 February 2020 BenevolentAI researchers reported in The Lancet that their knowledge graph had helped them identify a possible treatment for COVID-19. It was baricitinib, an arthritis drug.
- Ghost Murmur may have located a downed F-15E officer in Iran
The rescue of a downed American airman in Iran is established by the White House briefing of 6 April.[1] The claim that an AI-assisted quantum sensor located him comes from two anonymous sources in the New York Post. President Trump later told the Post that the tool was important but did not explain how it worked.[2] Physicists interviewed by the Post and by Scientific American questioned the capability as described.[3] The contribution attributed to Ghost Murmur remains unverified.
- Lynxight alerts a Perth lifeguard to a sinking swimmer, who is rescued
On 10 June 2025, a man suffered a medical episode and sank in the outdoor pool at Stirling Leisure - Inglewood in Perth. Lynxight is a camera system built to detect swimmers in distress.
- A nurse returns to full-time work after software detects her brain bleed
A 57-year-old nurse suffered a ruptured brain aneurysm. Her symptoms began on 2 February 2024.[1] According to her surgeon, software at Baptist Health's Marcus Neuroscience Institute flagged the bleeding on her scan and alerted clinicians.[2] The clinicians sealed the aneurysm with coils delivered through a catheter.
- A man with a rare blood disorder became well enough for a transplant
Public money now goes to the search for new uses of old drugs. A patient's case is told in support of that spending.[1] In January 2024 a 37-year-old man from Renton, Washington, was critically ill with POEMS syndrome.
- Fewer sepsis patients die after COMPOSER alerts UC San Diego nurses
COMPOSER is a deep-learning model that flags possible sepsis to nurses. It went live in two emergency departments at UC San Diego Health on 7 December 2022.
- Embrace wristband alerts a mother, who calls paramedics
In November 2016 a 26-year-old teacher from Fulton, Missouri, had an epileptic seizure. It left her unresponsive and not breathing.[1] Her Embrace wristband detected the seizure and alerted her mother, who called paramedics.
- Mayo credits AI with finding a woman’s ruptured brain aneurysm fast
A woman's recovery at Mayo Clinic in Jacksonville, Florida, is a story that Mayo tells about its use of AI in stroke care. Her brain aneurysm ruptured on 24 February 2022.
- Transpara helps Swedish breast screening find more cancers
Lund University researchers ran the MASAI trial within the Swedish national screening program. Between 12 April 2021 and 7 December 2022, they randomly assigned 105,934 women in southwest Sweden to Transpara-supported screening or to a control group.
- Insight MMG as a third reader finds more breast cancers in Stockholm
The ScreenTrustCAD study measured what an AI reader adds where breast screening already uses two radiologists.[1] Karolinska Institutet researchers ran it at Capio Sankt Göran Hospital in Stockholm. From 1 April 2021 to 9 June 2022, Lunit's Insight MMG ran in the background on the screening mammograms of 55,581 women.
- Emory diagnoses VT after a smartwatch ECG records a 165 bpm tachycardia
Authors in the Division of Cardiology at Emory University School of Medicine in Atlanta published a case report online on 12 March 2025. They describe a 64-year-old man with a replaced aortic valve.
- Doctors find a pulmonary embolism after Apple Watch records low oxygen
A 50-year-old woman who had recently had COVID-19 was referred to the hospital of Oita University in Yufu, Japan, for breathlessness. Four days before admission, a primary care physician had measured her oxygen saturation at 94 to 98%.
- Doctors find lung cancer after a smartwatch flags a fast heart rate
Clinicians at St George's Healthcare NHS Trust and East Surrey Hospital in England published a case report on 4 January 2022. They describe how a smartwatch alert led indirectly to a diagnosis of lung cancer.
- Smart wristwatch prompts a patient’s return to hospital for an embolism
Drynan, Leroux and Zywiel work in the orthopedic surgery division at Toronto Western Hospital. In a case report published online on 8 June 2021, they credit a smart wristwatch with prompting a patient's readmission to the emergency department.
- Exomiser ranks gene variants to help diagnose rare disease in England
The 100,000 Genomes Project ran a preliminary national study in rare disease. From May 2016 through April 2019, the study returned results to the Genomic Medicine Centers of the English National Health Service.
- Meta’s suicide alert brings Meerut police to a man within 8 minutes
Police said that Meta sent a suicide alert at 6:53 pm on 7 June 2026 about a 25-year-old Meerut man's Instagram video. By their account, a Sardhana police team reached him within eight minutes.
- Raebareli police arrive in 12 minutes after Meta flags a suicide post
The Raebareli rescue preceded a similar rescue in Meerut in June 2026. Both took place under an arrangement in which Meta has alerted the Uttar Pradesh Police since 2022.[1] On 29 March 2026 an 18-year-old posted a suicide video on Instagram.
- Facial recognition helps agents identify 311 probable child victims
Operation Renewed Hope ran from 17 July 2023 into early August at the Cyber Crimes Center of Homeland Security Investigations (HSI). U.S.
- On CT, PANDA detects pancreatic cancers that standard care missed
PANDA is a deep-learning model that identifies pancreatic ductal adenocarcinoma on non-contrast CT. In a study published in Nature Medicine on 20 November 2023, the authors report that PANDA exceeded mean radiologist sensitivity by 34.1% and specificity by 6.3%.
- ASENA flags a car at Yüksekova carrying 52.5 kg of explosives
On 20 March 2022 officers of the Hakkari Provincial Police Directorate stopped a car at a checkpoint in Yüksekova, Turkey. Work on the directorate's ASENA analysis program had flagged the car.
- SARUAV detects an unconscious man in tall grass, whom rescuers save
The Bieszczady Mountain Rescue Service called this the first find in Poland of a missing person using a drone and SARUAV.[1] The missing man was 65. He had Alzheimer's disease and had left home in Cergowa a day after a stroke.
- Spotlight may have led to a missing child’s recovery in weeks
This case rests on the word of one speaker. In a WIRED report of 19 June 2019, Thorn's chief executive, Julie Cordua, said that a California officer had run a missing child's photo through Spotlight.
- Medicare withholds $4.41 billion in suspected fraudulent payments
The Department of Justice reported on 30 June 2025 that its analysts and their collaborators had detected anomalous Medicare billing in Operation Gold Rush. According to the department, health agencies stopped all but about $41 million of about $4.45 billion scheduled for payment.
- AI may have helped Abu Dhabi surgeons separate twins joined at the head
On 5 June 2026 PureHealth reported that surgeons at Sheikh Khalifa Medical City in Abu Dhabi had separated Nigerian twins. More than 60 clinicians took part in four staged procedures.
- Kent reports fewer homeless households after targeted support
Maidstone Borough Council uses an AI-enabled Xantura platform to flag households for proactive homelessness support. In a presentation published on 4 November 2024, a Kent County Council presenter reports that fewer than 1% of flagged households that received support became homeless.
- An early resistance warning helped doctors change a woman’s antibiotic
A 78-year-old woman at China Medical University Hospital in Taichung developed severe pneumonia during a prolonged admission. Within an hour, the hospital's Intelligent Anti-Microbial System predicted from a bacterial mass spectrum that her Stenotrophomonas maltophilia isolate resisted ceftazidime.
- Vyntra’s AI risk models may have stopped a USD 500,000 ATM fraud
Vyntra, formerly NetGuardians, lists about two dozen short case studies at unnamed banks on its “Fraud Cases” page.[1] In one, Vyntra says that its AI risk models flagged the unusual timing and frequency of ATM withdrawals. Vyntra also says that fraudsters were arrested while attempting to withdraw USD 500,000 (KSh 58 million).
