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 flagged 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.
- Software flags a nurse’s brain bleed, and she returns to full-time work
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, and she 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.
- Smartwatch ECG captures a 165 bpm tachycardia, and Emory diagnoses VT
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.
- 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.
- 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 and helps 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.
- Meta flags a suicide post, and Raebareli police arrive in 12 minutes
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, and rescuers save him
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.
- 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.
- 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, and a couple has 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 a father and daughter to a 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 and has 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.
- An Apple Watch records low oxygen, and doctors find a pulmonary embolism
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%.
- Smartwatch flags a racing heart, and follow-up care finds lung cancer
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.
- Fitbit flags a racing heart, and doctors treat ventricular tachycardia
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.
- 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.
- 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.
- 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 target 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.
Suggestions and corrections receive a public tracking number on GitHub. A GitHub account is required. Email remains available for private material.
Public submissions and their status. Peter H. Vartanian reviews contributions before publication.
The editor’s guide gives every rule. In short:
- A benefit names the AI system, the beneficiary and the good.
- Each source records who wrote it: the makers, the deployer, an evaluator or someone independent.
- The grade (found, studied, reported or claimed) comes from the strongest independent source that credits the system.
- Depth runs from 1, time or money saved, to 5, a life or livelihood preserved.
- Counts describe what the database covers, not how beneficial AI is.
Briefs by Peter H. Vartanian examine the evidence and its gaps.
Not admitted
Excluded candidates and their reasons. An unresolved AI contribution can warrant further evidence. Intended benefits, aggregate results, and weak evidence can qualify.
Duplicate of a benefit
- Face2Gene, consumer wearables, speech neuroprostheses and the brain-spine interface AIBD-0037
These systems were not searched again because the database already records them. Face2Gene is № 0037, the Apple Watch and other wearables are № 0021 to № 0027, the speech neuroprostheses are № 0002 to № 0005, and the brain-spine interface is № 0001. A new entry would repeat those records.
- OECD AI Monitor entry on Casey Harrell's UC Davis speech neuroprosthesis AIBD-0002
This OECD AI Monitor entry covers the press coverage of June 2026 (Washington Post, 15 June; MIT Technology Review, 16 June) on the BrainGate2 speech neuroprosthesis from UC Davis that Casey Harrell uses. That case is already № 0002, and its record cites the NEJM paper of 2024 and the Nature Medicine paper of 2026 on long-term use.
- OECD AI Monitor entry on Meta's self-harm alert and Meerut police in eight minutes AIBD-0010
This OECD AI Monitor entry describes the same event as № 0010. Meta alerted the Social Media Center of the UP Police at 6:53 pm on 7 June 2026, and Sardhana police reached the home of the 25-year-old within eight minutes. The entry draws on PTI reports carried by Rediff, ETV Bharat, The Hans India and LatestLY.
- OECD AI Monitor entry on Mia flagging Barbara's 6 mm cancer at NHS Grampian AIBD-0062
This OECD AI Monitor entry describes the same event as № 0062: Mia flagged a 6 mm cancer that NHS Grampian radiologists had missed during the GEMINI evaluation. The entry draws on the BBC report of 21 March 2024 and a University of Aberdeen release. The Aberdeen release of 10 March 2026 has since been added to the evidence for № 0062.
- OECD AI Monitor entry on TrailGuard AI and the 39 poachers arrested in Similipal AIBD-0102
This OECD AI Monitor entry describes the same arrests of 39 poachers in Similipal as № 0102. Its sources are The New Indian Express, the Times of India and PTI, dated 20 and 21 February 2026. These reports date the entry of the poachers to the night of Thursday, 19 February. The account of the Field Director gives the night of Wednesday, 18 February.
- OECD AI Monitor entries on Indian Railways AI elephant detection AIBD-0009
Ten OECD AI Monitor entries on elephant detection by Indian Railways, dated from March 2023 to February 2026, repeat two cases that the database already records. These are the intrusion-detection case in Assam, № 0009, and the thermal-camera case at Madukkarai, № 0101.
