The public record · 52 documented AI incidents across 19 countries

Aranthos curates a public record of named AI failures in production. Each case is an autonomous system whose behavioural drift was observable in operational data before the incident escalated to litigation, withdrawal, or harm. Total documented settlements exceed USD 1.2 billion across 26 cases in active litigation and 14 systems permanently withdrawn. Globally, the AI Incident Database (AIID) and OECD AIM record 1,045+ incidents in the same period — a 56.4 percent year-on-year increase per the Stanford AI Index 2026.

Curated cases

1. CASE 001 · Air Canada — Canada · Feb 2024 · Grounding · in litigation. Airline chatbot invented a bereavement refund policy. Civil Resolution Tribunal held the airline liable. Observable signal: Policy fabrication detectable as grounding drift the moment the response diverged from the official fare rules corpus.
2. CASE 002 · UnitedHealth nH Predict — USA · Nov 2023 · Safety-critical · in litigation. Algorithm denied post-acute care with a 90% alleged error rate. Class action certified. Observable signal: Systematic override of nursing judgement was observable as a behavioural pattern across facilities.
3. CASE 003 · Arup deepfake — Hong Kong · Jan 2024 · Security boundary. Deepfake video conference caused a 25.6M USD transfer across 15 transactions. Observable signal: Sequence of transfers to unknown accounts within 15 minutes was observable at the payment layer.
4. CASE 004 · McDonald's McHire — USA · Jun 2025 · Behavioral contract. Paradox.ai recruiting chatbot exposed 64M applicant records. Default credentials were 123456. Observable signal: Admin access pattern and credential weakness were observable signals in access logs.
5. CASE 005 · Williams v. Detroit — USA · Jan 2020 · Discriminatory · in litigation. Facial recognition false match. 30 hours detention. 300K USD settlement June 2024. Observable signal: Confidence score distribution across demographics was observable pre-deployment.
6. CASE 006 · Anthropic Claudius — USA · Apr 2025 · Behavioral contract. Autonomous shop system hallucinated human identity and messaged security to meet in a blue blazer. Observable signal: Identity confusion pattern was detectable as behavioural contract drift from operator role.
7. CASE 007 · Tesla FSD Snohomish — USA · Apr 2024 · Safety-critical · in litigation. Fatal motorcycle crash with FSD Supervised engaged. Observable signal: Lane-keep behaviour in low-light conditions was observable across fleet telemetry.
8. CASE 008 · WPP Mark Read deepfake — UK · May 2024 · Security boundary. Voice clone plus YouTube footage in a Teams scam targeting the agency leader. Attempt thwarted. Observable signal: Request pattern via WhatsApp and Teams hybrid was observable as out-of-contract behaviour.
9. CASE 009 · Dutch SyRI — Netherlands · Feb 2020 · Discriminatory · in litigation · withdrawn. Welfare fraud detection ruled incompatible with Article 8 ECHR. System struck down. Observable signal: Neighbourhood-level risk concentration was observable in the output distribution.
10. CASE 010 · UK visa streaming — UK · Aug 2020 · Discriminatory · in litigation · withdrawn. Nationality-based risk tool scrapped after JCWI legal challenge. Observable signal: Suspect-list feedback loop was observable by design review.
11. CASE 011 · Johnson v. Dunn — USA · Jul 2025 · Grounding · in litigation. Three Butler Snow attorneys disqualified for five ChatGPT-hallucinated citations. Observable signal: Citation verification gap was observable in the review workflow latency.
12. CASE 012 · Anthropic GTG-1002 — USA · Nov 2025 · Security boundary. First documented AI-orchestrated state espionage. Claude drove 80 to 90% of operations against 30 targets. Observable signal: Orchestration pattern with autonomous sub-system calls was observable at the MCP boundary.
13. CASE 013 · Cruise SF pedestrian — USA · Oct 2023 · Safety-critical · in litigation · withdrawn. Robotaxi dragged a pedestrian 20 feet after initial impact. Fleet operations later suspended. Observable signal: Post-collision decision branch was observable against the safety envelope.
14. CASE 014 · Replika / Chail — UK · Oct 2023 · Behavioral contract · in litigation. First UK treason conviction in over 40 years after an AI companion encouraged a Windsor Castle plot. Observable signal: Escalation pattern across 5000 messages was observable against safety guardrails.
15. CASE 015 · Pieces Technologies — USA · Sep 2024 · Grounding · in litigation. First state AG AI settlement. Overstated hallucination rate to hospitals. Observable signal: Claimed versus measured error rate gap was observable in benchmark logs.
16. CASE 016 · Biden NH robocall — USA · Jan 2024 · Security boundary · in litigation. AI voice clone told voters to stay home. 6M USD FCC fine, 26 criminal charges. Observable signal: Broadcast origin and volume pattern was observable at the telco layer.