- WeChat Pay alert helps a man keep 190,000 yuan from a suspected scammer
On 14 December 2021 WeChat announced an “anti-fraud customer service reminder” for WeChat Pay. WeChat's example was a man in Yunnan, identified only by his surname.[1] ChinaZ reported that the man had met a woman on a video platform.
- Mother’s AI query prompts a rare-disease diagnosis in a 22-month-old
In a case report published on 16 January 2026, Hajebian and Friel describe a 22-month-old girl with skull defects. The treating physician did not initially recognize the defects on her computed tomography scan.
- DeepSeek-R1 flags familial Mediterranean fever, which gene tests confirm
A 34-year-old man of Egyptian origin had recurrent inflammatory attacks from early childhood. Despite almost three decades of symptoms, extensive investigations had not established a diagnosis.
- STAR finds two motile sperm, letting a couple have a girl after 19 years
Columbia University Fertility Center now charges $3,000 for a STAR analysis.[1] The first pregnancy that the center reported with STAR came from a couple who had tried for 19 years to have a child. The husband, who had non-obstructive azoospermia, gave a sample that had been judged sperm-free on manual examination.
- A smartwatch alert leads a 34-year-old to treatment for a weakened heart
Cardiologists in Thessaloniki published a paper on 19 December 2025. They note that premature atrial contractions are a major source of false-positive alerts for atrial fibrillation on smartwatches.[1] In this case, an alert led to the diagnosis of a heart condition that clinicians could treat.
- Smartwatch readings lead to a family’s long QT diagnosis
A man came to the cardiology clinic of University Medical Centre Mannheim, Germany, because of readings from his smartwatch. The visit led to a diagnosis.
- A 76-year-old spots a change on his Apple Watch ECG, leading to a bypass
In a case report published on 30 January 2024, Koster and colleagues at Amsterdam UMC describe a 76-year-old man with no symptoms. On rare occasions and out of interest, he recorded an electrocardiogram (ECG) in lead I on his Apple Watch right after exercise.
- Waymo’s driverless cars draw fewer insurance claims than human drivers
As driverless ride-hailing expands, insurance claims are one way to compare it with human driving. In a peer-reviewed paper of 14 July 2024, Swiss Re and Waymo authors reported zero bodily-injury claims over 3,868,506 driverless miles.
- Doctors treat ventricular tachycardia after Fitbit flags a fast heart
In October 2021 a 55-year-old man in Jamshedpur, India, had symptoms at work. His Fitbit Sense showed an irregular rhythm at 164 beats per minute.
- Camera alerts bring help sooner to residents with dementia who fall
Between May and August 2018, staff in six California facilities that provide memory care received real-time alerts from SafelyYou Guardian. The system uses AI to detect falls on video.
- Zasocitinib clears psoriasis in more patients than deucravacitinib
In the LATITUDE Atlas phase 3 trial, 36.5% of 301 participants on zasocitinib reached complete skin clearance at week 16. On deucravacitinib, 13.9% of 303 participants did so.
- ENAD-DET-01 helps Seoul endoscopists find more adenomas
A randomized trial has tested ENAD-DET-01, a polyp detector from the South Korean company Ainex. At two Seoul hospitals between May 2023 and April 2025, 998 patients were assigned at random to colonoscopy with or without the system.
- ALERTCalifornia AI may have helped contain the Trotter Fire
The Trotter Fire, west of Tehachapi in Kern County, California, started on 14 August 2023. It was held to 52 acres.
- GI Genius lifts adenoma detection in VA colonoscopy units
The Veterans Health Administration tested computer-aided polyp detection in routine colonoscopy at scale.[1] It assigned facilities at random to receive a device for polyp detection. The device is listed in the trial registry as Medtronic's GI Genius.
- Nigeria uses machine-learning poverty maps to direct COVID-19 cash aid
Nigeria had no urban social registry. Its National Social Safety Nets Coordinating Office therefore built the Rapid Response Register for a government-funded cash transfer during the COVID-19 pandemic.
- Vara MG helps German screening radiologists find more breast cancers
PRAIM is the largest of the non-randomized implementation studies in a 2026 meta-analysis of AI in screening mammography. Its result for the safety net appears in a second 2026 meta-analysis.[1] Across 12 German screening sites, radiologists who read with Vara MG found 6.7 cancers per 1,000 women.
Each suggestion gets a public number on GitHub, which needs an account. Private material can go by email instead.
Public submissions, each with its status. Peter H. Vartanian reviews contributions before publication.
The editor’s guide gives every rule. In short:
- A benefit shows the AI system that helped to produce the good.
- Each source is marked with who wrote it.
- The grade comes from the strongest source whose authors credit the system. The grades form three levels, from claimed to found.
- Depth runs from 1, a small gain, to 5, survival.
- A count covers only the benefits found so far. It is not a rate of success.
In briefs, Peter H. Vartanian weighs the evidence for some benefits.
Downloads
CloseEach link below downloads its file from the current edition. Download every file at once
The dataset is one file, records.json.gz, released under CC BY 4.0. Three flat tables come from it:
- records.csv, one row per benefit
- ratings.csv, one row for each rated kind of beneficiary
- echoes.csv, one row per attempt to make a benefit happen again
Query the tables with SQL in your browser
The schema defines every field. Each benefit is also available on its own page as JSON.
For one reader's questions
Each file puts the benefits under the questions of one document, in that document's words. A cell reads “not in the sources” when the sources were read for its question and do not answer it, and “not checked” when no one has read them for it yet.
- For procurement: OMB Memorandum M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, Section 4(b)(iii), Performance-Based Acquisition Techniques
- For impact assessment: OMB Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, Section 4(b)(ii), items A, E and F
- For clinical use: The Joint Commission and the Coalition for Health AI, Guidance on the Responsible Use of AI in Healthcare (RUAIH), Element 4, Ongoing Quality Monitoring
- For regulation: Regulation (EU) 2024/1689, the Artificial Intelligence Act, Article 7(2)(j)
- For court use: UNESCO, Guidelines for the Use of AI Systems in Courts and Tribunals, Principles 1.3, 1.6 and 1.13
The list of readers gives each document's passage and address.
For research
A schema defines this table.
Profiles of the benefits
These files can be downloaded:
Fructus
The monetary outcomes of AI. The financial claims, each with its sources. JSON CSV
Counting
A benefit is one good from one system in one setting. A benefit reported in several sources is one benefit with several sources. A benefit that happened again elsewhere is an echo of the first.
Fields
- id
- A stable identifier.
- title, summary, date_observed, account
- What happened. The title gives it in one line. The summary gives it in a paragraph. The account gives the full story. The date gives when it happened.
- brief
- A brief from Peter H. Vartanian, where he has written one. It gives his reading of the evidence. Its sidenotes are in brief_notes.
- headline, metric
- The headline is a short title that adds no fact. Asterisks around a name set it in italics. The metric is the measured effect of the AI, copied verbatim from the text of the benefit. It is null when the benefit gives no number.
- date_basis, date_precision
- The basis shows what date_observed marks. The precision is the unit of the date. A partial date uses the first day of its period only for sorting.
- benefit_status
- How far the benefit happened, from intended to realized. In the data, the value for realized is reported. The value for partly realized is mixed. The status is separate from the grade of the evidence.
- category
- The kind of benefit. The editor’s guide lists the five kinds.
- unexpected, purpose
- The purpose is what the system was built to do. The field unexpected shows whether the benefit differed from that purpose.
- system, model_version, developer, deployer, location, domain, segment
- The system, with its version where known. The developer made it. The deployer ran it. The remaining fields show where it was used.
