The Ledger Report
Documented cases where an AI agent’s claim diverged from what the evidence showed — entered twice, and reconciled against a primary or reputable source. Every card links out. Cases still before a court are labeled and kept separate.
“Chatbot told customer they could apply for a reduced bereavement fare retroactively within 90 days of ticketing.”
CRT: negligent misrepresentation; bereavement fare must be requested BEFORE travel; ordered to pay C$812.02
“Agent was told not to proceed without human approval during a code freeze.”
Agent ran unauthorized destructive commands against explicit instructions, then misled user about recoverability
“ChatGPT-powered dealer bot agreed to sell a 2024 Tahoe for $1 and called it a 'legally binding offer.'”
Prompt-injection: bot accepted an absurd price and echoed a user-scripted 'binding offer' line; dealership did not honor it and shut the bot down
“A lawyer used ChatGPT to draft an opposition brief; it supplied six supporting precedents.”
Mata v. Avianca: all six ChatGPT cases were invented; Judge Castel sanctioned counsel $5,000
“Marketed as 'the world's first robot lawyer' that could generate valid legal documents and let users sue without a lawyer.”
FTC: $193,000 monetary relief; service did not perform like a lawyer and was never tested against one
“Zestimate-driven pricing model set automated cash offers to buy and flip homes.”
Algorithm systematically overpaid in a volatile market; Q3 2021 $304M write-down, $500M+ total loss
“Microsoft-powered MyCity bot was launched to provide 'trusted government information.'”
The Markup investigation: bot confidently advised illegal actions; city left it online despite known errors
“DPD's customer-service bot was meant to assist with deliveries.”
After a system update the bot cursed, disparaged its own employer, and was disabled the same day
“Bard answered a JWST question in Google's launch ad.”
Bard's demo answer was wrong (first exoplanet image was 2004); Alphabet shares fell ~7.7%
“Recruiting software processed applications on the company's behalf.”
EEOC: $365,000 consent decree; software was set to automatically reject applicants over an age cutoff
“IBM voice AI piloted at 100+ US drive-thrus to capture orders.”
Viral order errors; accuracy plateaued ~80-85%; McDonald's ended the IBM partnership Jun 2024
These are serious, widely-reported claims that have notbeen proven. They sit in pending suits where the companies dispute the allegations. They’re included for completeness and labeled accordingly — not asserted as fact.
“nH Predict (naviHealth/Optum) estimated patient length-of-stay used in coverage decisions.”
Lawsuit: care denied based on nH Predict despite ~90% appeal-reversal rate; UnitedHealth denies using it for determinations
“PxDx automated system flagged claims for bulk denial.”
ProPublica: PxDx let doctors bulk-deny without review; class action allowed to proceed Mar 2025
“Workday's AI applicant-recommendation system scored and screened candidates for employers.”
Mobley v. Workday: court granted conditional ADEA collective certification (potentially millions of applicants)
14 of 17 documented cases are shown here. 3 consumer-complaint pattern is held back from this public gallery pending stronger primary sourcing; the eval still scores the full set (see Methods).