{ "site": "https://becauseofai.org", "as_of": "2026-10-22", "counts": { "because": 16, "not-because": 3, "not-yet": 4 }, "held": [ "gb-0895-phase-3", "maryland-ai-tutor-negative" ], "entries": [ { "id": "ambient-scribes-nejm-ai", "section": "because", "rank": 15, "headline": "Because of AI, physicians using ambient scribes reported modestly less burnout, exhaustion, and task load in a 238-physician randomized trial.", "impact": "Modest burnout reduction (RCT)", "summary": "NEJM AI (Dec 2025) randomized 238 outpatient physicians across 14 specialties to Microsoft DAX, Nabla, or usual care for two months; either scribe improved burnout, exhaustion, and task load, and Nabla reduced time-in-note.", "caveat": "Effects were modest, only one vendor saved documentation time, occasional note inaccuracies required vigilance, and the trial ran Nov 2024–Jan 2025.", "domain": "work", "ai_type": "generative", "grades": { "evidence": "A", "attribution": "isolated", "window": "pre" }, "comparison": "Scribe access vs usual care, covariate-constrained randomization", "effect": { "relative": "Modest improvements on burnout, exhaustion, task load (see paper for effect sizes)", "absolute": null }, "reach": "238 physicians, one academic health system", "dates": { "achieved_start": "2024-11-04", "achieved_end": "2025-01-03", "published": "2025-11-26", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "NEJM AI: Ambient AI Scribes in Clinical Practice — A Randomized Trial (PMC)", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12768499/", "type": "primary" }, { "title": "JAMIA (May 2026): randomized crossover of two ambient scribes, 160 clinicians (Duke)", "url": "https://scholars.duke.edu/publication/1706148", "type": "primary" } ], "conflicts": "verify: vendor funding or in-kind support.", "what_would_change": "Patient-outcome or quality measures, not just clinician self-report.", "notes": "Window is pre for the NEJM AI trial; verify dates of the Duke crossover trial — it may be in-window.", "related": [ "ucla-chart-summarizer-rct" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "china-ai-cal-null", "section": "not-because", "rank": 4, "headline": "AI-personalized computer-assisted learning did no better than non-personalized software for 8,647 rural Chinese primary students.", "impact": "Null vs active control", "summary": "Stanford REAP and collaborators randomized 315 classes in 105 Sichuan schools to AI-personalized CAL, non-personalized CAL, or non-CAL activities; the AI arm showed no improvement in any ability tercile and may have been worse than plain CAL for the middle tercile.", "caveat": "One AI product, one province; the active-control design is exactly what the Sierra Leone trial lacked.", "domain": "education", "ai_type": "classic-ml", "grades": { "evidence": "A", "attribution": "isolated", "window": null }, "comparison": "AI-CAL vs non-personalized CAL (active control) vs pure control", "effect": { "relative": "No significant effect of AI-CAL on math scores", "absolute": null }, "reach": "8,647 students, 105 schools, Sichuan", "dates": { "achieved_start": null, "achieved_end": null, "published": "2026-07", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "SSRN: Putting AI to the Test — Evidence from a Large-Scale RCT in Rural China", "url": "https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7195260", "type": "primary" }, { "title": "CESifo working paper version", "url": "https://www.ifo.de/DocDL/cesifo1_wp12837.pdf", "type": "primary" } ], "conflicts": "none found", "what_would_change": null, "notes": "verify: intervention dates to set window grade.", "related": [ "sierra-leone-gemini-math" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "columbia-star-birth", "section": "because", "rank": 3, "headline": "Because of AI, a baby girl was born to a couple who had been told there was no sperm to work with.", "impact": "2 sperm found in 2.5M images", "summary": "Columbia's STAR system scanned 2.5 million images of a 3.5 mL sample in about two hours and found viable sperm where manual microscopy found none; the resulting pregnancy was confirmed in 2025, published in The Lancet in October 2025, and a healthy daughter was born in late 2025.", "caveat": "One case; STAR finds sperm in just under 30% of men diagnosed with azoospermia, and larger trials are still underway.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "D", "attribution": "only-path", "window": "in" }, "comparison": "Manual microscopy by trained technicians on the same sample (found none) vs STAR", "effect": { "relative": "~40x more sperm recovered than manual search across cases reported by the team", "absolute": "1 live birth; sperm found in ~30% of azoospermia patients tested" }, "reach": "Columbia University Fertility Center only", "dates": { "achieved_start": "2025-03", "achieved_end": "2025-12", "published": "2025-10-31", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Columbia / The Lancet: First clinical pregnancy following AI-based microfluidic sperm detection (press release)", "url": "https://www.eurekalert.org/news-releases/1104028", "type": "primary" }, { "title": "National Geographic: The first AI-assisted IVF babies are now being born", "url": "https://nationalgeographic.com/health/article/ivf-ai-assisted-reproductive-technology", "type": "secondary" }, { "title": "Decrypt/BBC: Doctors use AI to spot hidden sperm (birth confirmed)", "url": "https://decrypt.co/366374/doctors-ai-hidden-sperm-men-study", "type": "secondary" } ], "conflicts": "Two authors are inventors on Columbia patent applications for STAR.", "what_would_change": "A multi-center series with live-birth rates would move evidence to B or C.", "notes": "verify: reports differ on which couple was the first (a 19-year infertility case in the Lancet letter vs a Klinefelter case in BBC coverage); confirm birth month.", "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "ethiopia-tb-ai-cxr", "section": "because", "rank": 4, "headline": "Because of AI, 1,660 TB cases were found in Ethiopia between January 2025 and March 2026 in screening programs that had no radiologist to read the X-rays.", "impact": "1,660 TB cases found", "summary": "AI-read chest X-rays (CAD) served as the triage reader in MSH's Eliminate TB Project across four regions, contributing to 1,660 detected cases including asymptomatic infectious cases that symptom screening misses.", "caveat": "Program report, not a controlled comparison; \"contributed to\" means the AI flagged people for confirmatory testing, so the counterfactual is how many would have been found by symptom screening alone.