- sector
- One of 14 sectors, in the CSV. The fields domain and segment give the detail.
- unrated_because
- For a benefit with no depth, the reason.
- ratings
- For each kind of beneficiary that benefited, a depth from 1 to 5 with its reason. The depth is null when it is not rated.
- ai_role
- What the system contributed, assessed apart from the rest. It can be marked as not assessed.
- learns_from_data
- Whether a source shows that the system learned from data how to produce its outputs. The value comes with the words of the source. It is empty when unknown.
- evidence, confidence
- Each source carries its details with the tie of its authors to the system. Only sources that support the benefit set confidence, the grade of the evidence. The value of confidence follows the three levels:
- high: found
- medium: independent, the level of the two middle grades
- low: claimed
- comparator
- What the benefit is better than, in these parts:
- against: the baseline that the benefit is measured against
- before: the figure without the AI
- after: the figure with the AI
- basis: whether the difference was measured
- quote: the passage that gives the comparison
- url: the address of that passage
- later
- What happened after the first report. The editor’s guide lists the outcomes. Each comes with its date, beside the passage that supports it. A challenge means that someone disputed a claim. It does not mean that the claim was reversed. Unknown means that no later outcome is established in the benefit.
- value
- A financial benefit that a source gives. The amount is quoted with its currency, beside the passage that supports it.
- claim_status, claim_basis
- The claim_status gives the authority of the strongest source. The claim_basis gives the cited evidence with its limitations. Both are computed from the evidence.
- replication
- The conditions that produced the benefit, which must hold again for it to recur. The echoes are the attempts to repeat the benefit, each with its outcome.
- benefit_categories, broad_benefits
- The broad benefits: up to two categories, the main one first. Ten take their names from the framework of SCSP and Johns Hopkins APL (2023). The other four are the database’s own. The definitions are the database’s own.
- scsp_example
- The example in SCSP’s list that the benefit matches, by name, when one does.
- size
- Four inputs of the magnitude. Each is a band from 1 to 5, with the quotation that it rests on. When the sources give none, the reason appears instead. The CSV gives the bands.
- probability, magnitude
- Both are computed, never entered. Probability is 3 for a realized benefit. It is 2 for a partly realized one. It gains one for each other deployment where the benefit recurred, up to two. Magnitude follows SCSP’s formula, which the editor’s guide gives. It is exact, a multiple of 0.005. It appears only when all six inputs are documented.
- registry_ids
- The numbers under which trials were registered, where a benefit rests on a registered trial.
- deepest, groups (CSV)
- deepest gives the highest depth across the kinds of beneficiary. groups lists the kinds that were rated, separated by semicolons.
- comparator_against, comparator_before, comparator_after, comparator_basis (CSV)
- The comparator above, flattened into four columns.
- later_status, later_date (CSV)
- The later outcome above, with its date.
- learns_from_data_source (CSV)
- The address of the source that shows whether the system learns from data.
- value_amount, value_what, value_estimate (CSV)
- The financial value above, flattened into three columns.
- sources (CSV)
- The number of distinct addresses of sources. Several quotations from one document count once, so it can be lower than the length of evidence in the JSON.
- The CSV of the list of benefits
- The benefits shown, in the order of the list. Its columns:
- ID
- Benefit
- Benefit status
- Date, to the precision shown
- Date basis
- Countries
- Sector
- Benefit type
- Depth
- Evidence, the label of the grade
- Sources
- Reported measure
- Link
Colophon
Artificial Intelligence Benefits Database is made by Peter H. Vartanian. Write to peter.vartanian@icloud.com.
This site is released under CC BY 4.0, with its dataset.
The banknotes beside amounts of money are drawn for the database. They copy no part of a real note.
The border copies the calibration strip that Smithsonian Libraries photographed beside a plate of E. A. Séguy’s Papillons.
Set in Newsreader, Public Sans and IBM Plex Mono, all used under the SIL Open Font License. The icons of the sectors are Bootstrap Icons, under the MIT License. The logos of organizations are their own marks, shown to identify them. The source of each logo is listed in logos.json. Built 6 November 2026.
-
Independently studied
Claim Use Fall Detection with Apple Watch
Independent source Wearable saving lives: a case report
United Statesto
NetherlandsSpecific harm prevented Assumed comparison
Persisted
WANT reported on 1 May 2026 that the runner had spent two months in rehabilitation and aimed to run the Doetinchem half marathon in October 2026. No later report on the race appears.
Trade-offs The device shortened the path from an unwitnessed collapse to human help. Survival still depended on CPR, defibrillation, and hospital care, and this case cannot measure the outcome without the alert.
Limits The authors give no date or place. The identification of the patient, the date and the street are inferred from interviews and a pager message. Men's Health says that the runner also pressed the SOS button himself and that a pulse returned within about three minutes. This account does not fit the timeline in the paper. The case cannot show how quickly he would have been found without the alert.
Tried again It happened again in one other setting.
Needs The watch has connectivity at the scene. The collapse produces a detectable fall signature.
As a 69-year-old runner collapsed in cardiac arrest, Fall Detection on an Apple Watch Series 10 called emergency services. It alerted his emergency contacts. He made a full neurological recovery.
-
Independently studied
Claim AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial
Independent source Randomized Controlled Trials Evaluating Artificial Intelligence in Cardiovascular Care: A Systematic Review
Without the system 4.3% mortality within 90 days
With it 3.6% mortality within 90 days
Trade-offs The alert could lead to more tests and treatment, and the trial team found higher mean medical costs in its economic analysis.
Limits The deployed model weights, longer-term survival and the clinical actions that caused the mortality difference remain unknown.
In a Taiwan trial of 15,965 patients, a neural network alerted physicians to high-risk electrocardiograms. Mortality at 90 days was 3.6%, against 4.3% with usual care.
-
Drug found with a knowledge graph lowered COVID-19 deaths in a trial
Independently studiedMagnitude 0.48
Magnitude 0.48
Claim Baricitinib as potential treatment for 2019-nCoV acute respiratory disease
Independent source Baricitinib in patients admitted to hospital with COVID-19 (RECOVERY): a randomised, controlled, open-label, platform trial and updated meta-analysis
United Kingdomto
Argentina, 
Brazil, 
Denmark10 moreWithout the system 8 days (95% CI, 7 to 9)
With it 7 days (95% confidence interval [CI], 6 to 8)
Challenged
FDA approved the COVID-19 indication on 10 May 2022. In Europe, Lilly withdrew its application on 7 December 2022 after EMA reached the provisional opinion that the benefit-risk balance was negative.
Trade-offs Serious adverse events affected 15% of baricitinib recipients and 18% of controls in COV-BARRIER. RECOVERY investigators reported 12 suspected serious reactions. Short follow-up leaves rare and long-term harms uncertain.
Limits The exact software release, a comparison without AI, the observed savings and the number of people helped worldwide are absent from the sources.
Tried again It happened again in both other settings.
Needs The clinical comparisons concern hospitalized patients and the eligibility criteria of each trial. Background treatment differed: remdesivir in ACTT-2, and mostly corticosteroids in COV-BARRIER and RECOVERY.
Fewer deaths
In RECOVERY, 12% of patients receiving baricitinib died within 28 days, compared with 14% receiving usual care alone. [2]
Faster recovery
In ACTT-2, patients receiving remdesivir recovered in a median of 7 days with baricitinib, compared with 8 days with a placebo. [5]
-
Independently reported
Claim Remarks on the Rescue of United States Military Personnel in Iran and an Exchange With Reporters
Independent source 'Ghost Murmur,' a never-used secret tool, deployed to find lost airman in Iran in daring mission
United Statesto
IranSpecific harm prevented No documented comparison
One rescued airman, with the claimed AI contribution unverified
Challenged
Physicists interviewed by the Post and Scientific American disputed the public technical description. The claim about the rescue has not been formally reversed or withdrawn.