- OECD AI Monitor entry on Tesla Autopilot avoiding a wrong-way driver AIBD-0109
This OECD AI Monitor entry, titled "Tesla Autopilot Prevents and Fails to Prevent Collisions: Contrasting Incidents," matches a case that is already recorded as № 0109. That case concerns a Tesla Model S and a wrong-way driver in Cabarrus County.
- OECD AI Monitor entry on Tesla Autopilot braking as an oak fell in Storm Dennis AIBD-0108
This OECD AI Monitor entry, titled "Tesla Autopilot Prevents Fatalities During Storm Dennis by Emergency Braking," is the case of a Tesla Model X and a falling oak tree. That case is already recorded as № 0108.
- Tesla Autopilot Involved in Both Crash and Harm Avoidance Incidents (OECD Monitor 2020-06-08-c7f5)
This OECD AI Monitor entry combines two events: the crash into an overturned truck in Taiwan, which is a harm listed as incident 218 in the AI Incident Database, and a video of a car avoiding a "pig." Reports in Electrek (8 June 2020) and the Daily Star (9 June 2020) show that the video is the swerve of Kenneth Van Hoesen around a wild boar in Belgium on 30 May 2020. That event is written up from a separate Monitor entry and has no case number yet.
- AI System Enables Major Criminal Gang Bust in Chile's 'Operación Influencer' (OECD Monitor 2023-05-12-99fe) AIBD-0233
№ 0233 already records Operación Influencer (11 May 2023), the Fiscal Heredia model and the ten detentions. That record also cites this OECD AI Monitor entry.
- Tesla Autopilot Saves Driver During Heart Attack Emergency (OECD Monitor 2024-04-16-ba4d) AIBD-0139
This is the same drive as № 0139: The owner's Tesla Model Y on FSD v12 took him to a VA emergency room on 2 April 2024. In the Bosnian and Croatian reports behind this entry (Raport.ba, 16 April 2024, citing a post by Elon Musk of 11 April), the story becomes a 20 km drive in Virginia during a heart attack. The record for № 0139 already notes these garbled syndicated versions.
- Police use facial recognition AI to catch county lines drug dealers (OECD Monitor 2024-07-02-ee42) AIBD-0232
№ 0232 already records the facial recognition match by Cheshire Constabulary to an 18-year-old, the stop with a 19-year-old on 17 May 2024 and the sentences of 1 July 2024. That record also cites this OECD AI Monitor entry.
- AI System in Japan Prevents Suicides at Train Stations and Commercial Buildings (OECD Monitor 2026-06-01-db16) AIBD-0234
№ 0234 already records the camera-analysis system by Asilla at about 40 stations and buildings in Japan and the two suicide interventions that Kyodo reported on 1 June 2026. This OECD AI Monitor entry covers the same event with the same sources: Kyodo, the Straits Times, Anadolu and syndicated copies.
- Paralyzed Man Controls Computer with Neuralink Chip, Spurs Safety and Privacy Worries (OECD Monitor 2025-03-23-f7ad) AIBD-0263
№ 0263 already records the Neuralink N1 implant that Noland Arbaugh received in January 2024 and his control of a cursor and a chess game by thought. This OECD AI Monitor entry covers the same person, system and event. The Monitor labels the entry an incident because of its material on malfunction and privacy, and that material describes no separate benefit.
No AI system took part
- Spelling by a completely locked-in ALS patient (Chaudhary et al. 2022)
The authors of the Nature Communications paper state that they did not implement a multichannel decoding algorithm. The patient selected letters through firing-rate thresholds with auditory neurofeedback. The only support-vector machine belonged to an earlier eye-tracking system, so no machine-learning system produced the benefit.
- ChatGPT as manuscript editor for a missed-pneumothorax case report (Brown 2023)
ChatGPT served only to draft and edit the Cureus paper about a delayed diagnosis of pneumothorax at Jersey General Hospital. The AI content of the paper is a proposal of how AI could have helped. No AI system had a role in the care of the patient.