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "C", "attribution": "only-path", "window": "ongoing" }, "comparison": "AI CXR triage vs routine symptom screening in the same facilities (descriptive)", "effect": { "relative": "Community AI screening yield >5% vs <0.1% routine clinical screening (program figures)", "absolute": "1,660 cases, Jan 2025–Mar 2026" }, "reach": "Four project-supported regions, Ethiopia", "dates": { "achieved_start": "2025-01", "achieved_end": "2026-03", "published": "2026-07", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "MSH technical brief: AI-powered chest X-ray screening for TB case finding, Ethiopia", "url": "https://msh.org/wp-content/uploads/2026/07/Eliminate-TB-Project_Technical-Brief_AI-Chest-X-Rays.pdf", "type": "primary" }, { "title": "MSH: AI-powered chest X-ray screening — field experience from Ethiopia", "url": "https://msh.org/resources/ai-powered-chest-x-ray-screening-for-tb-case-finding-field-experience-from-ethiopia/", "type": "secondary" } ], "conflicts": "Implementing NGO's own report (USAID-funded project).", "what_would_change": "A stepped-wedge or cluster comparison of facilities with and without CAD would move evidence to A/B.", "notes": null, "related": [ "india-tribal-tb-ai-cxr" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "every-cure-repurposing", "section": "not-yet", "rank": 3, "headline": "Every Cure's AI ranks 75 million drug–disease pairs; its patient wins so far came from clinicians or predate the window.", "impact": "$76M ARPA-H; 10 programs", "summary": "The nonprofit reviewed 9,000+ repurposing opportunities in 2025 and has ten active programs; the AI-identified adalimumab–Castleman case (NEJM) was February 2025, and the DFMO–Bachmann-Bupp program (six treated, five children improved) originated with the researchers who identified the disease.", "caveat": "No in-window patient outcome attributable to an AI-identified match yet.", "domain": "medicine", "ai_type": "hybrid", "grades": { "evidence": null, "attribution": null, "window": null }, "comparison": null, "effect": null, "reach": null, "dates": { "achieved_start": null, "achieved_end": null, "published": "2026-02-26", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Every Cure: up to $76M from ARPA-H", "url": "https://everycure.org/phase2/", "type": "primary" }, { "title": "CNBC: Every Cure's drug repurposing strategy", "url": "https://www.cnbc.com/2026/03/27/every-cure-drug-repurposing-rare-disease-treatment.html", "type": "secondary" } ], "conflicts": "Nonprofit self-reporting.", "what_would_change": "A documented patient outcome where the drug–disease match originated from the MATRIX platform, with dates.", "notes": null, "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "fido-northumbrian-leaks", "section": "because", "rank": 13, "headline": "Because of AI, Northumbrian Water stopped losing about 6.4 million litres of drinking water a day.", "impact": "~6.4 ML/day saved", "summary": "A six-month trial starting November 2025 placed FIDO Tech acoustic sensors on pipes, with AI classifying which underground sounds are leaks and directing crews; the utility reported early results of nearly 6.4 megalitres per day saved.", "caveat": "Operational report, no controlled comparison, and the saving is attributable to the detect-and-repair package; FIDO had already delivered a 37% leakage reduction across 15 target DMAs at the same utility in 2022–23.", "domain": "infrastructure", "ai_type": "classic-ml", "grades": { "evidence": "C", "attribution": "bundled", "window": "in" }, "comparison": "None (before/after reported by the utility)", "effect": { "relative": null, "absolute": "~6.4 ML/day (~25,000 bathtubs)" }, "reach": "Northumbrian Water network, North East England (4.4M customers)", "dates": { "achieved_start": "2025-11", "achieved_end": "2026-05", "published": "2026-06", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Northumbrian Water: AI technology helps to hunt down leaks", "url": "https://www.nwg.co.uk/news-and-media/news-releases/ai-technology-helps-to-hunt-down-leaks/", "type": "primary" }, { "title": "FIDO Direct / Northumbrian 37% reduction (2023)", "url": "https://www.wateronline.com/doc/fido-direct-launched-as-end-to-end-solution-to-solve-water-loss-0001", "type": "secondary" } ], "conflicts": "Utility and vendor joint announcement.", "what_would_change": "Leakage figures vs a matched set of DMAs using conventional detection in the same period.", "notes": null, "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "firesat-constellation", "section": "not-yet", "rank": 4, "headline": "Three FireSat satellites launched in July 2026 to spot 5×5 m fires from orbit; no fire-response outcome has been measured yet.", "impact": "3 satellites in orbit", "summary": "The Earth Fire Alliance / Google constellation's pilot already spotted small blazes invisible to existing satellites; the three operational satellites launched July 7, 2026, with 50+ planned by 2030.", "caveat": "Detection is not response; no published case yet of a fire contained earlier because of a FireSat