Trade-offs The rescue is documented, but the claimed heartbeat-sensing mechanism cannot be evaluated from the public evidence. The official reference to 40 miles concerns a camera observation.
Limits No measurement or technical account in the sources shows that the claimed system located the officer. Officials in their briefing and CBS in its reports name no tool, and CBS does not say that the technology used AI. Lockheed Martin declined to comment.
Two anonymous sources credit the CIA’s Ghost Murmur, quantum sensing paired with AI, with locating a downed F-15E officer in Iran. CBS reports that CIA technology pinpointed him. It names no tool.
-
Independently reported
Claim Use Crash Detection on iPhone or Apple Watch to call for help in an accident
Independent source 19-year-old rescued from car sinking in freezing water speaks out
Protective response, prevention unproven Assumed comparison
Persisted
WKRC reported on 11 March 2026 that the driver was alive and well and had thanked the Clermont County responders. ABC News (4 March) said that the accident report recorded only bruises.
Trade-offs The watch contributed an alert, and firefighters and deputies pulled her out. Clermont County dispatchers have not said whether the watch's call was the only alert or the earliest. Okayama University physicians found that 35 of 49 crash-detection calls to the Okayama City Fire Bureau from October 2022 to March 2023 were false alarms.
Limits Only the agencies' dispatch logs could separate the Apple Watch alert from the Life360 alert and show their timing. Reporters support the rescue and the minor injuries, but their accounts do not show whether the watch was needed, given the other alerts, the dispatchers and the responders.
Tried again It happened again in one other setting.
Needs The on-device crash classifier recognizes the impact as a severe car crash. The wearer does not dismiss the alert within the countdown.
On 22 February 2026 near Batavia, Ohio, a 19-year-old’s truck flipped into the Little Miami River. Clermont County officials say that her Apple Watch Crash Detection called 911. Crews freed her.
-
Independently reported
Claim Power of anti-drowning technology on display at Stirling Leisure - Inglewood in WA’s first AI-assisted pool rescue
Independent source AI pool monitoring system helps save life of swimmer suffering medical episode in the water
Israelto
AustraliaSpecific harm prevented Assumed comparison
Persisted
ABC News reported on 31 December 2025 that the system was still in use at the Inglewood pool more than a year after it started. ABC also reports that Royal Life Saving Australia is studying the system's impact in a research partnership with Lynxight.
Trade-offs Paul Salmon of the University of the Sunshine Coast asked how pools will prevent over-reliance on the technology and the loss of detection skills in lifeguards.
Limits Whether the man would have survived without the alert is unknown, because a fellow swimmer noticed the emergency independently and lifted him to the surface. The claim of a faster response comes from the city's own staff.
Tried again It happened again in one other setting.
Needs The system must detect distress and deliver an alert that lifeguards can act on promptly.
Lynxight, a camera system at a Perth pool, alerted a lifeguard’s smartwatch to a man who had sunk unconscious. Lifeguards rescued him with help from a swimmer. He recovered fully.
-
Independently reported
Claim Stroke Nurse Shares Her Experience After Brain Aneurysm
Independent source From caregiver to cared for: How an ICU nurse’s training saved her own life
Specific harm prevented Assumed comparison
Persisted
About 14 months after the hemorrhage, The Coastal Star reported that the nurse had returned to full-time work and her usual activities with no lasting effects of the stroke.
Limits Baptist Health has not confirmed which product, module or release flagged this patient's scan. No source isolates the software's effect on survival.
Software, apparently Viz.ai, detected a nurse’s brain bleed. It alerted her doctors. After treatment, the nurse returned to full-time nursing.
-
Independently reported
Claim Repurposing Portfolio
Independent source Doctors Told Him He Was Going to Die. Then A.I. Saved His Life.
Persisted
In its annual report for 2025, Every Cure states that Joseph remains in remission (page 15). The New Yorker also reported his remission in July 2025. In Penn Medicine’s documentary, published on 17 November 2025, Coates says that full remission followed five months after his transplant.
Editor’s note The clinician had already recognized the connection with myeloma. High scores supported that judgment, while remission followed further treatment.
Limits No detailed clinical case report, medical record or output of the model for this case is cited, and the authors of the Lancet Haematology vignette label its result unpublished data. Which algorithm or release scored the drugs for Coates is not known.
A consulting physician had considered three drugs for Joseph Coates’s POEMS syndrome. Every Cure’s early repurposing system scored all three highly. Coates responded. He later reached remission.
-
Independently reportedMagnitude 0.675
Magnitude 0.675
Claim Impact of a deep learning sepsis prediction model on quality of care and survival
Independent source Integrating artificial intelligence into healthcare systems: more than just the algorithm
Without the system 11.39% modeled counterfactual mortality
With it 9.49% observed post-deployment mortality
Persisted
The COMPOSER team reports that the alerts kept running in two emergency departments at UC San Diego Health from February 2023 through October 2025. A large language model layer was added to the alerts in July 2024.
Trade-offs Nurses received about 235 alerts a month across both sites. The sources show no evaluation of unnecessary treatment among patients without sepsis, or of its costs and harms.
Limits The authors did not measure harms from false alarms in patients without sepsis or financial returns, and the sources include no independent replication of the mortality effect.
Tried again It happened again in one other setting.
Needs Reliable data and prompt clinical assessment are likely requirements. Local evaluation is needed to establish whether the benefit recurs.
After UC San Diego Health deployed COMPOSER, a deep-learning model that alerts nurses to sepsis, mortality was 9.49%, against a modeled 11.39% without it.
-
Independently reported
Claim 510(k) Summary K172935: Embrace
Independent source This Seizure-Detecting Smartwatch Could Save Your Life
Specific harm prevented Measured comparison
Persisted
In an independent study at an epilepsy monitoring unit in Montreal, researchers later found that the Embrace2 detected all 16 convulsive seizures that occurred while it was worn and working. Alerts reached caregivers in 15 of 16 cases, with 29 false alarms over the study.
Trade-offs False alarms and broken connectivity could reduce the value of monitoring, and seizures without convulsions fell outside the demonstrated detection capability.
Limits The rescue rests on a single news account, and no clinical record of the event has been found. The same article covers a death despite a successful alert, which shows that a timely human response is a separate requirement.
Tried again It happened again in 2 of 3 other settings and did not in 1.
Needs Connectivity between watch, phone and caregiver (one of 16 alerts in Montreal failed on a Wi-Fi disconnection). Caregivers tolerate false alarms (29 in 373 monitoring days in Montreal, mostly from routine movements).
An Embrace wristband detected a seizure that stopped Morgan McGrath’s breathing. It alerted her mother, who called paramedics. IEEE Spectrum reports that McGrath, 26, of Fulton, Missouri, survived.
-
The deployer's account
Specific harm prevented Assumed comparison
Persisted
Mayo reported in January 2025 that two years after the stroke, the woman had returned to an active life. The woman was taking part in the DISCOVERY study of post-stroke cognitive impairment.
Trade-offs Mayo reports no costs or harms.
Limits Mayo does not name the software or compare her care with care without it, and it gives no measure of her treatment time.
Mayo Clinic says that unnamed AI algorithms let its care team find a woman’s ruptured aneurysm in minutes. After surgery, the woman recovered.
