AInvasion AI Failure Ledger · Last checked 15 October 2026 · 33 cases
Every row below is a documented AI or automation failure tied to a court order, regulator, filing, audit, or report. Rows are grouped by what the failure cost, from a few thousand dollars to more than a billion. Each row states what kind of money the figure is, how much of the failure belongs to the AI system itself, how well the number is sourced, and whether it was re-checked for this edition.
Jump to a cost band: Under $1k · $1k–$10k · $10k–$100k · $100k–$1M · $1M–$10M · $10M–$100M · $100M–$1B · $1B+ paid · Big numbers that are not payments · No figure and pending
Read this before quoting a number
- This is not a measure of how often AI fails. Cases appear here because a court, regulator, filing, or journal made them public. There is no denominator, so do not infer a failure rate or treat the list as representative of any industry.
- Costs are different kinds of money. A court sanction, a settlement, an accounting write-down, a share-price move and an investment are not the same thing. Every row carries a type of figure, and nothing in this database is added up.
- Cost bands are approximate US dollars, for sorting only. Conversions use rounded planning rates (A$1 ≈ US$0.65, C$1 ≈ US$0.72, €1 ≈ US$1.08, KRW 103.3M ≈ US$92,900 as stated by the source). The original figure always appears first. Quote the original, from the linked source.
- Unproven claims are not costs. Pending cases and amounts that were only sought sit under No figure and pending.
- Check the “Checked” column. It says whether the figure was matched to a primary document on 15 October 2026, to secondary reporting that day, or carried over unchanged from the 12 October Ledger. See the verification log.
Columns
| Column | Meaning |
|---|---|
| AI role | Core: the AI system’s own behaviour or oversight was the failure. Adjacent: AI is involved, but the failure lies elsewhere (a marketing claim, how data was acquired, a human who did not check output). Baseline: not machine learning; kept because the liability pattern recurs. |
| Grade | P: the figure appears in a primary document. R: reputable reporting, several outlets agree; primary document not retrieved. S: secondary sources only; verify before quoting. |
| Checked | Primary 15 Oct: matched to a primary document that day. Secondary 15 Oct: matched to news or legal commentary that day. Ledger 12 Oct: carried over, not re-retrieved. |
| Can change | An appeal, approval, payment or ruling is still pending. Re-check before quoting. |
What the ladder shows
- Only five rows with a paid or booked cost are failures of the AI system itself: the Air Canada chatbot, Zillow’s pricing model, Cruise’s robotaxi (penalised for its crash report, not the crash), Earnest’s underwriting model and the MD Anderson Watson project. Every other row in the ladder is AI-adjacent or a baseline: a human who did not check, a marketing claim, a data-sourcing or consent violation, or automation that was not machine learning. (Bard’s market-value drop also involves model behaviour, but it is not a cost anyone bore.)
- The $1B+ band contains no model-accuracy failure. Its two paid rows are a copyright settlement over how training data was acquired and a biometric-consent settlement. The two bigger numbers in the database are not payments at all (see below).
- Citation sanctions climb, but slowly. From US$2,000 (2024) to about US$110,000 (2026), with courts saying smaller fines had not deterred filings. These are chosen endpoints from a non-random sample, not a trend line.
- Regulators punish the claim or the cover-up more often than the failure. Cruise (the report), DoNotPay, and Air AI (capability claims).
- The bill arrives years later. Robodebt failed from 2015 to 2019 and was repriced in 2020 and 2026. Tesla’s 2019 crash produced a 2025 verdict and a 2026 appeal.
Coverage and skew
| Where the case sits | Cases |
|---|---|
| United States | 27 |
| Australia | 2 (Robodebt, Deloitte) |
| Canada | 1 (Air Canada) |
| South Korea | 1 (Lee Luda) |
| Netherlands (regulator) | 1 (Clearview AI) |
| Sweden (company) | 1 (Klarna) |
Roughly four in five cases are US-centred, and the rest are Australia, Canada, Western Europe and South Korea. This edition added the Korean and Dutch cases after targeted searches; it did not search for cases in China, India, Japan, Latin America, Africa or the Middle East, and none appear. Failures that stay internal or are quietly fixed never appear at all. Read the ledger as a record of what becomes public through courts and regulators, mostly in common-law and EU systems, not as a map of where AI fails.
Rows considered and left out are listed at the end.