alert.", "domain": "disaster", "ai_type": "classic-ml", "grades": { "evidence": null, "attribution": null, "window": null }, "comparison": null, "effect": null, "reach": null, "dates": { "achieved_start": null, "achieved_end": null, "published": "2026-07-07", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Google: Three new FireSat satellites launch", "url": "https://blog.google/innovation-and-ai/models-and-research/google-research/firesat-satellites/", "type": "primary" } ], "conflicts": "Google.org funds the constellation.", "what_would_change": "A documented detection with a timestamp ahead of other systems and a fire-agency response outcome.", "notes": null, "related": [ "hurricane-melissa-2025" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "flood-hub-flash-floods", "section": "not-yet", "rank": 5, "headline": "Google's urban flash-flood model now gives up to 24 hours' notice worldwide; whether anyone acted on an alert hasn't been measured.", "impact": "24h flash-flood alerts, global", "summary": "Launched March 2026, the model is trained on news-derived ground truth, with precision and recall in South America and Southeast Asia comparable to rich countries; Flood Hub covers 150+ countries.", "caveat": "Evaluated against flood records, not against outcomes for people.", "domain": "disaster", "ai_type": "classic-ml", "grades": { "evidence": null, "attribution": null, "window": null }, "comparison": null, "effect": null, "reach": null, "dates": { "achieved_start": null, "achieved_end": null, "published": "2026-03-09", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Google Research: Protecting cities with AI-driven flash flood forecasting", "url": "https://research.google/blog/protecting-cities-with-ai-driven-flash-flood-forecasting/", "type": "primary" }, { "title": "EarthArXiv preprint: AI expands urban flash flood forecasts globally", "url": "https://eartharxiv.org/repository/object/12082/download/21671/", "type": "primary" } ], "conflicts": "Google authored the evaluation.", "what_would_change": "An independent impact evaluation (evacuations, losses) tied to specific alerts.", "notes": null, "related": [ "hurricane-melissa-2025" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "fortell-ai-hearing-aids", "section": "because", "rank": 12, "headline": "Because of AI, 30 hearing-aid users understood far more words in heavy noise — according to the company that makes the device.", "impact": "31% → 80% words understood (company trial)", "summary": "Fortell's neural source-separation hearing aid (launched Dec 2025) reports a pre-registered double-blind 30-person trial against Phonak's Audéo Infinio Sphere showing a 9.2 dB vs 3.1 dB signal-to-noise improvement on QuickSIN, alongside a 20-person blinded preference study published in the International Journal of Audiology (May 2026).", "caveat": "Both studies are Fortell-funded with a Fortell employee as first author, the control device also uses neural-network noise reduction, and the 30-person intelligibility trial was posted on the company's site rather than confirmed in a journal as of October 2026.", "domain": "accessibility", "ai_type": "classic-ml", "grades": { "evidence": "C", "attribution": "ai-vs-ai", "window": "in" }, "comparison": "Fortell Spatial AI vs leading competitor AI hearing aid, within-subject crossover", "effect": { "relative": "9.2 dB vs 3.1 dB SNR improvement; ~19x odds of correct word recognition (company figures)", "absolute": "Company-reported: 69%→90%, 31%→80%, 7%→50% words correct across noise levels" }, "reach": "Prescription product, 30+ US states as of Oct 2026", "dates": { "achieved_start": "2025-11", "achieved_end": null, "published": "2026-05-11", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "IJA: Spatial AI consistently preferred to state-of-the-art hearing aids in multitalker noise", "url": "https://www.tandfonline.com/doi/full/10.1080/14992027.2026.2663345", "type": "primary" }, { "title": "Fortell evidence page (intelligibility trial)", "url": "https://www.fortell.com/evidence", "type": "primary" }, { "title": "Independent summary: what is actually known about Fortell", "url": "https://www.audiologyassociates.org/fortell-hearing-aids-what-we-know", "type": "secondary" } ], "conflicts": "Fortell Research funded both studies; first author is a Fortell employee.", "what_would_change": "Independent replication, or a comparison against a non-AI hearing aid, would move attribution and evidence.", "notes": "verify: a claimed Oct 8, 2026 IJA publication of the 30-person trial (PubMed 42850968) could not be confirmed.", "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "gb-0895-phase-3", "section": "not-yet", "rank": 2, "headline": "An AI-optimized antibody for severe asthma is in Phase 3 with ~1,600 patients — no AI-designed antibody has been approved.", "impact": "Phase 3 enrolling", "summary": "Generate:Biomedicines' GB-0895 started the SOLAIRIA-1 Phase 3 study in December 2025.", "caveat": "The antibody was optimized with AI, not designed from scratch; readout timing unannounced.", "domain": "medicine", "ai_type": "generative", "grades": { "evidence": null, "attribution": null, "window": null }, "comparison": null, "effect": null, "reach": null, "dates": { "achieved_start": null, "achieved_end": null, "published": "2025-12-03", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "IntuitionLabs: AI-discovered drugs in clinical trials 2026", "url": "https://intuitionlabs.ai/articles/ai-discovered-drugs-clinical-trials-2026", "type": "secondary" } ], "conflicts": "Sponsor announcements via secondary tracker.", "what_would_change": "Phase 3 primary endpoint result; then approval; then treated patients.", "notes": "verify: primary source (ClinicalTrials.gov entry for SOLAIRIA-1).", "related": [ "rentosertib-phase-3" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "golden-bridge-ii-stroke", "section": "because", "rank": 7, "headline": "Because of AI-assisted decision support, 10 fewer stroke patients per 1,000 had another stroke, heart attack, or vascular death within three months.", "impact": "26% fewer vascular events", "summary": "GOLDEN BRIDGE II (BMJ, March 2026) cluster-randomized 77 Chinese hospitals and 21,603 acute ischemic stroke patients; new vascular events at three months were 2.9% with the AI-imaging decision-support system vs 3.9% with usual care (adjusted HR 0.71), with the benefit holding through 12 months.", "caveat": "Hospitals, not patients, were randomized, and the intervention bundled AI image analysis with guideline-based treatment prompts, so the AI's isolated contribution isn't separable; enrollment ran January 2021 to June 2023.