-
Confirmed by U.S. Food and Drug Administration
Official document Summary of Safety and Effectiveness Data (SSED), P230022: Jewel Patch Wearable Cardioverter Defibrillator (P-WCD)
Without the system Below 2.0 per 100 patient-months
With it 0.357 per 100 patient-months
Trade-offs Patients got inappropriate shocks at 0.357 per 100 patient-months. Seven of 305 patients had skin reactions severe enough to be withdrawn. There were 155 device-related adverse events, 25 needing medical treatment, though no serious ones. Only 30 percent of patients surveyed at the first visit reported overall satisfaction.
Limits The study was open-label with no control group, so no comparison with another defibrillator or with no device is possible. Only nine patients received an appropriate shock, and the Jewel's effect on survival rests on those eight conversions, with no outcome after the shocks reported. Patients wore the patch for weeks, not years.
In a study at 29 US sites that FDA reviewed, the Jewel patch defibrillator gave 0.357 inappropriate shocks per 100 patient-months. It converted 8 of 9 shockable rhythms with one shock.
-
Confirmed by OSTEBAMagnitude 0.525
Magnitude 0.525
Claim Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study: a randomised, controlled, non-inferiority, single-blinded, population-based, screening-accuracy trial
Independent source Artificial intelligence-supported double reading in European population breast cancer screening: A systematic review and meta-analysis of prospective programs
Netherlandsto
SwedenWithout the system 5.0 screen-detected cancers per 1,000 women
With it 6.4 screen-detected cancers per 1,000 women
Trade-offs HTA syd, the Region Skåne health technology assessment, found that AI in mammography screening brings no cost savings over double reading. Its model of the MASAI approach in Region Skåne added an estimated 3.3 million kronor a year for 133,775 women at an AI price of 30 kronor per woman.
Limits A 2026 review notes that MASAI is the only randomized trial to have reported, from one country with one AI system. The radiologists knew which arm each examination was in. Race and ethnicity were not collected. The lower interval cancer rate was not statistically significant. Whether the extra cancers found mean fewer deaths or more overdiagnosis is still in follow-up.
Tried again It happened again in all 3 other settings.
Needs Readers must have the AI scores and marks in front of them when they read. The AI must perform on the local equipment and population as it did in the trial.
In the MASAI trial of 105,934 women, Transpara-supported screening found 6.4 cancers per 1,000 against 5.0 with standard double reading. It cut screen reading by 44.2%.
-
Confirmed by HTA sydMagnitude 0.525
Magnitude 0.525
Claim Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study
Independent source Artificial intelligence in breast cancer screening: A systematic review and meta-analysis of integration strategies
South KoreatoSwedenWithout the system 250 screen-detected cancers
With it 269 screen-detected cancers with AI as a third reader
Persisted
Karolinska Institutet's news release says that since June 2023 Capio S:t Göran has read screening mammograms with one radiologist supported by the AI, so the system stayed in routine use after the study, though as a partner to one reader rather than as a third reader.
Trade-offs Triple reading by two radiologists plus Insight MMG raised recalls by 5 percent compared with two radiologists, so more women were called back and more went to consensus discussion without a cancer. The authors say that the extra radiologist workload, cost and worry for women must be weighed against the 19 added cancers.
Limits The ScreenTrustCAD study was single-arm and paired, so it cannot compare interval cancer rates between reading strategies. Its planned 23-month follow-up had not yet been done. It ran at one hospital with one mammography vendor and one AI system, which limits generalisability. Lunit, the maker of Insight MMG, was one of the funders. No quality assurance protocol existed to detect data drift in the AI's performance.
Tried again It happened again in both other settings.
Needs A consensus step must exist where AI-flagged cases are reviewed by radiologists. The AI threshold must be set to match radiologist sensitivity on local data.
In Stockholm, Lunit’s Insight MMG read 55,581 mammograms beside two radiologists, and the three together found 269 cancers against 250 by the radiologists alone.
-
Independently studied
Editor’s note Admission rhythm was normal and the device did not continuously record ECG.
Limits The authors do not name the watch's brand or model, and they give no dates for the events. The watch recorded an ECG and gave a generic warning about the heart rate. Clinicians identified the ventricular tachycardia. No source quantifies how much earlier treatment occurred.
Tried again It happened again in both other settings.
Needs The arrhythmia occurs while the wearer is able to trigger a recording (the watch has no continuous ECG). The QRS morphology on a single lead is clear enough for a clinician to suspect VT and pursue confirmation.
A smartwatch ECG function of an unnamed brand recorded a 165 bpm wide-complex tachycardia during palpitations. Cardiologists diagnosed ventricular tachycardia. They implanted a defibrillator.
-
Independently studied
United Statesto
JapanTrade-offs The watch kept a record of a low-oxygen episode outside the hospital, which normal spot readings had missed. The authors of the systematic review of 2023 report that readings from the Apple Watch can differ from medical-grade pulse oximeters by up to 15 points.
Limits This is a single clinical case, and clinicians made the diagnosis with computed tomography angiography. The authors do not describe how the watch's oxygen algorithm works, and they describe no automatic alert. They do not say who reported the reading. A systematic review of 2023 covered five studies of the Series 6 with 973 patients. Its authors found no strong systematic bias against medical-grade pulse oximeters, with outliers of up to 15 points.
Tried again It happened again in 1 of 2 other settings and did not in 1.
Needs The reading is not dismissed as artifact. Confirmatory testing (CT angiography, co-oximetry) is available.
The Blood Oxygen app on an Apple Watch Series 7 recorded an oxygen saturation of 88%, which led clinicians in Japan to computed tomography angiography that found a pulmonary embolism.
-
Independently studied
United Statesto
SwitzerlandPersisted
At the last follow-up, four weeks after the start of glucocorticoids, she reported that she was well and her resting heart rate and blood tests were normal. No later report on the patient was found.
Trade-offs No cost or harm of the tracker is reported in the sources. At the last follow-up she reported a weight gain of 2 kg, from fluid retention caused by the glucocorticoids.
Limits The patient noticed the trend, and clinicians did the diagnostic work. A reading from a smartwatch is not enough to show that an AI inference caused the benefit. One case cannot show that such screening works.
Tried again It happened again in one other setting.
Needs A stable personal baseline exists. The clinician treats the trend as a signal rather than noise.
A woman in Zurich saw in her Fitbit Alta HR app that her resting heart rate had risen from about 55 to about 75 beats per minute. The trend led to an early diagnosis of large vessel vasculitis.
-
Independently studied
Trade-offs After the lobectomy the man had a small prolonged air leak and left hospital on the tenth day. The authors warn clinicians not to trust fully a device made outside healthcare standards. They cite a study in which only 153 of 2,161 smartwatch alerts of an irregular rhythm led to a confirmed arrhythmia.
Limits The watch's brand and algorithm are unspecified, so whether AI was involved remains open. The lung cancer was found on a chest X-ray after the procedure, and the watch did not detect it. The claim that surgery would otherwise have become impossible is a conjecture with no follow-up evidence.
Tried again It happened again in one other setting.
Needs The arrhythmia is one the watch records while hospital monitoring misses it. The incidental finding is in the field of the post-procedure X-ray and caught early enough for curative surgery. The authors cite an estimated 1 in 25 cancers diagnosed incidentally.
A smartwatch of unstated brand alerted a 58-year-old man to an irregular heart rate above 180 beats per minute. In the care that followed, doctors found a lung adenocarcinoma. Surgeons removed it.
-
Independently studied
Trade-offs The surgeons say that surrounding issues must be considered in implementation, and they do not name them. They report no cost or harm in this case.
