Medical AI errors are found to have contributed to 100 deaths

Where things stand (as of 16 Oct 2026)

I found no qualifying publication dated on or after 16 Sep 2026. No court, regulator or public inquiry anywhere has published a finding that errors by a medical AI system contributed to 100 or more patient deaths, and I am not aware of any such finding before the window either. The cases closest to the question are all far from a qualifying finding:

• UnitedHealth's nH Predict (Estate of Lokken v. UnitedHealth, D. Minn.), plus the similar Humana and Cigna suits. The Lokken case is still in discovery: a magistrate judge ordered broad document production in March–April 2026 (ArentFox Schiff, 23 Apr 2026; Vox, Sep 2026). The surviving claims are for breach of contract and good faith, so the court never has to make a finding about how many people died.
• Lampard Inquiry (a statutory inquiry into more than 2,000 Essex mental-health deaths, 2000–2023). It is looking at Oxevision, a camera-based monitoring system. Interim recommendations may follow a hearing on 20 Oct, and the final report is not expected until 2028 (BBC, Oct 2026; Health Tech World; BBC, 24 Jul 2026). The Oxevision-linked deaths raised so far are individual cases, many involving staff muting or ignoring alerts. A finding that Oxevision's own errors contributed to 100 or more Essex deaths is implausible.
• Australia's aged-care funding algorithm (the Integrated Assessment Tool). A Senate committee inquiry into the Support at Home program is due to report on 24 Nov 2026 (APH), and the Commonwealth Ombudsman is also investigating (Guardian, 19 Aug 2026; Guardian, 9 Sep 2026). This is the only scheduled report in the near term, but it is very unlikely to qualify, for three reasons:
  • The department says the tool is rule-based, not AI (HelloLeaders, Nov 2025), and published analysis describes a points-and-thresholds decision tree.
  • The mortality data the department is preparing counts people who "died while waiting" for a care package (Ageing Agenda, 23 Sep 2026). That shows association, not deaths caused by the tool's errors.
  • It is unclear whether aged-care home support counts as "medical care".
• Medicare's AI prior-authorisation pilot (WISeR). Delays and errors have been reported, and a Senate move to repeal the pilot failed in July 2026 (MedPage Today, 17 Jul 2026). No official body has made findings about deaths.
• ChatGPT wrongful-death suits. These are individual lawsuits. A general-purpose chatbot is arguably not "used in medical care", and no court has made findings across cases.
• UK Covid Inquiry Module 3 (healthcare systems, including NHS 111). It was published in March 2026, before the window opened.

Reference class and base rate

I looked for official findings that attribute 100 or more patient deaths to a single software or automated system in healthcare, and found none in about 40 years:

• Therac-25 caused 3 deaths.
• In the 2018 UK breast-screening failure, "up to 270" was an upper bound, which the question explicitly excludes, and the software was conventional.
• Estimates of deaths from the 2020 Excel test-and-trace error came from researchers, not an official body.

Even a heavily investigated, widely deployed AI system has an officially attributed toll far below 100: when NHTSA closed its Tesla Autopilot investigation it counted 13 fatal crashes and 14 deaths (NHTSA EA22-002, Apr 2024). Major algorithmic scandals outside medicine, such as Robodebt and the Dutch childcare-benefits scandal, ended in inquiry reports that did not put a number on deaths. Official death counts of 100 or more usually come from acute, biologically clear disasters, such as contaminated products or blood, which is not how AI harms tend to show up.

Paths to YES and how I built the forecast

Rough probabilities of a qualifying finding by the end of 2031, by route:

• US insurance or Medicare AI denials, through a court, regulator or congressional inquiry: about 0.5%
• A systematic failure of a clinical AI system (triage, diagnosis, monitoring or dosing) followed by a quick regulator or inquiry finding: about 0.8%
• Government care-access algorithms that are also accepted as AI: about 0.4%
• Consumer health or mental-health chatbots: about 0.3%
• Existing inquiries (Lampard, Scottish Covid inquiry and similar): about 0.2%
• Other countries and unknown cases: about 0.4%

I set the hazard (the chance the first qualifying finding appears in a given quarter, if none has yet) at about 0.10% for the part of Q4 2026 that remains. It rises from about 0.07% to 0.09% per quarter in 2027 to about 0.18% per quarter in 2031. The rise reflects AI spreading into high-volume care decisions and the time it takes for inquiries and court cases to produce findings. Accumulated in Python, this gives about 0.4% by end-2027, 1.4% by end-2029 and 2.7% by end-2031.