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "A", "attribution": "bundled", "window": "pre" }, "comparison": "Hospitals with AI-CDSS vs usual care; primary endpoint composite vascular events at 3 months", "effect": { "relative": "26% relative reduction (HR 0.71)", "absolute": "3.9% → 2.9%: 10 fewer events per 1,000 patients" }, "reach": "21,603 patients, 77 hospitals, 23 provinces", "dates": { "achieved_start": "2021-01", "achieved_end": "2023-06", "published": "2026-03-20", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "BMJ 2026;392:e085810 (press release with trial figures)", "url": "https://www.eurekalert.org/news-releases/1120567", "type": "primary" }, { "title": "ClinicalTrials.gov NCT04524624", "url": "https://clinicaltrials.gov/study/NCT04524624", "type": "primary" }, { "title": "Urban Health Today summary (enrollment dates)", "url": "https://www.urbanhealthtoday.com/post/use-of-ai-stroke-tool-leads-to-reduced-vascular-events-in-chinese-hospitals", "type": "secondary" } ], "conflicts": "verify: CDSS vendor and funding disclosures in the BMJ paper.", "what_would_change": "A factorial design separating AI imaging from guideline prompts would resolve attribution.", "notes": "Worked example on /method for the achieved-vs-published rule.", "related": [ "nhs-stroke-ai-imaging" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "hurricane-melissa-2025", "section": "because", "rank": 1, "headline": "Because of AI, the Hurricane Center called a Category 5 landfall in Jamaica about three days out, while Melissa was still a Category 1.", "impact": "~3 days' warning on a Cat 5", "summary": "NHC forecasters explicitly forecast Category 5 intensity while Melissa was an 80-mph Category 1, largely on the strength of Google DeepMind's cyclone model, which posted the lowest track and intensity errors of any model for the 2025 Atlantic season and kept the official forecast below its five-year average error while NOAA's GFS had its worst year in two decades.", "caveat": "There is no lives-saved number and cannot be; the measured outcome is forecast skill, and the final call was made by human forecasters using the model.", "domain": "disaster", "ai_type": "classic-ml", "grades": { "evidence": "B", "attribution": "isolated", "window": "in" }, "comparison": "DeepMind ensemble vs all other operational models and the NHC official forecast, full 2025 Atlantic season verification", "effect": { "relative": "Lowest season track and intensity error of any model; roughly one extra day of forecast accuracy in 2023–2025 hindcasts", "absolute": null }, "reach": "Jamaica (pop. ~2.8M), Haiti, Cuba, Bahamas; 2025 Atlantic season", "dates": { "achieved_start": "2025-10-25", "achieved_end": "2025-10-28", "published": "2025-11-29", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "NWS: AI hurricane forecasting Q&A (2025 season)", "url": "https://www.weather.gov/news/261102-AI-Hurricane-Forecasting", "type": "primary" }, { "title": "Michael Lowry: Two forecast models stand out (2025 season verification)", "url": "https://michaelrlowry.substack.com/p/this-hurricane-season-two-forecast", "type": "primary" }, { "title": "NPR: The future of hurricane forecasting is AI", "url": "https://www.npr.org/2025/11/29/nx-s1-5604290/atlantic-hurricane-season-forecast-ai-google", "type": "secondary" }, { "title": "CNN: Why NHC bet big on AI to forecast Isaias (Oct 2026)", "url": "https://www.cnn.com/2026/10/08/weather/hurricane-isaias-ai-models-google", "type": "secondary" }, { "title": "Gizmodo: WeatherNext Cyclones gives an extra day of warning", "url": "https://gizmodo.com/google-deepminds-hurricane-forecast-model-provides-a-whole-extra-day-of-warning-2000800635", "type": "secondary" } ], "conflicts": "Google partnered with NHC to train and evaluate the model; verification stats are from NHC and an independent former NHC branch chief.", "what_would_change": "A published NHC verification report quantifying the official forecast's error with and without the DeepMind input would move attribution from isolated toward a hard number.", "notes": "verify: exact NHC advisory time of the first explicit Cat 5 forecast; final Jamaica death toll and any evacuation-timing analysis.", "related": [ "flood-hub-flash-floods" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "india-tribal-tb-ai-cxr", "section": "because", "rank": 5, "headline": "Because of AI, TB case notifications rose 80% in a tribal district of Chhattisgarh, India, including 20 confirmed cases in people with no symptoms.", "impact": "+80% TB notifications", "summary": "A prospective multi-centre study of qXR v3 chest X-ray reading found an 80.21% increase in TB notifications versus the Aug 2022–Mar 2023 baseline; of 51 asymptomatic people the AI flagged, 20 were confirmed to have TB, and over 40% of all AI-flagged patients were confirmed.", "caveat": "Before-and-after design in a high-burden setting; national notifications were also rising from other case-finding efforts.