Under $1k
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Air Canada chatbot (Moffatt v. Air Canada) | 2022 → 2024 | C$812.02 | ≈585 | Tribunal award | Core | P | Primary 15 Oct | No |
Air Canada. C$650.88 damages, C$36.14 interest and C$125 in tribunal fees, ordered on 14 February 2024. The customer paid C$1,630.36 for two flights; the tribunal’s estimate of the bereavement fare was C$979.48, and the difference is the C$650.88 in damages. The chatbot had told him he could claim the bereavement discount after booking, which the airline’s policy did not allow. The tribunal rejected the argument that the chatbot was a separate legal entity. Caveats: a small-claims tribunal ruling, persuasive and not binding precedent. The figure excludes legal time and reputational cost. Some press reports gave the bereavement fare as C$949.48; the decision text gives C$979.48, and the arithmetic confirms it. Sources: BC Civil Resolution Tribunal decision, 2024 BCCRT 149 · MobileSyrup
$1k–$10k
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Gauthier v. Goodyear (E.D. Tex.) | 2024 | US$2,000 | 2,000 | Court sanction | Adjacent | R | Secondary 15 Oct | No |
| Mata v. Avianca (S.D.N.Y.) | 2023 | US$5,000 | 5,000 | Court sanction | Adjacent | R | Secondary 15 Oct | No |
| Wadsworth v. Walmart (D. Wyo.) | 2025 | US$5,000 total | 5,000 | Court sanction | Adjacent | R | Secondary 15 Oct | No |
| Coomer v. Lindell (D. Colo.), defense counsel | 2025 | US$6,000 total | 6,000 | Court sanction | Adjacent | R | Secondary 15 Oct | No |
Gauthier. The court ordered US$2,000 paid into the court registry, one hour of generative-AI legal education, and a copy of the order sent to the client, after a summary-judgment response cited nonexistent cases and quotations and was not corrected when the defendant raised the problem. Caveat: the defendant’s cost of checking the citations was weighed but is not in the figure. Sources: Court memorandum and order (CourtListener) · Texas Lawbook
Mata. US$5,000, joint and several on the firm and two lawyers, for a brief with fake opinions and quotes produced by ChatGPT that the lawyers defended after the court questioned them. They also had to notify each judge falsely named as an author. The injury case was dismissed as time-barred. An early, widely cited sanction; later ones are larger. Sources: Bloomberg Law · Reed Smith
Wadsworth. US$3,000 on the drafting attorney (whose pro hac vice status was revoked) and US$1,000 each on two others, for a motion citing eight nonexistent cases. The firm was not sanctioned: it withdrew the motion, paid the other side’s fees for it, and added verification training. Sources: LawNext · Bloomberg Law
Coomer v. Lindell. US$3,000 on lead counsel and his firm jointly, and US$3,000 on a second attorney, for a brief with nearly 30 defective or fabricated citations. Lead counsel said a draft had been run through AI tools. The sanctions did not extend to Lindell or his companies, whom counsel had not told about the AI use. Caveat: the judge treated the cause as either improper AI use or gross carelessness. Sources: News report (Yahoo) · AI Incident Database
$10k–$100k
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Noland v. Land of the Free (Cal. Ct. App.) | 2025 | US$10,000 | 10,000 | Court sanction | Adjacent | R | Secondary 15 Oct | No |
| Lee Luda chatbot (ScatterLab, South Korea) | to 2021 | KRW 103.3M | ≈92,900 | Regulator fine | Adjacent | R | Secondary 15 Oct | No |
| Deloitte report for Australia’s DEWR | 2025 | A$97,587 refund | ≈63,000 | Refund | Adjacent | R | Ledger 12 Oct | No |
| Whiting v. City of Athens (6th Cir.) | → 2026 | US$15,000 each, plus fees | ≈30,000+ | Court sanction | Adjacent | R | Ledger 12 Oct | No |
Noland. US$10,000 payable to the court clerk, a State Bar referral, and service of the opinion on the client, after nearly all quotations in an appellate brief were fabricated by AI tools. California’s first published opinion on the subject. The court called the amount conservative and awarded nothing to the other side, which had not flagged the fakes. The figure sits exactly on the band edge and is listed here, in the higher band. Sources: Proskauer · McGuireWoods