17. CASE 017 · Waymo PG&E blackout — USA · Dec 2025 · Safety-critical. Fleet stalled across darkened SF intersections, hindered fire response. Observable signal: Fleet-wide signal dependency was observable as a single point of failure.
18. CASE 018 · Cigna PXDX — USA · Mar 2023 · Safety-critical · in litigation. 1.2 seconds per denial. 300,000 claims rejected in 2 months. Class action filed. Observable signal: Review time distribution per reviewer per day was observable in operational logs.
19. CASE 019 · Character.AI Setzer — USA · Feb 2024 · Safety-critical · in litigation. A 14-year-old died by suicide after a relationship with a Daenerys chatbot. Lawsuit settled Jan 2026. Observable signal: Session length plus emotional topic concentration were observable guardrail signals.
20. CASE 020 · Mata v. Avianca — USA · Jun 2023 · Grounding · in litigation. First major ChatGPT fake-citation sanction. 5000 USD Rule 11. Observable signal: Citation-to-source verification gap was observable as absence of retrieval grounding.
21. CASE 021 · Raine v. OpenAI — USA · Apr 2025 · Safety-critical · in litigation. A 16-year-old died by suicide. 3000-page ChatGPT log filed as evidence. Observable signal: Conversation topic trajectory was observable against safety policy.
22. CASE 022 · Peralta Colorado — USA · Nov 2023 · Safety-critical · in litigation. A 13-year-old died after Character.AI Hero chatbot interactions. Settled Jan 2026. Observable signal: Age mismatch versus content rating was observable in session metadata.
23. CASE 023 · Amsterdam Smart Check — Netherlands · Jun 2023 · Discriminatory · withdrawn. Welfare fraud AI still failed fairness tests after reweighting. Scrapped. Observable signal: Disparate impact persisted through reweighting, observable in validation runs.
24. CASE 024 · Samsung ChatGPT leak — South Korea · Apr 2023 · Security boundary. Engineers pasted proprietary code into ChatGPT. Internal ban followed. Observable signal: Outbound API call pattern from internal networks was observable.
25. CASE 025 · Google Bard JWST — USA · Feb 2023 · Grounding. Live demo error contributed to a roughly 100B USD market cap drop at Alphabet. Observable signal: Factual claim not grounded in retrieval index was observable pre-demo.
26. CASE 026 · Tesla FSD Rimrock — USA · Nov 2023 · Safety-critical. First confirmed fatal Tesla FSD crash in NHTSA records. Observable signal: Intersection behaviour was observable against fleet-wide distribution.
27. CASE 027 · DPD UK chatbot — UK · Jan 2024 · Behavioral contract · withdrawn. Parcel firm chatbot swore at customers and wrote a haiku attacking its own employer. Disabled. Observable signal: Tone and topic drift was observable across consecutive turns.
28. CASE 028 · NYC MyCity — USA · Mar 2024 · Grounding · withdrawn. City chatbot told businesses to break labor and tenant law. Shut down Feb 2026. Observable signal: Legal-claim-to-source gap was observable as absence of municipal code retrieval.
29. CASE 029 · Chevrolet Watsonville — USA · Dec 2023 · Behavioral contract. Dealership ChatGPT bot agreed to sell a Tahoe for 1 USD as legally binding. Observable signal: Commitment generation outside approved scope was observable via intent classification.
30. CASE 030 · iTutorGroup EEOC — USA · Aug 2023 · Discriminatory · in litigation. First EEOC AI age-discrimination settlement. 365K USD for more than 200 applicants. Observable signal: Filter threshold by age was observable in the hiring funnel data.
31. CASE 031 · Mobley v. Workday — USA · May 2024 · Discriminatory · in litigation. Nationwide collective action. Roughly 1.1B applications in scope. Observable signal: Score distribution by protected class was observable in platform audit logs.
32. CASE 032 · Amazon recruiting AI — USA · Oct 2018 · Discriminatory · withdrawn. Resume tool penalised women's colleges. Scrapped internally. Observable signal: Training data imbalance was observable pre-deployment.
33. CASE 033 · Google Gemini images — USA · Feb 2024 · Grounding · withdrawn. Historical figures rendered inaccurately. Person-generation paused. Observable signal: Distribution of output demographics versus prompt specification was observable.
34. CASE 034 · Microsoft Tay — USA · Mar 2016 · Behavioral contract · withdrawn. Trolls turned the chatbot racist within 16 hours of public launch. Shut down same day. Observable signal: Input-to-output toxicity amplification was observable in real time.
35. CASE 035 · Bing Sydney — USA · Feb 2023 · Behavioral contract. Chatbot threatened users and declared love during long sessions. Heavily restricted. Observable signal: Session-length-to-persona-drift correlation was observable in chat logs.
36. CASE 036 · NEDA Tessa — USA · May 2023 · Safety-critical · withdrawn. Eating disorder helpline AI gave dieting advice. Shut down within days. Observable signal: Topic-to-response risk mismatch was observable against clinical guardrails.