Strongest case for a higher number: AI is being deployed fast in care-access decisions that affect vulnerable elderly people, partisan legislative reports may adopt dramatic estimates, and the definitions are broad: a contributing cause is enough, decision support reviewed by a human counts, and an adopted estimate counts. Why I don't move much: a finding must meet every condition at once. It has to identify the system as AI, attribute the errors to it, make its own causal finding, and count 100 or more distinct deaths from the same system. Inquiries also take years. In past scandals, official bodies have not done this, even when the evidence was much clearer.

Question. By the last day of each quarter, has a court, regulator, or public inquiry in any country published a finding that errors by an AI system used in medical care contributed to the deaths of at least 100 patients?

Definitions.

• AI system: a machine-learning model or a knowledge-based AI system that infers predictions, recommendations, decisions, or actions from inputs. Any developer's system qualifies; the covered-developer list does not apply. Ordinary database, scheduling, billing, or other conventional software does not qualify solely because it uses an algorithm or contains a programming error. The official finding or accompanying technical documentation must identify the relevant system as AI or describe its qualifying functionality.
• Used in medical care: actually used in screening, diagnosis, triage, monitoring, treatment, prescribing, or decisions about patients' access to or continuation of care, including insurance authorisation where it affects care delivered. Human-reviewed decision support qualifies. Research-only evaluations and drug-discovery tools not used in patients' care do not.
• Errors: incorrect outputs, recommendations, decisions, actions, or failures to act attributed by the finding to the AI system. An AI error may be one of several contributing causes, including human or institutional failures; sole causation, negligence, and autonomous operation are not required.
• Contributed to deaths: the authority affirmatively finds a causal contribution to actual patient deaths, including hastening death. Association, exposure to risk, and statements that the system might have caused deaths do not suffice.
• At least 100 patients: 100 distinct deceased patients attributed to the same identifiable AI system. Multiple deployments, related versions, hospitals, incidents, and time periods may be included if the authority treats them as the same system and states the combined death toll. Do not sum unrelated systems or independent findings. An adopted numerical estimate of attributable deaths qualifies if its point estimate is at least 100, or, absent a point estimate, its stated range has a lower bound of at least 100. An upper bound such as “up to 270,” an exposure count, and projections of future deaths do not qualify.
• Court, regulator, or public inquiry: a legally constituted court; a governmental or statutory authority with regulatory, enforcement, or patient-safety investigative jurisdiction over the conduct; or a formally government- or legislature-commissioned public inquiry. Private accreditation bodies, academic researchers, and developer-appointed reviews do not qualify on their own.

Resolves YES if a qualifying authority publishes a judgment, adopted factual findings, final regulatory investigation or enforcement report, or final or formally adopted interim public-inquiry report that establishes all the above elements. The authority must adopt the AI-error attribution, causal conclusion, and death count as its own findings. Supporting incorporated annexes may supply these details. A merits finding counts even if appeal or review remains available; a later reversal does not undo a qualifying YES under the Resolution window convention.

Exclusions (resolve NO). Complaints, allegations, filings by litigants, testimony merely reproduced in a report, investigation announcements, settlements without adopted findings, recall notices reporting only possible harm, and rulings that merely assume allegations true do not count. Nor do studies or press reports, even if they estimate more than 100 deaths, unless a qualifying authority adopts the conclusions in a qualifying finding. Resolve NO for a horizon with no qualifying publication by its end.

Resolving source. The issuing court's official judgment or docket publication; the regulator's official decision, report, or enforcement publication; or the public inquiry's published report on its official site or that of its commissioning government or legislature. Officially published copies of the underlying documents count; a press release alone does not substitute for the finding. These are the narrower resolving sources for this question, rather than the broader Public confirmation convention A2; developer agreement or absence of dispute is not required. Apply the Resolution date convention A3 and the Resolution window convention. Earlier deaths may qualify when the official finding is first published within the window; mere republication of a pre-opening finding does not count.