Limits Only the abstract is available. It reports one patient, with no age, sex, operation or outcome given. It does not say what the wristwatch measured or what it alerted on, and it does not name the device. There is no comparison and no follow-up.
Orthopedic surgeons at Toronto Western Hospital report that a patient’s smart wristwatch, of an unnamed brand, prompted readmission for a pulmonary embolism after surgery.
-
Independently studied
Claim 100,000 Genomes Pilot on Rare-Disease Diagnosis in Health Care - Preliminary Report
Independent source An optimized variant prioritization process for rare disease diagnostics
Expanded
Exomiser was run again across the full 100,000 Genomes Project. The run found diagnoses for 463 of 24,015 patients who had been left unsolved by the earlier analysis, after manual interpretation of the ranked candidates.
Trade-offs The Exomiser team and Genomics England report no costs or harms, and the record has no comparable incident.
Limits The pilot's diagnostic yield and clinical actionability apply to the combined workflow of sequencing, panels, prioritization and clinical review. They cannot all be attributed to Exomiser. The authors do not say whether Exomiser or the applied panels found the variants in the two patient cases that they describe. Changes in clinical decisions do not show better health or survival across the pilot.
Tried again It happened again in both other settings.
Needs Genome or exome sequencing with family data where available. Structured HPO phenotype terms of adequate quality.
In a pilot of England’s National Health Service, Exomiser ranked candidate variants for 2,183 families. A quarter of probands got a genetic diagnosis. A quarter of those were actionable at once.
-
Independently reported
Specific harm prevented No documented comparison
Limits The account is one patient's and his clinic's. No source compares the reading with and without the tool.
Diario de Mallorca reports that QP Prostate marked a tiny lesion on a 58-year-old man’s MRI at Clínica Girona. A biopsy confirmed cancer, and after surgery he says he is free of disease.
-
Independently reported
Claim How Facebook AI Helps Suicide Prevention
Independent source When every minute counts: Uttar Pradesh Police save youth attempting suicide in 8 minutes after Meta alert over Instagram post
United Statesto
IndiaSpecific harm prevented No documented comparison
Within eight minutesA Sardhana police team reached the man's home.
Persisted
The Times of India reported on 22 October 2026 that Uttar Pradesh Police credit the alerts with more than 4,000 lives saved from 1 January 2023 to 10 October 2026. The figure comes from the police, and no outside audit of it is known.
Trade-offs Decode reported a false alarm in Kanpur Dehat, where police traced a college student who said that she had posted a video with tablets on Instagram to get views. Neither Meta nor the Uttar Pradesh Police has published an assessment of other costs or harms to the people whom police visit.
Limits The police have not said whether Meta's classifier or a user's report flagged this post. The police are also the only source of the statewide figures of 3,011 and more than 4,000 lives saved, and no outside audit of them is known. Meta publishes no data on false alarms in India.
Tried again It happened again in 6 of 7 other settings and did not in 1.
Needs Account details resolve to a reachable address within minutes. Local officers are available to respond immediately.
After Meta alerted police to a 25-year-old man’s video on Instagram, officers in Meerut reached his home within eight minutes. His condition improved after treatment in hospital.
-
Independently reported
Claim Meta Flagged A Suicide Post. Police Had Minutes To Save The Man
Independent source Teen's life saved in 12 minutes: UP Police act swiftly after social media suicide alert
United StatestoIndiaSpecific harm prevented Measured comparison
Within 12 minutesA police team reached the man's home.
Persisted
The Times of India reported on 22 October 2026 that UP Police say that they saved more than 4,000 lives from 1 January 2023 to 10 October 2026 by acting on suicide alerts linked to social media posts. The figure is the police's own count, and no outside audit of it has been published.
Editor’s note The mechanism depends on a visible post and a rapid response. A stable condition after treatment is a narrower result than long-term recovery.
Limits Reporters establish that a Meta alert led police to his home within 12 minutes and that he was treated in hospital. They do not say whether software or a user report started this alert. The reading of the case as a life saved rests on the police account, and the sources contain no medical assessment of what would have happened without the alert.
Tried again It happened again in both other settings.
Needs The post is flagged and reviewed while the person can still be treated. Here the alert came at about 8:58 pm and police arrived within 12 minutes. The account leads to a phone number and a reachable address. Decode reports that UP police usually receive a phone number and trace the SIM and live location.
Meta’s system for detecting suicide risk flagged an 18-year-old’s Instagram post for review. Raebareli police reached him within 12 minutes. He was treated after taking rat poison.
-
Independently reported
The camera trace gave police his route and pictures for a public appeal, and a passer-by found him at about 3 a.m. and took him home
Expanded
On 29 June 2026 Hanoi police opened phase 2 of the command center. VnExpress reports that the city then had nearly 5,000 AI cameras, up from the 1,837 installed in December 2025. Hanoi police have reported no other lost person found through the system.
Trade-offs No assessment of costs or harms appears in the police accounts. Because the vendor and the matching method are undisclosed, the trace cannot be checked beyond the police account.
Limits No source shows a link between the passer-by's discovery and the camera trace or the appeal on social media. The documented contribution is the reconstruction of his route, so the rescue itself cannot be credited to the system. The vendor and the matching method are undisclosed.
Hanoi police traced a lost 78-year-old on AI traffic cameras on 15 December 2025. A passer-by who found him at 3 a.m. the next day took him home.
-
Independently reported
Claim Try it! Last year, Grok saved the life of a man in Norway
Independent source Man credits Grok AI with saving his life after ER missed near-ruptured appendix
United Statesto
NorwayProtective response, prevention unproven Assumed comparison
Persisted
In a comment in August 2026 the man wrote that he was alive and could not complain. He wrote that the surgeons deserved the credit and that the AI had helped in part. He also wrote that the operation lasted six hours and that his appendix had ruptured. In December 2025 he had written that it was about to rupture.
Trade-offs The likely benefit was renewed medical attention after a visit that sent him home without imaging. Grok ranked a perforated ulcer first, and by the man's account the cause was the atypical appendicitis that it ranked second. The man's account does not show that the scan depended on the chatbot's wording. Pediatricians in Istanbul reported in February 2026 that Grok 3 over-triaged 36.3% of cases in a children's emergency department.
Limits The story after the operation is self-reported, and Grok drafted the Reddit post. The post and the later comments differ on where the first visit was and on how he got back. In August 2026 he wrote that his appendix had ruptured, and in December 2025 he had written that it was about to rupture. No clinician or hospital has confirmed the CT findings or the surgery, and no record is public.
A Norwegian man had been sent home with an acid blocker. The advice of xAI’s Grok was to demand a computed tomography scan. His appendix was removed on 3 December 2025.
-
Independently reported
Claim Alert train drivers 'avert collision' with herd of elephants in Assam
Independent source Assam: AI Technology Generated Alerts Prevent Train From Colliding With Elephants
Protective response, prevention unproven Measured comparison
A herd of about 60 wild elephants
Expanded
By December 2025 the system covered 141 route kilometers of Northeast Frontier Railway and tenders had been awarded for 981 more route kilometers nationwide.
Trade-offs No cost or harm from the stop is on record. The system covers only part of the network: On 20 December 2025 a Rajdhani Express killed seven elephants at Sangjurai, where the railway said that no such system was installed.
Limits The crew saw the elephants and also received an alert. The railway has not shown that the AI warning caused the stop or that a collision would otherwise have occurred. The railway counts elephants saved without saying how many elephants were near trains, so its counts cannot be compared across periods or sections.
Tried again It happened again in one other setting.