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "B", "attribution": "only-path", "window": "pre" }, "comparison": "AI implementation period vs pre-implementation baseline (same sites)", "effect": { "relative": "+80.21% notifications (p<.001)", "absolute": "162 confirmed cases; 44.6% positivity among AI-flagged" }, "reach": "High-risk tribal population, Chhattisgarh", "dates": { "achieved_start": "2023-04", "achieved_end": null, "published": "2026-01-07", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Evaluating AI-based chest X-ray screening for TB in tribal Chhattisgarh (PMC)", "url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12810203/", "type": "primary" } ], "conflicts": "verify: author affiliations and vendor (Qure.ai) involvement.", "what_would_change": null, "notes": "verify: end date of the AI implementation period.", "related": [ "ethiopia-tb-ai-cxr" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "maryland-ai-tutor-negative", "section": "not-because", "rank": 3, "headline": "A GPT-4o tutor built into a university learning platform was associated with lower grades in a 2,379-student randomized trial.", "impact": "Lower grades (RCT)", "summary": "A University of Maryland working paper (Oct 2026) randomized 30 instructors in fall 2025 to a course-integrated GPT-4o tutor drawing on their own materials; students with the tutor performed worse.", "caveat": "Working paper, not yet peer-reviewed; effect sizes and mechanisms to verify.", "domain": "education", "ai_type": "generative", "grades": { "evidence": "A", "attribution": "isolated", "window": "in" }, "comparison": "Tutor access vs none, instructor-level randomization", "effect": { "relative": "Lower performance in the tutor arm (verify magnitude)", "absolute": null }, "reach": "2,379 undergraduates, 30 instructors", "dates": { "achieved_start": "2025-08", "achieved_end": "2025-12", "published": "2026-10", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "World Assessment Council summary of EdWorkingPaper 26-1598", "url": "https://worldassessmentcouncil.org/spotlight/ai-tutors-beat-good-teaching-and-cut-grades-design-decides/", "type": "secondary" } ], "conflicts": "none found", "what_would_change": null, "notes": "verify: locate the EdWorkingPaper itself and record the primary effect size; currently only a secondary source.", "related": [ "sierra-leone-gemini-math" ], "history": [ { "date": "2026-10-22", "note": "Added — needs primary source before it renders." } ] }, { "id": "masai-ai-mammography", "section": "because", "rank": 9, "headline": "Because of AI, Swedish mammography screening found 29% more cancers with 44% less radiologist reading time, and 12% fewer cancers surfaced between screens.", "impact": "29% more cancers, 44% less workload", "summary": "MASAI, the first RCT of AI in cancer screening (105,934 women; Lancet, Jan 2026), reported interval cancers of 1.55 vs 1.76 per 1,000 with AI-supported reading vs standard double reading, with fewer invasive, T2+, and non-luminal-A interval cancers, on top of earlier results showing higher detection and lower workload at the same specificity.", "caveat": "The trial was designed for non-inferiority; the 12% interval-cancer reduction was not statistically significant (ratio 0.88, 95% CI 0.65–1.18, p=0.41), mortality wasn't measured, and screening took place in 2021–2022.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "A", "attribution": "isolated", "window": "pre" }, "comparison": "AI-supported single/double reading vs standard double reading, individually randomized", "effect": { "relative": "12% fewer interval cancers (n.s.); 29% more screen-detected cancers; 44% workload reduction", "absolute": "1.55 vs 1.76 interval cancers per 1,000 (0.21 per 1,000 difference)" }, "reach": "105,934 women, Sweden national screening programme", "dates": { "achieved_start": "2021-04", "achieved_end": "2022-12", "published": "2026-01-29", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Lancet 2026;407:505-514 (ScienceDirect)", "url": "https://www.sciencedirect.com/science/article/abs/pii/S014067362502464X", "type": "primary" }, { "title": "ASCO Post summary with interval cancer rates", "url": "https://ascopost.com/news/february-2026/interval-cancer-rate-with-ai-supported-mammography-screening/", "type": "secondary" }, { "title": "Cancer (Wiley) summary", "url": "https://acsjournals.onlinelibrary.wiley.com/doi/10.1002/cncr.70440", "type": "secondary" } ], "conflicts": "ScreenPoint Medical (Transpara) provided the software and publicized results; academic sponsor Lund University.", "what_would_change": "Breast-cancer-specific mortality follow-up, or a replication powered for superiority.", "notes": "verify: exact screening enrollment dates from the Lancet methods.", "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "nhs-derm-skin-cancer", "section": "because", "rank": 6, "headline": "Because of AI, about a third of urgent skin-cancer referrals at some NHS hospitals got a same-day all-clear without waiting weeks for a dermatologist.", "impact": "25,000+ cancers found; ~1/3 discharged same day", "summary": "Skin Analytics' DERM, the only AI with Class III CE certification for autonomous use, has assessed 250,000+ NHS patients, identified 25,000+ histology-confirmed cancers, and is expanding to all 18 NHS South East trusts (about 9 million more people) from autumn 2026.", "caveat": "Operational data from the vendor and trusts; at one district hospital the autonomous pathway discharged 37% of referrals and missed one basal cell carcinoma.