Lee Luda. South Korea’s Personal Information Protection Commission fined ScatterLab KRW 103.3M (about US$92,900) in April 2021, finding eight violations of the Personal Information Protection Act. They included using messages collected through the company’s other apps to build the chatbot beyond the purpose users agreed to, failing to delete or encrypt personal information in training data, publishing models on GitHub containing identifiable messages, and collecting data from children under 14 without parental consent. The first reported case of that Act being applied to an AI system. Caveat: a data-protection case, not a model-accuracy failure, hence Adjacent. Source: Future of Privacy Forum (the Commission’s decision was not retrieved)
Deloitte. A$97,587 refunded, the final instalment of an A$440,000 contract, as reported at a Senate estimates hearing on 8 October 2025; the department confirmed over A$97,000. A 237-page report on an automated welfare-penalty system, drafted with Azure OpenAI GPT-4o, contained a fabricated Federal Court quote and nonexistent references, and the corrected version disclosed the AI use. Caveats: under a quarter of the fee was returned; Deloitte did not say AI caused the original errors; the department said recommendations were unchanged. The AusTender record was not retrieved. Sources: Malay Mail · Australian Greens release
Whiting. US$15,000 each on two attorneys as punitive sanctions, plus the appellees’ full appellate fees, double costs and a disciplinary referral (13 March 2026). Caveats: the court did not expressly find that generative AI produced the fake citations. The US$15,000 is read two ways (per lawyer in total, or per appeal); this entry uses the quoted opinion text, and under the wider reading plus fees the row could reach the next band. Sources: Volokh Conspiracy, quoting the opinion · LawNext · ABA Litigation News
$100k–$1M
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Couvrette v. Wisnovsky (D. Or.) | 2025 → 2026 | ≈US$110,204 | 110,204 | Court sanction | Adjacent | R | Ledger 12 Oct | Appeal possible |
| DoNotPay (“robot lawyer”) | 2021–2023 → 2025 | US$193,000 | 193,000 | FTC order | Adjacent | P | Ledger 12 Oct | No |
| iTutorGroup (hiring software) | 2020 → 2023 | US$365,000 | 365,000 | Settlement | Adjacent | R | Ledger 12 Oct | No |
Couvrette. US$15,500 in fines (US$500 per fake case, US$1,000 per fabricated quote) plus US$94,704 in the other side’s fees: 15 fake cases and 8 fabricated quotations. Lead counsel owes about US$96,000 and a second attorney about US$14,200. Reported as the largest such sanction on record. Caveats: the court found persuasive evidence that the client, not the lawyers, drafted the filings, and lead counsel said he did not write them and acted pro bono. Sources differ on whether dismissal came in December 2025 or on 30 March 2026; the AI tool was never identified, and the client said she would appeal. The per-lawyer split comes from a tracker. Sources: ABA Journal · Opinion and order (CourtListener)
DoNotPay. US$193,000 in monetary relief and notice to 2021–2023 subscribers under the final FTC order (January 2025). The charge was an unsubstantiated claim: the product was never tested against human lawyers, and no attorneys were retained to test its quality. Customer harm was not measured. Sources: FTC final order · DLA Piper
iTutorGroup. US$365,000 among 200+ rejected applicants and a five-year consent decree. The EEOC alleged the software automatically rejected women over 55 and men over 60. Caveat: an alleged hard-coded age filter; whether any machine learning was involved is not established, and the case settled without a finding. Sources: Duane Morris · Bloomberg Law
$1M–$10M
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Cruise robotaxi (crash report) | 2023 → 2024 | ≈US$2.1M combined | 2,112,500 | Criminal fine and regulatory penalties | Core (penalty for the report) | P/R | Ledger 12 Oct | No |
| Earnest Operations (AI-assisted underwriting) | n/s → 2025 | US$2.5M | 2,500,000 | Settlement | Core | R | Ledger 12 Oct | No |
Cruise. US$500,000 criminal fine under a deferred prosecution agreement (primary), plus a US$1.5M NHTSA penalty and US$112,500 to the California PUC (reported). After a robotaxi struck and dragged a pedestrian in October 2023, the company gave federal investigators an incomplete crash report and admitted to submitting a false record. Caveats: the penalty was for the report, not the crash. It excludes the loss of the California licence and the grounded fleet. A reported victim settlement of “at least US$8M” rests on one outlet and is not counted. Sources: US Attorney’s Office, N.D. Cal. · TechCrunch