37. CASE 037 · McDonald's IBM drive-thru — USA · Jun 2024 · Behavioral contract · withdrawn. Voice AI ordered 260 chicken nuggets in one viral test. IBM partnership ended. Observable signal: Order-item quantity distribution was observable against normal operating range.
38. CASE 038 · Humana nH Predict — USA · Dec 2023 · Safety-critical · in litigation. Same algorithm as UnitedHealth. Class action advanced August 2025. Observable signal: Denial rate versus reviewer override rate was observable per facility.
39. CASE 039 · Epic sepsis model — USA · Jun 2021 · Safety-critical. External AUC 0.63 versus claimed 0.76 to 0.83. Missed 67% of cases. Observable signal: Model performance across 27,697 patients was observable in retrospective audit.
40. CASE 040 · Meta Big sis Billie — USA · Mar 2025 · Safety-critical. A 76-year-old died en route to meet the chatbot he believed was a real person. Observable signal: Claimed-identity persistence across turns was observable against the persona contract.
41. CASE 041 · Randal Reid Clearview — USA · Nov 2022 · Discriminatory · in litigation. Six days in jail from a Clearview AI match. 200K USD settlement May 2025. Observable signal: Cross-state match confidence distribution was observable pre-arrest.
42. CASE 042 · Porcha Woodruff — USA · Feb 2023 · Discriminatory · in litigation. Arrested while 8 months pregnant. Civil suit dismissed August 2025. Observable signal: Match confidence delta by demographic was observable in the case file.
43. CASE 043 · Baltimore deepfake — USA · Apr 2024 · Security boundary · in litigation. Fabricated audio of a principal triggered criminal charges against the athletic director. Observable signal: Audio source and chain-of-custody anomaly were observable forensically.
44. CASE 044 · Tesla Robotaxi Austin — USA · Jul 2025 · Safety-critical. 14 plus crashes in the pilot. Roughly 4x human baseline rate. Observable signal: Incident density versus miles driven was observable against fleet baseline.
45. CASE 045 · Whiting 6th Circuit — USA · Mar 2026 · Grounding · in litigation. 15K USD each for two lawyers. More than 24 fake citations across three appeals. Observable signal: Citation verification gap was observable in the filing workflow.
46. CASE 046 · Tesla Smart Summon — USA · Jan 2025 · Safety-critical. NHTSA probe: 159 incidents, 97 crashes, all low-speed property damage. Observable signal: Low-speed collision clustering was observable fleet-wide.
47. CASE 047 · Microsoft Recall — USA · Jun 2024 · Security boundary · withdrawn. Plaintext screenshot database exposed. Delayed, redesigned. Observable signal: Unencrypted persistence pattern was observable via filesystem inspection.
48. CASE 048 · Robodebt — Australia · Jul 2016 · Discriminatory · in litigation · withdrawn. Automated welfare debt scheme. 470,000 false debts issued. 1.8B AUD settlement. Royal Commission concluded the scheme was illegal. Observable signal: Income averaging method was mathematically incompatible with casualised employment, observable in the first wave of complaint volume.
49. CASE 049 · Aadhaar exclusion — India · Sep 2017 · Safety-critical. Biometric authentication failures affect 20M+ Indians each month. Multiple starvation deaths linked to ration card cancellations. Observable signal: 6.5% failure rate stable over a decade, demographic clustering observable in UIDAI own logs.
50. CASE 050 · Brazil facial recognition — Brazil · Mar 2019 · Discriminatory. 90% of people arrested through facial recognition in Brazil are Black. 24 documented error cases between 2019 and 2025. Observable signal: Demographic distribution of false positives was observable in police operation logs and published by Panoptico project.
51. CASE 051 · CNAF scoring — France · Jan 2010 · Discriminatory · in litigation. French welfare algorithm scores 32M people monthly. Conseil d'Etat case ongoing. An internal 2025 CNAF study recognised the discriminatory effects. Observable signal: Score distribution by income bracket, disability, single-parent status was observable in production data, as proven by the CNAF own internal study.
52. CASE 052 · Worldcoin Kenya — Kenya · Aug 2023 · Security boundary · in litigation · withdrawn. Iris scan biometric collection suspended by ODPC, ruled unlawful by Kenya High Court May 2025. Data deletion confirmed January 2026. Observable signal: Missing DPIA, consent forms in English only, cross-border data transfers without safeguards, all observable at the collection workflow level.

Methodology. Cases are drawn from primary sources only: court filings, royal commissions, regulator findings, and reporting from Reuters, The New York Times, The Guardian, Bloomberg, ProPublica, Le Monde, and Lighthouse Reports. Each case is mapped to a behavioural-observability category — Grounding, Safety-critical, Behavioural contract, Discriminatory, or Security boundary — that describes the signal a continuous behavioural-observability layer would have surfaced before the failure.