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "C", "attribution": "only-path", "window": "ongoing" }, "comparison": "Autonomous AI triage vs dermatologist face-to-face review (capacity-constrained usual care)", "effect": { "relative": "Discharge-without-appointment rate 26% with dermatologist second read → 37% autonomous (one hospital)", "absolute": "250,000+ assessed; 25,000+ cancers confirmed" }, "reach": "25 NHS trusts; +18 trusts from autumn 2026; Northern Ireland (1.9M)", "dates": { "achieved_start": "2022-01", "achieved_end": null, "published": "2026-10-09", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "NICE early value assessment of DERM (May 2025)", "url": "https://www.healthtechdigital.com/?p=237510", "type": "primary" }, { "title": "NIHR HTA: DERM and Moleanalyzer Pro clinical and cost-effectiveness (Feb 2026)", "url": "https://njl-admin.nihr.ac.uk/document/download/2049990", "type": "primary" }, { "title": "BJD abstract: implementing DERM at an NHS district hospital", "url": "https://www.citedrive.com/en/discovery/ai04-implementing-ai-derm-technology-into-the-urgent-skin-cancer-pathway-at-an-nhs-district-hospital-is-it-worth-it/", "type": "primary" }, { "title": "HLTH: NHS South East expands DERM across 18 trusts", "url": "https://hlth.com/insights/news/nhs-england-south-east-expands-skin-analytics-ai-skin-cancer-screening-across-18-trusts", "type": "secondary" } ], "conflicts": "Patient and cancer counts come from the vendor's monitoring data.", "what_would_change": "Published missed-cancer rates across all autonomous sites, or NICE's post-evidence-generation decision, would move evidence to B.", "notes": null, "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "nhs-stroke-ai-imaging", "section": "because", "rank": 11, "headline": "Because of AI, stroke patients in England were 57% more likely to get a thrombectomy and reached specialist centres 64 minutes faster.", "impact": "EVT odds 1.57x; 64 min faster transfer", "summary": "Across ~453,000 stroke admissions at 107 English hospitals over five years, patients whose imaging was reviewed with Brainomix 360 had higher thrombectomy odds (OR 1.57), median door-in-door-out of 128 vs 192 minutes, and a good outcome (mRS 0–2) at discharge of 47.1% vs 45.6% (OR 1.16).", "caveat": "Observational, not randomized; long-term outcomes couldn't be evaluated and reasons for non-use of the AI in some cases weren't captured.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "B", "attribution": "isolated", "window": "pre" }, "comparison": "Patients reviewed with AI software vs without, adjusted for age, severity, premorbid disability, timing", "effect": { "relative": "OR 1.57 for thrombectomy; OR 1.16 for good outcome", "absolute": "Good outcome 47.1% vs 45.6% (15 more per 1,000); DIDO 64 min shorter" }, "reach": "107 hospitals, England; patient-level data on 71,017 ischemic strokes", "dates": { "achieved_start": 2019, "achieved_end": 2024, "published": "2025-12", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Lancet Digital Health: AI imaging decision support for acute stroke in England", "url": "https://www.thelancet.com/journals/landig/article/PIIS2589-7500(25)00109-8/fulltext", "type": "primary" } ], "conflicts": "verify: Brainomix involvement in authorship or funding.", "what_would_change": null, "notes": "verify: exact observation period and publication date; window grade assumes the 5-year data period ended before Oct 2025.", "related": [ "golden-bridge-ii-stroke" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "oman-ai-retinopathy", "section": "because", "rank": 14, "headline": "Because of AI, Oman screened 43,000 diabetics for retinopathy, flagged 12,000 for treatment, and cut specialist waiting lists by 87%.", "impact": "Waitlists cut 87%", "summary": "Oman's national AI diabetic-retinopathy screening program reported 43,000+ patients screened, ~30% referred for specialist examination, an 87% reduction in eye-clinic waiting lists, and 43.2% lower annual operating costs.", "caveat": "Ministry-reported program figures with no comparison group; vision outcomes (sight saved) are not yet measured.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "C", "attribution": "bundled", "window": "in" }, "comparison": "None (program before/after)", "effect": { "relative": "−87% waiting list; −43.2% operating cost", "absolute": "43,000+ screened; 12,000+ referred" }, "reach": "Oman national program", "dates": { "achieved_start": null, "achieved_end": "2026-08", "published": "2026-08-24", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Muscat Daily: Oman turns to AI to tackle diabetes-related blindness", "url": "https://www.muscatdaily.com/2026/08/24/oman-turns-to-ai-to-tackle-diabetes-related-blindness/", "type": "primary" } ], "conflicts": "Ministry of Health self-report.", "what_would_change": "Rates of sight-threatening retinopathy treated in time, vs the prior program.", "notes": "verify: program start date; which AI vendor.", "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "rentosertib-phase-3", "section": "not-yet", "rank": 1, "headline": "The first generative-AI-discovered drug entered Phase III in September 2026 — no AI-discovered drug has yet been approved anywhere.", "impact": "Phase III started; 0 approvals", "summary": "Insilico dosed the first patient in GENESIS-IPF-3 (320 patients, 47 Chinese sites, 52-week FVC primary endpoint) for rentosertib in idiopathic pulmonary fibrosis after a Phase IIa signal published in Nature Medicine (2025).", "caveat": "Primary completion is estimated for late 2029; Phase IIa was 71 patients over 12 weeks.", "domain": "medicine", "ai_type": "generative", "grades": { "evidence": null, "attribution": null, "window": null }, "comparison": null, "effect": null, "reach": null, "dates": { "achieved_start": null, "achieved_end": null, "published": "2026-09-10", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Insilico: first patient dosed in GENESIS-IPF-3", "url": "https://insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3", "type": "primary" }, { "title": "ClinicalTrials.gov NCT07687459", "url": "https://clinicaltrials.gov/study/NCT07687459", "type": "primary" } ], "conflicts": "Sponsor announcements.", "what_would_change": "A positive 52-week FVC result, then regulatory approval, then patients treated outside a trial — each step moves it closer; approval