- BirdCast and the lights-out schemes of Raleigh and Bird Safe Philly
Raleigh's Lights Out (27 August 2026) follows a fixed seasonal schedule from 11 p.m. to 6 a.m. and lists BirdCast only as a public resource. The Bird Safe Philly page has no mention of BirdCast alerts and no collision results, so no AI output changed either decision. Houston's Lights Out Nights is the case that BirdCast drove, and it is № 0103.
- Titian catheter-robot ablation at Shaoxing People's Hospital (OECD AI Monitor entry)
The only event in this OECD AI Monitor entry is a Titian robot ablation on a 64-year-old woman with paroxysmal atrial fibrillation at Shaoxing People's Hospital, reported by Xinhua Zhejiang on 13 December 2025. The surgeon steered the teleoperated catheter robot from a console, and the report has no mention of AI, autonomy or inference. The CA-GPT item in the entry is a comparative preprint with 96 patients and reports no event. The third source concerns an unrelated neurosurgical robot in Fushun.
- Wandercraft Atalante exoskeleton walking for people with complete paraplegia
Independent studies exist, including a 2026 trial with 11 participants at Nimes University Hospital in the Journal of NeuroEngineering and Rehabilitation. The exoskeleton walks on trajectories that nonlinear optimization computes in advance, and a hybrid zero dynamics controller tracks them. No machine learning or other AI is involved.
- Estonia's small-claims 'robot judge'
The Estonian Ministry of Justice stated on 16 February 2022 that it does not develop an AI robot judge to replace the human judge in small-claims or general court proceedings. There is no system and no outcome to verify.
- Rajesh Dandotiya bank-fraud intervention
Human bank staff made the intervention in the Rajesh Dandotiya case. No AI system was involved, so there is nothing to record.
- Singapore Police Force Project A.S.T.R.O. scam alerts
The Singapore Police Force says that Project ASTRO uses robotic process automation to send SMS alerts to scam victims (MHA Minister's Awards page, 29 October 2024). The Straits Times reported on 16 April 2023 that officers and OCBC staff analyzed fund flows and used the tool to send SMS messages in bulk. The S$442 million that police say has been averted since April 2023 is a total for automated messaging and does not measure model-based detection.
- AI-Enabled Halo System Prevents Serious Injury in Formula 1 Crash (OECD Monitor 2021-09-12-c119)
The Halo is a titanium bar over the cockpit that the FIA made compulsory from 2018. Neither Wikipedia nor the DER STANDARD report of 13 September 2021, in which Wolff says "Der Halo hat definitiv Lewis' Leben gerettet," has any mention of a sensor, software or an AI component. The sources that the Monitor lists are match reports that call the Halo a safety system. The label "AI-enabled" comes from the Monitor alone.
Nothing to write up
- AI embryo selection or AI newborn screening changing an identifiable patient's care
No peer-reviewed case report was found in which AI embryo selection or AI newborn screening changed the care of an identifiable patient. There was no system or case to examine.
- An individual patient helped by a sepsis early-warning alert
No individual case of a patient helped by a sepsis early-warning alert was found in a source independent of the developer. Searches returned only cohort studies and implementation stories. The cohort evidence for TREWS is № 0066.
- ChatGPT run after the fact on a necrotizing otitis externa case (Pugliese 2023)
ChatGPT-3 and ChatGPT-4 were run on the Milan case only after the diagnosis had been made. Neither model suggested the correct diagnosis, and neither had any part in the care of the patient. The result is a failed test of the models, so there is no benefit to record.
- A Garmin watch alert leading to a diagnosis
No peer-reviewed case report was found of a Garmin alert that led to a diagnosis. Bogár et al. (2024) searched for Garmin cases and found none, and case-report searches in Europe PMC returned no qualifying case.