Needs Train speed and sighting distance allow a stop after the alert. Crew act on the alert. The operator attributes this stop partly to the crew's own sighting.
On 16 October 2024 Northeast Frontier Railway’s AI-based Intrusion Detection System alerted the crew of the Kamrup Express, who also saw about 60 elephants crossing in Assam. The crew braked.
-
Independently reported
Claim HSI, partners launch first US-based international victim identification surge
Independent source Homeland Security identifies 311 child victims of sexual exploitation in 'cold cases'
311 probable identifications of previously unknown victims, 14 positive contacts
Expanded
HSI repeated the operation each year. ICE reports that Operation Renewed Hope III (24 February to 7 March 2025) positively identified 36 victims. ICE calls Operation Renewed Hope IV (2 to 13 March 2026) the largest such operation ever conducted. It produced more than 100 positive identifications. ICE counts more than 600 children identified and rescued across the four operations. These figures come from the deployer.
Trade-offs The likely contribution was narrowing a difficult search. Mistaken facial matches remained a reason to verify identities before investigators contacted families or suspects. Crikey reports that Australian case material was analyzed with Clearview AI, which the Australian Federal Police was not allowed to use.
Limits The 311 identifications are probable, and ICE gives no number of confirmed rescues. In its release, ICE does not mention AI. Homeland Security Investigations has not confirmed any product. No figure separates identifications that came from AI from those made with other techniques.
Tried again It happened again in 3 of 4 other settings.
Needs Legal authority to use web-scraped facial recognition (banned for this use in Australia, per ABC). Human investigators to verify tentative identifications before action.
U.S.-led investigators credited facial recognition, reportedly Clearview AI, for leads in a 2023 operation that made 311 probable identifications of child abuse victims. AI image enhancement helped.
-
Independently reportedMagnitude 0.20
Magnitude 0.20
Claim Use Case 27: Multi-Cancer Early Screening with Non-contrast CT and Deep Learning.
Independent source In China, A.I. Is Finding Deadly Tumors That Doctors Might Miss
Expanded
Xinhua reported in January 2026 that DAMO PANDA had been deployed in urban tertiary hospitals and in remote county facilities in Zhejiang, Yunnan, Hunan and Shanxi provinces.
Trade-offs In the first real-world cohort, PANDA gave false-positive findings for 76 patients (0.5%), and radiologists easily ruled out 70 of them. PANDA missed 28 lesions, mostly benign cysts but also one neuroendocrine tumor. A commentator warns that false positives may lead to unnecessary procedures and higher costs, and that training mostly on East Asian data raises generalizability concerns.
Limits The real-world test was retrospective. No survival gain has been reported. The authors of an independent 2026 review report that no study shows that the use of PANDA improves stage at diagnosis or survival.
Alibaba DAMO’s PANDA model exceeded the mean sensitivity of radiologists for pancreatic cancer on CT scans by 34.1%. It found five cancers that the initial standard of care had missed.
-
Independently reported
Claim Bakan Soylu'dan polis aracında bulunan patlayıcıyla ilgili flaş açıklama: Hedef İstanbul'du
Independent source Görüntüler ortaya çıktı: Aracında 52 kilo patlayıcı çıkan polis tutuklandı, Valilik açıklama yaptı
Specific harm prevented No documented comparison
133 packets, 52.5 kg, of factory-made explosive
Taken to court
On 20 November 2023 the Hakkari 2nd Heavy Penal Court sentenced the police officer A.K. to 21 years in prison. It sentenced the four co-defendants to 17 years 6 months each.
Trade-offs The police first suspected drugs and were wrong about the cargo, though officers found explosives. So the flag was a useful lead that pointed to the wrong kind of threat.
Limits How many attacks or casualties the seizure prevented is unknown. The police have not disclosed the model, its inputs or the part played by other intelligence, so what AI added remains uncertain.
The Turkish police’s ASENA analysis program flagged a car at Yüksekova on 20 March 2022. Officers found 52.5 kg of explosives. The interior minister said that they were bound for Istanbul.
-
Independently reported
Claim First Successful Rescue of a Lost Person Using the Human Detection System: A Case Study from Beskid Niski (SE Poland)
Independent source Bieszczady. Dron odnalazł zagubionego człowieka! Pierwszy raz w Polsce
Specific harm prevented Measured comparison
Expanded
The University of Wrocław hosts the developer. It reports that a team of mountain rescuers in South Tyrol used SARUAV to find a missing person at the beginning of August 2024. The person had died. The university calls this the second automatic detection in a real search, and it states that rescue services in Poland, other European countries and Asia have adopted the system. Only the developer's side reports this.
Trade-offs Two authors from the Main School of Fire Service in Warsaw wrote that each use of SARUAV needs spatial data prepared for the area. They also wrote that detection works only in the area that the recipient of the system has bought. The detection also required that enough of the man was visible from above in the photographed area.
Limits The authors document one successful detection and rescue. They do not measure how much sooner SARUAV found the man than a search without it would have. The authors include developers of SARUAV and a member of the Bieszczady Mountain Rescue Service. Researchers at Texas A&M wrote in 2023 that the makers of SARUAV give no details on its training data or the design of its model.
Tried again It happened again in one other setting.
Needs The person is within the imaged area and visible from above. Processing completes while the person can still be helped.
SARUAV is software that uses convolutional neural networks. On drone images it detected an unconscious 65-year-old man in tall grass after more than 24 hours. Mountain rescuers retrieved him alive.
-
Independently reported
United StatestoUnited KingdomProtective response, prevention unproven Assumed comparison
Trade-offs The Brentwood driver's wife had two black eyes and bruising, and the other car's passengers were sore from the seatbelts. Both cars were written off.
Limits One driver credited automatic braking, while the other was unsure whether it was enabled. Without vehicle logs or an engineering reconstruction, the drivers' accounts cannot show how the car braked or how many lives the braking saved.
Two Tesla Model X cars stopped as an oak fell across the A31 in Storm Dennis. All eight occupants walked away. One driver credits Autopilot’s emergency braking.
-
Independently reported
The child was recovered within weeks
Persisted
On 10 May 2024 Thorn announced that it was moving Spotlight to a new home led by Kristin Boorse, its former Vice President of Victim Identification. Canary NGO Ltd now builds the tool and says that more than 8,000 investigators use it. All later figures come from the tool's operators.
Trade-offs Spotlight matches faces with Amazon's facial recognition. WIRED notes accusations that police use of that technology endangers privacy and that the technology is inaccurate. The sources report no cost or harm from this recovery itself.
Limits Thorn's CEO reports the recovery, and no corroboration from police or courts has been found. Matching a photo to advertisements and recovering a child are separate steps. The aggregate identification and rescue counts cannot be substituted for each other or used to verify this case.
In April 2019 Thorn’s Spotlight, using Amazon facial recognition, matched a missing child’s photo to online sex ads. Thorn’s chief executive said that the child was recovered within weeks.
-
Independently reportedMagnitude 0.39
Magnitude 0.39
Claim Evaluation of the Recovery Engagement and Coordination for Health-Veterans Enhanced Treatment Suicide Risk Modeling Clinical Program in the Veterans Health Administration
Independent source Inside VA's yearslong AI effort to uncover veterans at high risk of suicide
Persisted
Nextgov reported in July 2025 that REACH VET still scans records monthly with the 61-variable model and that VA is updating it to a 2.0 model. In a cohort study published on 8 July 2025, VA researchers found no association between inclusion in REACH VET and suicide or all-cause mortality. That matches the mortality subanalyses of the 2021 evaluation.