Earnest. US$2.5M settlement with the Massachusetts Attorney General, plus AI governance, annual fair-lending testing and removal of two variables. The attorney general alleged disparate impact on Black, Hispanic, and non-citizen borrowers and a failure to test the models. Caveats: settlement of allegations; the underwriting period is not specified, and the number of borrowers harmed is not established. Sources: ConsumerAffairs · Today’s General Counsel
$10M–$100M
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Air AI (FTC proposed order) | n/s → 2026 | US$18M, largely suspended | 18,000,000 (headline) | Proposed order | Adjacent | P/R | Ledger 12 Oct | Yes |
| Clearview AI (Dutch DPA) | to 2024 | €30.5M | ≈33,000,000 | Regulator fine | Adjacent | R | Secondary 15 Oct | Unknown |
| IBM Watson at MD Anderson | 2012–2016 → 2017 | more than US$62M | 62,000,000 | Project spend (audit) | Core | R | Secondary 15 Oct | No |
Air AI. Proposed FTC order (24 March 2026): US$18M monetary judgment, largely suspended for inability to pay, US$50,000 for consumer relief and a ban on marketing business opportunities. The FTC alleged false earnings and refund claims, with buyers losing about US$19M. Caveats: the order was proposed and subject to court approval in D. Ariz.; entry was not verified. The charge is marketing, not model failure. The amount actually payable is far below the headline. Sources: FTC release · CFO Dive
Clearview AI. The Dutch Data Protection Authority fined the company €30.5M (decision 16 May 2024, announced 3 September 2024) for building a database of more than 30 billion scraped photos converted to biometric codes and sold for facial recognition, which it found unlawful under the GDPR. It also found the company did not inform individuals or cooperate with access requests, and ordered it to stop, with a further penalty for non-compliance. Caveats: a data-protection case, not a model-accuracy failure. Whether the fine was paid, contested or enforced was not established in the source retrieved. The Dutch decision itself was not retrieved. Source: Hunton Andrews Kurth privacy blog
IBM Watson at MD Anderson. More than US$62M spent over about four years on an oncology advisor that was never used with patients: about US$39.2M to IBM and US$21.2M to PwC, with the rest to other vendors, per the University of Texas System audit as reported. Caveats: the audit examined procurement and contract handling, not clinical accuracy; reported problems also included failure to integrate with the centre’s new records system. IBM called the R&D project a success. The audit itself was not retrieved, and this is project spend, not a penalty. Sources: Fierce Biotech · Network World · Engadget
$100M–$1B
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Tesla Autopilot (Benavides v. Tesla) | 2019 → 2025–26 | US$242.57M (Tesla’s share) | 242,570,000 | Jury verdict | Adjacent | R | Primary docket, 4 May snapshot | Yes |
| Zillow Offers (home-pricing model) | 2021 → 2021–22 | US$304.4M (Q3 2021 write-down) | 304,400,000 | Accounting write-down | Core | P | Primary 15 Oct (Q3 figure) | No |
| Robodebt (Australia) | 2015–2019 → 2020, 2026 | A$548.5M (2026 settlement) | ≈356,000,000 | Class settlement | Baseline | P/R | Ledger 12 Oct | Yes |
| RealPage renter class (proposed) | → 2026 | US$359.9M | 359,900,000 | Proposed settlements | Adjacent | R | Ledger 12 Oct | Yes |
| Knight Capital trading software | 2012 → 2012–13 | about US$440M | 440,000,000 | Trading loss | Baseline | P | Primary 15 Oct | No |
Tesla. The jury awarded about US$329M in total and found Tesla 33% responsible; Tesla’s share is US$42.57M compensatory plus US$200M punitive. Autopilot is driver assistance, not autonomous driving. Tesla appealed (No. 26-10858, filed 16 March 2026). The docket snapshot retrieved is dated 4 May 2026 and shows Tesla’s brief due 2 July 2026 after an extension, with no argument date; nothing later was retrieved, so the appeal’s current stage is unconfirmed. The verdict amounts are carried from the Ledger. An earlier tracker’s reply-brief date of 6 November 2026 was not confirmed. Sources: Business Insurance (Reuters) · Eleventh Circuit docket (Justia)
Zillow. The Q3 2021 write-down of US$304.4M is stated in the 10-Q and the 2 November 2021 8-K, which also expected a further US$240–265M of charges in Q4 and a roughly 25% workforce cut. The Ledger gives a full-year 2021 write-down of US$407.9M from the 10-K, and total 2021 wind-down charges of US$491.2M that include it; those two were not re-retrieved because the 10-K’s financial pages could not be read. Do not add the figures together, and treat a reported US$881M loss on Offers as a different measure. Caveat: management cited home-price unpredictability and a cooling market, so model error and market conditions cannot be separated. Sources: Form 10-Q, Q3 2021 · Form 8-K, 2 Nov 2021 · Form 10-K, FY2021