- A WHOOP alert leading to a benefit
No peer-reviewed case report was found of a WHOOP alert that led to a benefit. Searches returned only validation studies and vendor blog posts.
- AI-detected fast radio bursts at CHIME in 2024
Searches returned only studies that reclassified the CHIME/FRB catalogs with machine learning. The searches found no AI-detected burst at CHIME in 2024 with independent confirmation, so there is no system or event to record. The Berkeley neural network that detected pulses from FRB 121102 is № 0018.
- An AI painting attribution accepted by a museum
No statement in which a museum accepts an AI-based attribution was located, and the candidate included no named system, painting or museum. The other half of the candidate, Ithaca, is № 0017. The one named attribution case, the de Brécy Tondo in Bradford, came up under another theme and is № 0276.
- AI identification of the language of the Voynich manuscript
No peer-reviewed acceptance was found of any AI identification of the language of the Voynich manuscript. No system, source or case was examined.
- Machine-learning attribution of a work to Bach
No machine-learning attribution of a work to Bach was found that is peer-reviewed and accepted by scholars. No system, source or case was examined.
- Sign-language recognition or translation deployed for deaf users
No deployed system for sign-language recognition or translation with an independently evaluated outcome for deaf users was found. The authors of a 2026 systematic review in JMIR on such systems for health-care communication with deaf people report that every included system was at the stage of development and testing. Authors whom they contacted confirmed that none had been deployed or commercialized.
- AI interception of a fraud or scam with a named victim's loss prevented
Searches returned only aggregate statistics from industry and regulators on authorized push payment fraud. No court or regulator record was found of a specific AI interception with a named victim, and the candidate included no named system, victim or event. The cases of AI fraud interception now in the database, such as the TIAA case (№ 0210), came from other candidates.
- An individual automatic emergency braking intervention that prevented a crash
No regulator, court or peer-reviewed record was found of an individual intervention by automatic emergency braking in which the crash did not occur. The NTSB documents that were located, on Uber in Tempe in 2018 and Ford BlueCruise in 2026, concern failures or disabled braking. The aggregate IIHS study, once № 0008, was removed because no source shows that the system learns from data. The Tesla incidents recorded from press reports (№ 0108 and № 0109) came from other candidates.
- UIC trial of AI trespass detection on railways
No UIC document naming an AI detection system or a trial result could be located. The safety pages on uic.org list the TreSP network, the TRESPAD events and project names such as AI4SAFEBEHAVE with no results, and the UIC page on trespass and suicide returned a 404 error. The only identifiable deployment is the one by ProRail, which is recorded separately as № 0309.
- An evacuation or emergency decision taken on an AI flood or nowcast forecast
No statement was found in which an authority documents an evacuation or emergency decision taken on an AI flood or nowcast forecast. The candidate included no named system, place or event of its own, and the cases that it pointed to are already № 0031 (flood alerts in Bihar) and № 0092 (skill evaluation in Mali). Records of action on AI forecasts from other candidates include the anticipatory cash payments in Kogi (№ 0095) and the activation by the Philippine Red Cross (№ 0098).
A harm, not a benefit
- Netherlands SyRI welfare-fraud risk scoring
SyRI is a documented harm. The District Court of The Hague invalidated the system in 2020, so it was set aside and was not examined as a benefit.
AI contribution to the benefit not established
- LGA-005 - Dorset council: the housing service digital transformation journey
In the procurement assessment of the council and the service documents of Huume, the housing improvements come from configurable workflows, policy calculations, portals and dashboards. None of the reviewed sources ties an identified AI function to those gains. Dorset's separate pilots of meeting summaries and transcription do not show that AI contributed to this housing case.
Data
The dataset is one file, records.json, released under CC BY 4.0. Three flat tables are derived from it: records.csv, one row per benefit; ratings.csv, one row per beneficiary group rated; and echoes.csv, one row per attempt to make a benefit happen again. The schema describes every field. Each benefit is also available on its own page as JSON.