alone does not make it a \"because\" entry.", "notes": null, "related": [ "gb-0895-phase-3" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "sierra-leone-gemini-math", "section": "because", "rank": 8, "headline": "Because of AI (plus tablets and pair work), junior secondary students in Sierra Leone gained about 0.26 standard deviations in math in eight weeks.", "impact": "+0.26 SD in 8 weeks", "summary": "A preregistered classroom-level cluster RCT of 1,763 Grade 7–8 students in 48 classrooms across 12 Port Loko schools had treatment classes use Gemini's Guided Learning in pairs for about two of four weekly math lessons; the treatment group gained +0.258 SD on independently administered tests.", "caveat": "The control arm got neither tablets nor structured pair work, so the package was tested, not the AI; results are a Google DeepMind/Fab AI technical report, not a peer-reviewed paper.", "domain": "education", "ai_type": "generative", "grades": { "evidence": "A", "attribution": "bundled", "window": "in" }, "comparison": "Gemini-integrated lessons vs business-as-usual instruction, classroom-level randomization", "effect": { "relative": "+0.258 SD (described as ~1.2–1.7 years of typical progress)", "absolute": null }, "reach": "1,763 students, 12 schools, Port Loko District", "dates": { "achieved_start": "2025-10", "achieved_end": "2025-12", "published": "2026-05", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Google DeepMind & Fab AI technical report (May 2026)", "url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf", "type": "primary" }, { "title": "AEA RCT Registry 16651", "url": "https://www.socialscienceregistry.org/trials/16651", "type": "primary" }, { "title": "DeepMind blog summary", "url": "https://deepmind.google/blog/measuring-the-impact-of-learning-with-ai-in-sierra-leone-and-beyond/", "type": "secondary" } ], "conflicts": "Google authored the report and makes the product.", "what_would_change": "An active-control arm (tablets + pair work, no AI) would isolate attribution — see the China CAL trial for why that matters.", "notes": "verify: exact intervention dates (registry lists Sept 2025–Mar 2026 for the study overall).", "related": [ "china-ai-cal-null", "maryland-ai-tutor-negative" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "tricorder-ai-stethoscope-null", "section": "not-because", "rank": 1, "headline": "An AI stethoscope rolled out to 205 NHS GP practices did not significantly increase heart-failure detection at 12 months.", "impact": "Null on intention-to-treat", "summary": "TRICORDER (Lancet, Jan 2026) cluster-randomized practices covering 1.5 million patients; on intention-to-treat, heart failure, AF, and valve-disease detection did not significantly rise (1,342 vs 1,984 new HF cases overall), though patients actually examined with the device had roughly twice the HF detection and three times the AF detection.", "caveat": "The tool worked when used; the deployment didn't get it used — a workflow failure, not a model failure.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "A", "attribution": "isolated", "window": "pre" }, "comparison": "Practices with AI stethoscope vs usual care, cluster-randomized, 12 months", "effect": { "relative": "HF IRR ~0.94 (n.s.); subgroup examined: ~2x HF, ~3x AF detection", "absolute": null }, "reach": "205 practices, North West London", "dates": { "achieved_start": 2023, "achieved_end": 2024, "published": "2026-01-29", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Lancet: TRICORDER cluster-randomised implementation trial", "url": "https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(25)02156-7/fulltext", "type": "primary" }, { "title": "Imperial College news summary", "url": "https://www.imperial.ac.uk/news/articles/medicine/nhli/2026/ai-enabled-stethoscopes-show-promise-for-improving-diagnosis-of-cardiovascular-conditions-uk-trial-finds/", "type": "secondary" } ], "conflicts": "Eko Health supplied devices and publicized subgroup results before the primary analysis.", "what_would_change": null, "notes": "verify: exact implementation period.", "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "ucd-als-speech-bci", "section": "because", "rank": 10, "headline": "Because of AI, a man with ALS has spoken with his family and kept working from home for two years through a brain implant, at up to 99% word accuracy.", "impact": "2 years of independent speech", "summary": "UC Davis's intracortical BCI uses deep-learning brain-to-text and cursor decoders with continuous fine-tuning; Nature Medicine (June 2026) documented Casey Harrell's near-daily unassisted use at home for roughly two years.", "caveat": "One participant, an experimental device implanted in 2023, not commercially available; some BCIs in other ALS patients have later stopped working.", "domain": "accessibility", "ai_type": "classic-ml", "grades": { "evidence": "D", "attribution": "only-path", "window": "ongoing" }, "comparison": "No alternative communication path at comparable speed existed for this participant", "effect": { "relative": null, "absolute": "Up to 99% word accuracy; ~2 years independent home use" }, "reach": "One participant", "dates": { "achieved_start": "2023-07", "achieved_end": null, "published": "2026-06-15", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Nature Medicine: Long-term independent use of an intracortical BCI for speech and cursor control", "url": "https://www.nature.com/articles/s41591-026-04414-6", "type": "primary" }, { "title": "UC Davis Health news release", "url": "https://health.ucdavis.edu/news/headlines/brain-computer-interface-enables-independent-accurate-communication-for-man-living-with-als/2026/06", "type": "secondary" } ], "conflicts": "Academic; verify device-maker relationships in the paper.", "what_would_change": "Multi-participant results with durability data would move evidence to