Robodebt. A$475M compensation, up to A$60M administration and A$13.5M legal costs: A$548.5M, approved on 23 June 2026. That sits on top of A$112M paid in the first class action and A$1.76B in forgiven debts. Rule-based income averaging raised welfare debts automatically with no human review between 2015 and 2019; it was found unlawful. Caveats: rule-based automation, not machine learning. Reported totals run from “well over A$2B” to more than A$2.4B depending on which components are added, so this row uses only the 2026 package. The Department of Social Services release (2 September 2026) says Gordon Legal expects to start payments before the end of 2026; no payment is confirmed. Sources: Department of Social Services · Canberra Times
RealPage renter class. US$359.9M in proposed settlements with RealPage and landlords over rent-pricing software and shared landlord data; the defendants deny wrongdoing. Needs court approval, and claims are open until 29 January 2027. Not a finding of liability. Landlord settlements are reported separately; do not add rows together. Sources: Open Class Actions · KSHB
Knight Capital. The company’s 8-K reported a realised pre-tax loss of about US$440M after a software installation on 1 August 2012 sent erroneous orders for roughly 140–150 stocks into the market in the first 45 minutes of trading. The SEC’s 2013 release states a loss of more than US$460M and a US$12M penalty. Settled without admitting or denying. Caveat: a baseline case, not AI, included because the pattern (automation at scale with no effective stop) recurs. Sources: Knight Capital Form 8-K · SEC press release 2013-222
$1B+ paid or ordered
| Case | Failed → billed | Figure | ≈US$ | Type of figure | AI role | Grade | Checked | Can change |
|---|---|---|---|---|---|---|---|---|
| Meta facial recognition (Texas) | to 2021 → 2024 | US$1.4B over five years | 1,400,000,000 | State settlement | Adjacent (consent) | P/R | Primary 15 Oct | No |
| Bartz v. Anthropic (training-data piracy) | 2021–22 → 2026 | US$1.5B | 1,500,000,000 | Class settlement | Adjacent (data sourcing) | P/R | Ledger 12 Oct; status Secondary 15 Oct | Yes |
Neither row is a failure of a model’s accuracy or behaviour. Both are paid-or-ordered figures tied to how data was obtained.
Meta. Announced 30 July 2024: US$1.4B paid to Texas over five years, resolving claims under the state’s biometric-identifier law and its deceptive-practices law. The state alleged that the “Tag Suggestions” feature, launched in 2011 and turned on automatically, ran facial recognition on faces in uploaded photos and captured facial geometry without the required notice and consent. Press reports say Meta did not admit wrongdoing; the attorney general’s release does not address it. Meta shut the feature in 2021 and said it deleted face-scan data on more than a billion people. Sources: Texas Attorney General release · NBC News · Texas Tribune
Bartz v. Anthropic. US$1.5B settlement fund plus interest for 482,460 works (about US$3,000 per work), with 440,490 claimed (91.3%), per the Ledger. A judge held that training on lawfully acquired books was fair use but that building a library from pirated copies was not. Final approval came on 20 July 2026, overruling 53 objections. The 30-day appeal window closed on 19 August; whether an appeal was filed was not confirmed in the source retrieved, though the Ledger notes appeals of the approval and fee award were reported. Class counsel says payments will not begin before the end of 2026, in two stages of roughly 70% and 30%, with Anthropic paying in installments through September 2027. A Ledger note of first payments by 15 November came from a single source and conflicts with this; the class-counsel account is used. Disclosure: Anthropic, which makes Claude, is the defendant, and this edition was drafted with AI assistance from Claude. The row rests on court-record summaries and press, not company statements. Sources: Authors Guild · JURIST · Textbook & Academic Authors Association
Big numbers that are not payments
These are real events and large figures, but nobody was ordered to pay them, and they mix model behaviour with business decisions and investor reaction. They sit outside the ladder on purpose, so they do not inflate the top band.