Feed of new and revised benefits
Dataset editions are archived at Harvard Dataverse as numbered versions of one dataset, with a manifest of file checksums. Each benefit’s citation names its version and the edition that holds it.
Research tables
Measured outcomes: CSV or JSON. Taxonomy mappings: CSV or JSON. The research schema defines these tables.
Benefit profiles
Field definitions and symbols.
Download the public benefit profiles, field dictionary, and collection profiles. The coverage table lists each public field and its status for every benefit.
Fructus
AI outcomes, in money. A separately curated collection of financial claims, with original currencies, periods, recipients, comparisons, and supporting sources. JSON and CSV.
Counting
One system, one setting, one benefit is one benefit. A benefit reported in several sources is one benefit with several sources. A benefit that happened again elsewhere is an echo on the original benefit, not a new benefit.
Fields
- id
- A stable identifier.
- title, summary, date_observed, account
- What happened, in one line and in one paragraph, and when.
- brief
- A brief from Peter H. Vartanian: his reading of the evidence and its gaps, where one has been written, with its sidenotes in brief_notes.
- headline, metric
- A short title that adds no fact (asterisks around a name set it in italics), and the AI's quantified effect, copied verbatim from the summary or account (null when the benefit states no number).
- date_basis, date_precision
- Whether date_observed identifies an event, source publication, source information cutoff, or source access. Precision is day, month, year, or unassessed. A partial date uses the first day of its month or year only as a sorting anchor.
- benefit_status
- Reported, intended, partly reported (mixed) or unclear. This is separate from evidence confidence.
- category
- The kind of benefit: benefit created, harm prevented, protective response (prevention unproven), aggregate result (measured across many people or cases), or program (described as a whole, without a measured result).
- unexpected, purpose
- Whether the benefit was what the system was built to do, and what it was built to do.
- system, model_version, developer, deployer, location, domain, segment
- The system, its version where known, who made and ran it, and where it was used.
- sector
- One of 14 sectors, in the CSV. The domain and segment give the detail.
- benefit_type, taxonomy_note
- The benefit type from the taxonomy, with a note where the fit is loose.
- unrated_because
- For an event with no depth, the reason in a clause: the good is denied, or the AI's part is undocumented.
- ratings
- For each beneficiary group that benefited, a depth from 1 to 5 or null for unrated, and the reason.
- ai_role
- What the system contributed, assessed separately, or marked not assessed.
- learns_from_data
- Whether a source shows that the system learned from data how to produce its outputs: shown or not shown, with the source’s words and address. Empty when not yet checked.
- evidence, confidence
- The sources include their kind, citation, link, date, a short quote, and author relationship: maker, deployer, beneficiary, evaluator, independent author, or a relationship that remains unassessed. Only sources that support the benefit determine confidence (found, studied, reported, or claimed). Shade (high, medium, or low) corresponds to a solid, hatched, or outlined fill. The method explains these grades. An evaluator's role alone establishes neither independence nor a conflict of interest.
- comparator
- Better than what: the baseline the benefit is measured against (against), the figure without and with the AI (before, after), whether the source measured or assumed the difference (basis), and the passage that states it (quote, url). Null when no comparison is recorded. The benefit profile distinguishes an unassessed question from a reviewed absence.
- later
- What happened after the first report: persisted, scaled, challenged, reversed, withdrawn, litigated or unknown, with the date, a note, and the passage that states it. A challenge disputes a claim without establishing that it was reversed. Unknown means that no later outcome is established in the benefit.
- value
- A financial benefit stated by a source: the amount verbatim with its currency, what it measures, who gains or pays, whether it is an estimate, and the supporting passage.
- claim_status, claim_basis
- The strongest source's authority, the cited evidence by kind and author, and its limitations. Both are computed from the evidence.
- replication
- The conditions that produced the benefit, the conditions that must hold again, and the echoes (replications): each attempt to repeat it, with setting, model version, and outcome.