C/B.", "notes": null, "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "ucla-chart-summarizer-rct", "section": "because", "rank": 16, "headline": "Because of AI, 284 UCLA clinicians reported modestly lower task load and burnout from a chart summarizer — while saving no time and mostly not recommending it.", "impact": "Lower task load, zero time saved", "summary": "A pragmatic RCT (Feb–May 2026) of Epic's generative chart summarization tool among 284 ambulatory clinicians found modestly reduced physician task load and lower burnout, no time savings, declining engagement, and a net promoter score of −22.", "caveat": "Preprint (medRxiv, Aug 2026), self-reported outcomes, three-month window, one health system.", "domain": "work", "ai_type": "generative", "grades": { "evidence": "A", "attribution": "isolated", "window": "in" }, "comparison": "Tool access vs usual pre-charting workflow, 1:1 randomization for 90 days", "effect": { "relative": "Modest task-load reduction (primary endpoint met)", "absolute": null }, "reach": "284 clinicians, UCLA Health", "dates": { "achieved_start": "2026-02", "achieved_end": "2026-05", "published": "2026-08-26", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "medRxiv: A pragmatic randomized trial of an EHR-integrated generative AI chart summarization tool", "url": "https://www.medrxiv.org/content/10.64898/2026.08.26.26361496.full.pdf", "type": "primary" }, { "title": "ClinicalTrials.gov NCT07438743", "url": "https://clinicaltrials.gov/study/NCT07438743", "type": "primary" } ], "conflicts": "Epic's tool evaluated by its customer; no vendor funding reported — verify.", "what_would_change": null, "notes": "Included as the honest floor for generative AI in clinical workflow.", "related": [ "ambient-scribes-nejm-ai" ], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "uva-comet-deterioration-null", "section": "not-because", "rank": 2, "headline": "Displaying AI deterioration-risk scores on a cardiology ward did not improve patient outcomes across 10,422 admissions.", "impact": "No change in primary outcome", "summary": "A cluster-randomized trial at UVA (Scientific Reports, Feb 2026) showed CoMET risk trajectories on screens for intervention beds; event-free hours did not differ, and clinicians moved 11% of patients into or out of display beds, undermining the design.", "caveat": "Passive display with no mandated response, run during COVID-era staffing; sicker patients were preferentially moved into AI-monitored beds.", "domain": "medicine", "ai_type": "classic-ml", "grades": { "evidence": "A", "attribution": "isolated", "window": "pre" }, "comparison": "Display-on vs display-off beds, cluster-randomized by room", "effect": { "relative": "No significant difference in event-free hours or 21-day mortality", "absolute": null }, "reach": "10,422 inpatient visits, 85 beds", "dates": { "achieved_start": "2021-01-04", "achieved_end": "2022-10-04", "published": "2026-02-05", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Scientific Reports: A randomized controlled trial of AI-based analytics for clinical deterioration", "url": "https://www.nature.com/articles/s41598-026-39051-z", "type": "primary" } ], "conflicts": "In-kind support from Nihon Kohden; two authors consult for the vendor.", "what_would_change": null, "notes": null, "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] }, { "id": "waymo-injury-crashes", "section": "because", "rank": 2, "headline": "Because of AI, there were an estimated 841 fewer injury-causing crashes across 270 million driverless miles.", "impact": "82% fewer injury crashes", "summary": "Waymo's safety hub through June 2026 reports 82% fewer injury-causing crashes and 95% fewer serious-or-worse crashes than a geographically matched human benchmark, with roughly 134 of the 841 avoided injury crashes accruing in Q2 2026 alone (841 minus the 707 reported through March 2026).", "caveat": "Company-produced, modeled comparison against an estimated human benchmark with an adjustment for roughly 32% underreporting of human injury crashes; not a randomized comparison of identical trips.", "domain": "transport", "ai_type": "classic-ml", "grades": { "evidence": "C", "attribution": "isolated", "window": "ongoing" }, "comparison": "Waymo rider-only miles vs human-driver crash rates in the same ZIP codes and period, adjusted for underreporting", "effect": { "relative": "82% fewer injury crashes; 95% fewer serious-or-worse", "absolute": "0.67 injury crashes per million miles vs ~3.77 benchmark; 841 fewer injury crashes lifetime" }, "reach": "Phoenix, SF Bay Area, Los Angeles, Austin, Atlanta; ~4M miles/week by mid-2026", "dates": { "achieved_start": "2019-01", "achieved_end": "2026-06-30", "published": "2026-09-24", "added": "2026-10-22", "updated": "2026-10-22" }, "sources": [ { "title": "Waymo Safety Impact hub", "url": "https://waymo.com/safety/impact/", "type": "primary" }, { "title": "Waymo: safety data update, Sept 2026", "url": "https://waymo.com/blog/shorts/safetydata-september26/", "type": "primary" }, { "title": "Waymo: safety data update, June 2026 (through March 2026)", "url": "https://waymo.com/blog/shorts/safetydata-june26/", "type": "primary" }, { "title": "Parameter: IIHS commentary on Waymo data release", "url": "https://parameter.io/waymo-autonomous-vehicles-slash-injury-crashes-by-more-than-80/", "type": "secondary" } ], "conflicts": "Waymo owns the data and the benchmark methodology; IIHS has commended the transparency but not independently audited the counts.", "what_would_change": "An independent (NHTSA, IIHS, academic) replication of the benchmark on the same miles would move evidence to B.", "notes": "The 841 figure is lifetime. For a strict 12-month view use the quarterly deltas: ~134 avoided injury crashes in Q2 2026 across ~51M miles.", "related": [], "history": [ { "date": "2026-10-22", "note": "Added." } ] } ] }