| Case | Date | Figure | Type of figure | AI role | Grade | Checked |
|---|---|---|---|---|---|---|
| GM exits Cruise robotaxi program | Dec 2024 | more than US$10B | Capital invested over the program’s life | Adjacent | R | Secondary 15 Oct |
| Google Bard launch demo error | Feb 2023 | more than US$100B | One-day drop in Alphabet’s market value | Core | R | Secondary 15 Oct |
GM and Cruise. When GM said on 10 December 2024 that it would stop funding robotaxi work, it said it had invested more than US$10B since 2016. AP framed a similar figure differently, as more than US$10B in operating losses per GM shareholder reports, against under US$500M of revenue and a US$581M purchase of a controlling stake in 2016; the two measures are not identical. GM’s stated reasons were the time and cost to scale and a competitive market. Do not attribute the figure to the 2023 pedestrian incident. Sources: Reuters via Khaleej Times · AP via WGBH
Bard. A promotional demo of Google’s chatbot said the James Webb Space Telescope took the first pictures of a planet outside the solar system; an earlier telescope did. On 8 February 2023 Alphabet’s market value fell by more than US$100B, with shares closing about 7.7% lower after sliding as much as 9% during the day. Reuters tied the drop to the error and to a launch event that failed to impress. Caveats: not a realised loss and not money paid to anyone. Investors were also reacting to competitive worries about Microsoft, and the drop is a single day’s move. Sources: Gulf News (Reuters-sourced) · NPR via KAXE
No figure and pending
None of these has a settled dollar cost. They are kept because they show something the priced rows cannot: an accuracy measurement, a non-monetary order, or litigation that can still reprice the ledger.
Measured or non-monetary
| Case | What is known | AI role | Grade | Checked |
|---|---|---|---|---|
| Epic Sepsis Model (external validation) | An external study of 38,455 hospitalisations found AUC 0.63 against 0.76–0.83 in the vendor’s documentation; sensitivity 33%, positive predictive value 12%. It missed 67% of sepsis cases while alerting on 18% of all hospitalised patients. Measured in accuracy, not dollars or patient outcomes; Epic disputed the framing. Wong et al., JAMA Internal Medicine, 2021 | Core | P | Ledger 12 Oct |
| Klarna AI assistant | Klarna said in February 2024 that its assistant did the work of 700 agents. In May 2025 its CEO told Bloomberg that cost had been weighted too heavily and quality had suffered, and humans were rehired. No public accounting nets savings against quality loss; the headline efficiency claims rest on company materials. Entrepreneur | Core | R | Ledger 12 Oct |
| Rite Aid facial recognition | FTC ban for five years, plus deletion of images and derived models; no fine stated. The FTC found no accuracy testing before or after deployment; Rite Aid disagrees and calls it a limited pilot. Subject to bankruptcy-court approval. TechCrunch | Core | R | Ledger 12 Oct |
Pending litigation
| Case | Status | Can change | Grade | Checked |
|---|---|---|---|---|
| Character.AI and Google teen-harm suits | Five suits over teen deaths and harm, including the Setzer case, settled in January 2026. Terms undisclosed, no admission. Fortune | No | R | Ledger 12 Oct |
| UnitedHealth nH Predict (Lokken, D. Minn.) | Class action over a length-of-stay model allegedly used in place of physicians. Most claims dismissed February 2025; contract-based claims proceed; no class certified. UnitedHealth denies the model makes coverage decisions. The plaintiffs’ “over 90% reversed” figure is their reading of appealed denials, not a model error rate. ArentFox Schiff | Yes | P | Ledger 12 Oct |
| Mobley v. Workday | Applicants allege AI screening was biased by age, race, and disability. Age collective conditionally certified May 2025; most claims survived dismissal June 2026; in discovery. No finding of liability. Akin | Yes | R | Ledger 12 Oct |
| RealPage (DOJ, nine states, landlords) | DOJ consent judgment with a monitor and no money, entered 2026. Nine states’ claims survived a motion to dismiss on 30 September 2026; no liability finding. Landlord settlements of US$7M each by Greystar and LivCor are reported from secondary sources only (grade S). California AG | Yes | R/S | Ledger 12 Oct |
Re-check register
| Item | Status as of 15 Oct 2026 | Next known step |
|---|---|---|
| Tesla appeal (No. 26-10858) | Pending. Latest docket retrieved is a 4 May 2026 snapshot. | Check the Eleventh Circuit docket |
| Bartz payments | Approved 20 Jul 2026. Class counsel: no payments before end of 2026; installments through Sep 2027. | Settlement website; docket |
| Robodebt payments | Administrator expects to start before end of 2026; none confirmed. | Department of Social Services; Gordon Legal |
| Air AI order | Proposed; court entry not verified. | D. Ariz. docket |
| RealPage class settlements | Awaiting approval; claims open to 29 Jan 2027. | Settlement website |
| RealPage state case, UnitedHealth, Workday | Pending; no liability findings. | Dockets |
Verification log
What was actually retrieved on 15 October 2026, so the dates above cannot be over-read.