- benefit_categories, broad_benefits
- The broad benefits: one or two categories, the main one first. Ten take their names from the framework of SCSP and Johns Hopkins APL (2023) and four are the database’s own (knowledge discovery, culture and the arts, accessibility and animal health). The definitions are the database’s own.
- scsp_example
- The example in SCSP’s list that the benefit matches, by name, when one does.
- benefit_dimensions, iso_dimensions
- One to three of the eight kinds of benefit that ISO/IEC TR 21221:2025 names (functional, economic, environmental, social, societal, cultural, intellectual and personal), the main one first. The descriptions are the database’s own.
- size
- Reach, duration, distribution and frequency, each a band from 1 to 5 with the fact, the verbatim quotation and the source address that it rests on, or the reason that the sources give none. The CSV gives the bands.
- probability, magnitude
- Computed, never entered. Probability: 3 for a reported benefit and 2 for a partly reported one, one more for each of up to two other deployments where it recurred. Magnitude: SCSP’s benefit magnitude, the sum of scale (the deepest rating), reach, duration and distribution over 20 times the sum of probability and frequency over 10, given only when all six are documented.
- work_activities
- The O*NET work activities that earlier editions used as broad benefits, kept in the data. They come from the O*NET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA.
- source_collection, source_entry
- The imported collection and the stable entry key that AIBD assigns within it. These entry keys are not publisher-assigned identifiers.
- registry_ids
- Trial registration numbers, where a benefit rests on a registered trial.
- revisions
- Every change, with its date, source, and editor.
- deepest, groups (CSV)
- The highest depth across the beneficiary groups, and the groups rated, separated by semicolons.
- comparator_against, comparator_before, comparator_after, comparator_basis (CSV)
- The comparator above, flattened: the baseline, the figures without and with the AI, and whether the baseline was measured or assumed.
- later_status, later_date (CSV)
- The later outcome above and 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: the amount as stated, what it measures, and whether it is an estimate.
- sources (CSV)
- The number of distinct source addresses. Several quotations from one document count once, so it can be lower than the length of evidence in the JSON.
- updated (CSV)
- The date of the latest revision.
- The browse page’s CSV
- The benefits shown, in list order: ID, Benefit, Benefit status, Date (as shown, to its precision), Date basis, Countries, Sector, Benefit type, Depth, Evidence (the grade label), Sources, Reported measure and Link. Lists inside a cell use semicolons.
Colophon
Artificial Intelligence Benefits Database is made by Peter H. Vartanian. Write to peter.vartanian@icloud.com.
The dataset and this site are released under CC BY 4.0. Benefit types and beneficiary groups follow Trustible's benefits taxonomy.
The border adapts the color and grayscale reference strip shown beside a plate from E. A. Séguy’s Papillons in Smithsonian Libraries’ digitization.
Set in Newsreader, IBM Plex Sans and IBM Plex Mono, all used under the SIL Open Font License. Sector icons are Bootstrap Icons, under the MIT License. Organization logos are the organizations' own marks, shown to identify them; each one's source is listed in logos.json. Built 2 November 2026.
-
Fall Detection on an Apple Watch Series 10 called emergency services and alerted emergency contacts when a 69-year-old runner collapsed in cardiac arrest, and he made a full neurological recovery.
-
In a Taiwan trial of 15,965 patients, a neural network flagged high-risk electrocardiograms for physicians, and 90-day mortality was 3.6% versus 4.3% with usual care.
-
Drug found with a knowledge graph lowered COVID-19 deaths in a trial
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]
-
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 and names no tool.
-
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, and crews freed her.
-
Lynxight, a camera system at a Perth pool, alerted a lifeguard's smartwatch to a man who had sunk unconscious. Lifeguards and a swimmer rescued him, and he recovered fully.
-
Software, apparently Viz.ai, detected a nurse’s brain bleed and alerted her doctors. After treatment, the nurse returned to full-time nursing.