Trade-offs The Fuller Project reported on 30 May 2024 that the model favors men and that officials left military sexual trauma out of its 61 variables. Nextgov reported in 2024 that research found that the model “considered being a white man more of an indicator of potential self-harm than other factors”. Nextgov also reported that a REACH VET 2.0 adding military sexual trauma and intimate partner violence as risk factors was planned for early 2025.
Limits The model triggered a package of clinician outreach and enhanced care, so the observed association is not an isolated effect of prediction. Historical and subthreshold comparisons require assumptions about trends and confounding. The retained null mortality findings must not be converted into a claim that deaths were prevented.
Tried again It was tried in 2 other settings, with no known outcome.
Needs Providers act on the list: care evaluations rose from 46.1% to 87.5% of identified veterans between March 2017 and December 2018. The model's variables fit the population. The original model under-weighted factors relevant to women veterans, per Nextgov.
The Veterans Affairs office that runs REACH VET studied the veterans that its suicide-risk model flagged. They had fewer suicide attempts. Their admissions fell. Deaths did not change.
-
The deployer's account
Specific harm prevented Assumed comparison
Published Retirement savings of more than $3 million
Trade-offs An alert alone would not have resolved the scam: The participant at first refused to believe that he had been deceived. Staff time and family contact were central to the reported outcome.
Limits Duckett credits AI with the first flag, and a portfolio manager, the fraud team and the participant's daughter with the successful intervention. Neither Duckett nor the firm has described the model's design or shown what would have happened without its flag. The amount at risk rests on the CEO's words and on Fortune's report that the participant tried to withdraw his entire $3 million portfolio. The sources include no transaction record to confirm it.
TIAA’s CEO said that an unnamed AI tool flagged a 76-year-old’s unusual withdrawal. Staff then kept his retirement savings of more than $3 million from moving.
-
The beneficiary's account
Protective response, prevention unproven Assumed comparison
Trade-offs Storyboard18 notes that Full Self-Driving (Supervised) still requires driver supervision. Tesla did not design it for an unconscious driver.
Limits No Tesla vehicle data, record from the police or emergency services, hospital statement, freeway location, software version or footage of the pull-over has been published. Tesla has not said whether it designed the stop as a fallback for an unresponsive driver.
Tried again It happened again in one other setting.
Needs FSD (Supervised) requires driver supervision, per Storyboard18. An unconscious driver is outside its intended operating condition. Whatever stopped the truck was a fallback, not a substitute for an attentive driver. The account is Vohra's alone. No Tesla vehicle data, police or emergency-service record, hospital statement, freeway location, FSD version or footage of the pull-over has been published. The replies from Elon Musk and Tesla's account acknowledge the outcome without confirming the mechanism.
Rishi Vohra wrote that Tesla’s Full Self-Driving pulled his Cybertruck onto a freeway shoulder after he passed out on 18 February 2026. No vehicle data has been published.
-
The makers' claim
Specific harm prevented Assumed comparison
Expanded
Daewoong completed thynC installations covering 558 beds at Chonnam National University Hospital and Hwasun Chonnam National University Hospital. It was the largest deployment at a national university hospital so far, among several hospital rollouts in 2026.
Editor’s note The alarm brought staff to a patient who was no longer breathing normally. Earlier recognition is the plausible mechanism, while the available account ends after resuscitation.
Limits The account comes through the supplier, with no clinical record and no response time that anyone independent has assessed. Staff resuscitated the patient, and the time that would have passed without the alert is unknown. A restored circulation leaves long-term recovery unknown.
On 14 January 2026 the thynC monitoring system flagged ventricular tachycardia in an 85-year-old on a general ward in Gwangju. Staff resuscitated the patient.
-
The deployer's account
Without the system the approximately $4.45 billion that was scheduled to be paid by Medicare
With it all but approximately $41 million
Published Held $4.41 billion in escrow from about $4.45 billion scheduled for payment
Taken to court
DOJ reported 35 people charged and 16 convicted by 14 August 2026, after a guilty plea to money laundering the previous day. Neither what AI contributed nor what finally happened to the $4.41 billion in escrow is known.
Trade-offs About $41 million reached the organization through Medicare, and supplemental insurers paid about $900 million. Enforcement costs and effects on legitimate payments are not quantified.
Limits The algorithm, what AI contributed, the operating costs, the rate of false positives and the final fate of the escrow remain unknown.
Justice Department analysts found anomalous Medicare billing through proactive data analytics. $4.41 billion was held in escrow. The software is unnamed. The department has not said that it used AI.
-
The deployer's account
Persisted
XRlabs reported that the twins had recovered and returned home. In the release, XRlabs gives no clinical follow-up interval or independent assessment of what the AI contributed.
Editor’s note These were combined elements of one complex treatment, with no comparison that removed the AI component.
Limits No source compares the result with planning without AI, and no clinical evaluation separates the model's contribution from the surgical team's work. The reports differ on the twins' age and on the timing of the operations, and they give no exact operation date. XRlabs gives no clinical follow-up interval.
Tried again It happened again in one other setting.
Needs A surgical team able to act on AI-assisted plans and patient-specific implants. Imaging of sufficient quality to model the twins' shared vasculature and skull.
XRlabs says that its Surgical Intelligence platform designed skull implants for conjoined twins. The hospital credits AI planning in their successful separation in four staged procedures.
-
The beneficiary's account
Protective response, prevention unproven Assumed comparison
Challenged
NHTSA raised the owner's story with Tesla on 15 April 2024 and in writing on 14 May 2024. It said that Tesla's repost conflicted with Tesla's own message that drivers must stay in control. NHTSA published the email with its information request of 5 November 2024. NHTSA questioned the promotion of the story and did not rule on whether the drive took place. No hospital or press confirmation of the owner's account has appeared since.
Trade-offs A medical crisis can impair the supervision on which this driving system depends. NHTSA told Tesla that the company's repost of the story conflicted with the company's own message that drivers must stay in control.
Limits The story rests on the driver's post and its republication. The sources include no vehicle logs from Tesla, no confirmation from the hospital and no comparison of the drive with emergency transport. Tesla and Musk reposted the story without verifying it.
Tried again It happened again in both other settings.
Needs FSD (Supervised) requires an attentive driver with hands on the wheel and eyes on the road (Hindustan Times; Tesla Oracle quoting owner James Locke). A driver too impaired to supervise would be outside its intended use. The account is the owner's alone. No hospital, emergency-service, police or vehicle-log record of the drive has been found. The photos attached to the post do not show the route, a timestamp or the hospital.
By his own uncorroborated account on X, Tesla Full Self-Driving (Supervised) v12 drove its owner 13 miles to a VA emergency room at 2 a.m. on 2 April 2024.
-
The deployer's account
Persisted
The government's National Plan to End Homelessness featured the Maidstone and Kent program as a case study and reported results through January 2025. The department also funds a randomized trial of data-led outreach with Xantura in four other councils.
Editor’s note The reported comparison arose because some households could not receive support. Better prediction without additional response capacity can increase identified need without preventing additional homelessness.
Limits The less-than-1% versus 40% comparison sets households that received proactive support against a group left unsupported for lack of capacity, and allocation was not randomized. The councils and Xantura have published no sample sizes, comparable baseline risk or observation windows. The government does not state how it calculated its figures. With this design, the effect of AI targeting cannot be separated from the effect of the support itself.
Tried again It happened again in 1 of 2 other settings.
Needs The presenter states that protected characteristics are not included in the live prediction model. The source does not establish causal conditions for reproducing the prevention result.
Kent County Council reports that a Xantura predictive-analytics platform identified households at risk. Under 1% of those given proactive support became homeless, against 40% of an unsupported group.
No benefits match.