- Matched to a primary document that day: Air Canada (the tribunal decision); Knight Capital (the company’s 8-K and the SEC press release); Zillow (the Q3 2021 10-Q and the 2 November 2021 8-K); Meta (the Texas Attorney General’s release); Tesla (the Eleventh Circuit docket, a 4 May 2026 snapshot); Robodebt (the Department of Social Services release, for the payment timing only).
- Matched to secondary reporting or legal commentary that day: Gauthier, Mata, Wadsworth, Coomer, Noland, Lee Luda, Clearview, MD Anderson, GM and Cruise, Bard, Bartz’s current status.
- Carried over unchanged from the 12 October Ledger, not re-retrieved: Whiting, Couvrette, Deloitte, DoNotPay, iTutorGroup, Cruise’s penalties, Earnest, Air AI, Tesla’s verdict amounts, Zillow’s full-year US$407.9M, Robodebt’s components, Bartz’s work counts, RealPage and all watchlist rows.
- Corrected in this edition: Air Canada’s bereavement-fare estimate is C$979.48 (decision text), not C$949.48 (as in some press reports and in the earlier Ledger and HTML version). Zillow’s headline is now the Q3 figure that was matched to a filing, with the full-year figure flagged. Bartz’s payment timing now follows class counsel, not a single-source date.
- Not retrieved in either edition: the Whiting opinion, the Couvrette fee order, the MD Anderson audit, the Hansard or AusTender record for Deloitte, and the Dutch and Korean decisions themselves.
<a id=”left-out”></a>
Considered and left out
- Nippon Life v. OpenAI and the PocketOS coding-agent deletion were in the earlier version and are removed here: each rested on a single trade source and had no cost figure beyond amounts sought or none disclosed.
- Google’s US$1.375B Texas settlement (May 2025) bundles location tracking, Incognito mode, and biometric claims, so the AI-related share cannot be separated.
- The Dutch childcare-benefits scandal carries a reported compensation estimate near €14B, but the source retrieved does not describe an algorithm or risk-profiling system, so an algorithmic link is not established here.
- Single-outlet figures (for example, a Cruise victim settlement of “at least US$8M”) are noted inside rows and never counted.
FAQ
What is the most expensive AI failure in this database? Among payments, the largest are the Anthropic training-data settlement (US$1.5B) and Texas’s biometric settlement with Meta (US$1.4B); neither is a model-accuracy failure. Among failures of the AI system itself with a priced cost, the largest is Zillow’s Q3 2021 write-down of US$304.4M, with a reported full-year figure of US$407.9M. The two larger numbers, GM’s Cruise investment and Bard’s market-value drop, are not payments.
How much did the Air Canada chatbot cost? C$812.02: C$650.88 in damages, C$36.14 in interest, and C$125 in tribunal fees, ordered on 14 February 2024.
Can I add the costs up? No. They mix currencies and kinds of money, and some figures overlap or are proposed. The database never totals them.
Why do some rows have no cost? Their cost is undisclosed, unmeasured, or only alleged. They stay because they carry other evidence, such as the Epic Sepsis accuracy measurements, or because pending litigation can still change the ledger.
How often is this updated? Rows marked “Can change” are re-checked when a docket or payment date passes. Each row’s Checked entry says when and how it was last verified. This is research, not legal or financial advice; check dockets and primary documents before relying on a figure in a filing.