-
A consulting physician had considered three drugs for Joseph Coates’s POEMS syndrome, and Every Cure’s early repurposing system scored all three highly. Coates responded and later reached remission.
-
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.
-
An Embrace wristband detected a seizure that stopped Morgan McGrath's breathing and alerted her mother, who called paramedics. IEEE Spectrum reports that McGrath, 26, of Fulton, Missouri, survived.
-
In VA's trial across 140 medical centers, patients whose STORM risk scores triggered a mandated case review had 22% lower odds of death within four months. Serious adverse events did not change.
-
Mayo Clinic says that unnamed AI algorithms let its care team find a woman's ruptured aneurysm in minutes. The woman had surgery and recovered.
-
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.
-
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 and cut screen reading by 44.2%.
-
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.
-
The Veterans Affairs office that runs REACH VET studied the veterans that its suicide-risk model flagged. They had fewer suicide attempts and admissions and no change in deaths.
-
A smartwatch ECG function of an unnamed brand recorded a 165 bpm wide-complex tachycardia during palpitations, and cardiologists diagnosed ventricular tachycardia and implanted a defibrillator.
-
Across 463,094 women at 12 German screening sites, radiologists who used Vara MG found 6.7 cancers per 1,000 against 5.7 without it, with no rise in recalls.
-
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.
-
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.
-
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, and a quarter of those were actionable at once.
-
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.
-
Meta's system for detecting and reviewing suicide risk flagged an 18-year-old's Instagram post. Raebareli police reached him within 12 minutes, and he was treated after taking rat poison.
-
Hanoi police traced a lost 78-year-old on AI traffic cameras on 15 December 2025. A passer-by found him at 3 a.m. the next day and took him home.
-
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 and braked.
-
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.
-
U.S.-led investigators credited facial recognition, reported as Clearview AI, and AI image enhancement for leads in a 2023 operation that made 311 probable identifications of child abuse victims.
-
Alibaba DAMO's PANDA model exceeded mean radiologist sensitivity for pancreatic cancer on CT scans by 34.1% and found five cancers that the initial standard of care had missed.
-
The Turkish police's ASENA analysis program flagged a car at Yüksekova on 20 March 2022. Officers found 52.5 kg of explosives, which, the interior minister said, were bound for Istanbul.
-
SARUAV, software that uses convolutional neural networks, detected an unconscious 65-year-old man in tall grass on drone images after more than 24 hours, and mountain rescuers retrieved him alive.
-
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.
-
On 14 January 2026 the thynC monitoring system flagged ventricular tachycardia in an 85-year-old on a general ward in Gwangju, and staff resuscitated the patient.
-
Justice Department analysts found anomalous Medicare billing through proactive data analytics, and $4.41 billion was held in escrow. The department has not named the software or said that it used AI.
-
XRlabs says that its Surgical Intelligence platform designed skull implants for conjoined twins, and the hospital credits AI planning in their successful separation in four staged procedures.
-
Kent County Council reports that under 1% of households flagged by a Xantura predictive-analytics platform and given proactive support became homeless, against 40% of an unsupported group.
-
China Medical University Hospital’s Intelligent Anti-Microbial System flagged ceftazidime resistance in 1 hour, and doctors changed a 78-year-old woman’s antibiotic before she recovered.
-
Vyntra credits its AI risk models with blocking payments of USD 7,500 to CHF 1,000,000. It also credits them with stopping a USD 500,000 (KSh 58 million) ATM fraud that ended in arrests.
-
In a Colorado county's randomized trial with 4,681 children, those whose referrals were screened with an algorithm's risk score had 21 percent fewer injury admissions.
-
WeChat says that its WeChat Pay anti-fraud reminder put a Yunnan man in touch with customer service. An agent advised him to cancel a 190,000 yuan transfer to a “goddess” whom he had met online.
-
Researchers found that people hit by trains in suicidal acts fell from 1.17 to 0.50 a quarter at 14 Stockholm metro stations after SL installed AI-CCTV. Other stations saw no fall.
No benefits match.