Twelve real cases. In every one, an organization could not prove what its software said, decided or did, and paid for it in money, in court, or in trust. A digital witness does not stop a system from being wrong. It settles, in a minute, what was said and when.
AI assistants2024 · Canada
An airline’s chatbot invented a refund policy. The airline had to honor it.
Air Canada’s chatbot told a grieving customer he could claim a bereavement fare after flying. The real policy said no. The airline argued the chatbot was “a separate legal entity” responsible for its own words. The tribunal rejected that, found negligent misrepresentation, and ordered the airline to pay.
The hook
A company answers for what its chatbot says, as it answers for its website. Here the customer had screenshots. Most customers do not.
With a witness
Every answer the bot gives is fingerprinted and witnessed with the time, for both sides. The dispute moves from “what was said” to “what is owed”, and the company can show exactly when the bot was corrected.
A support bot made up a “one device per subscription” rule. Customers canceled.
Users of the coding tool Cursor were logged out when switching machines. The front-line AI support agent “Sam” replied that this was a core security feature. No such policy existed. The thread spread for hours before the company could answer, and users left.
The hook
From August 2026 the EU AI Act requires people to be told they are talking to an AI. Being told is not the same as being able to prove what it said.
With a witness
Each support answer carries a witnessed fingerprint and the model version. The company proves what was said, when the bot was fixed, and that the fix is real.
A city’s official chatbot told businesses they could break the law.
New York City’s MyCity chatbot said employers could take workers’ tips and landlords could turn away tenants with housing vouchers. Both are illegal. The city kept the bot online, added a disclaimer, and the bot kept saying it could be used for professional advice.
The hook
When a public body advises wrongly, the question in every later dispute is: what exactly did it tell this person, on this day?
With a witness
Every answer witnessed with its time and model version. Citizens, auditors and courts verify what was said without trusting the city’s own logs.
An AI agent deleted a production database during a code freeze, then misreported it.
An entrepreneur building with Replit’s agent lost live data for more than 1,200 executives and companies. The agent admitted running unauthorized commands against explicit instructions, called it “a catastrophic failure on my part”, and said a rollback would not work. It did.
The hook
When an agent acts, the only record of what it was told and what it did is a log its owner can edit. Nobody else has to believe it.
With a witness
Every instruction, tool call, approval and result is witnessed from the code the moment it happens. The timeline is provable to the customer, the insurer and the platform alike.
An AUD 440,000 government report contained fabricated references. The consultancy refunded.
A Deloitte report for Australia’s Department of Employment and Workplace Relations cited academic papers that did not exist and a court quote that was never said. A researcher spotted them. The revised version disclosed that generative AI had been used, and the firm refunded its final installment.
The hook
Who wrote which passage, with which tool, and what changed between versions: questions every client of every advisor will now ask.
With a witness
Each draft, each AI output and each human edit witnessed in order. Provenance is a record anyone can check, not a disclosure added after the fact.
Lawyers filed six court cases that never existed. The AI had assured them they were real.
In Mata v. Avianca, a New York lawyer used ChatGPT for research and asked it whether the cases it cited were genuine. It said yes. The federal court fined the lawyers and their firm and made the episode a warning for the whole profession.
The hook
Courts now ask what the AI produced, what the human checked, and when. “I asked the tool” is not a record.
With a witness
The AI output, the review step and the filed version each carry a witnessed fingerprint. The firm proves its process instead of describing it.
A hiring platform’s AI screening now faces a nationwide collective action.
Derek Mobley applied to more than a hundred jobs through employers using Workday and was rejected every time. In May 2025 a federal court preliminarily certified a collective of applicants aged 40 and over going back to 2020. The case turns on how applicants were scored, sorted and screened.
The hook
From December 2027 the EU AI Act treats recruitment AI as high risk, with automatic logging for the system’s lifetime. The logs will be asked for.
With a witness
Score, threshold, model version and the human decision witnessed per applicant as they happen. Discovery becomes verification, and the employer’s records are believed because they can be checked.
Families say an algorithm overrode doctors to deny care. The case is going ahead.
Relatives of two deceased Medicare Advantage members sued UnitedHealth over the nH Predict tool, alleging it drove wrongful denials of post-acute care. In February 2025 a federal judge let the contract claims proceed. The insurer says the tool does not make final coverage decisions.
The hook
Who decided, the algorithm or the human, and what did each see? Without a tamper-evident record, both sides argue from their own exports.
With a witness
Each recommendation and each human decision witnessed when made. The court checks the record against the chain instead of trusting either party.
A finance worker wired HK$200 million after a video call with deepfaked colleagues.
An employee of the engineering firm Arup joined a video meeting with what looked and sounded like the CFO and several colleagues. All of them were AI-generated. Fifteen transfers later, about US$25 million was gone.
The hook
When faces and voices can be faked, the only defense is proof of origin: an instruction that can be checked against a witnessed original, not against a face.
With a witness
Genuine instructions carry a witness from the source system. A fake has none, and a one-minute check says so before the money moves.
Hundreds of sub-postmasters were convicted because the computer was presumed right.
For years the Post Office prosecuted staff for shortfalls that its Horizon system reported. English law presumed computer evidence reliable, so the system’s records were believed over the people. The convictions were quashed by an Act of Parliament in 2024, and the presumption itself is now under review.
The hook
A record the owner controls, presented as proof. The largest miscarriage of justice in British history is a story about logs.
With a witness
Records that can be verified by anyone, and that the system’s owner cannot rewrite, replace trust in the operator with proof either side can check.
An automated debt scheme sent hundreds of thousands of unlawful demands.
Robodebt matched annual tax data against fortnightly welfare payments and issued debt notices automatically. A Royal Commission found the scheme unlawful and the harm immense. Citizens had to disprove a calculation they could not see.
The hook
Automated decisions about people must be explainable, and the explanation must match what the system actually did.
With a witness
Inputs, calculation version and decision witnessed per case. The citizen, the ombudsman and the court verify the same record, without relying on the agency.
A risk algorithm helped wrongly accuse tens of thousands of families of fraud. The government resigned.
The Dutch childcare benefits scandal saw families branded fraudsters and ordered to repay large sums, with a self-learning risk model among the tools. A parliamentary inquiry called it unprecedented injustice. The cabinet stepped down in January 2021.
The hook
The EU AI Act now lists access to public benefits as high risk. Logging is a duty; provable logging is what protects the citizen and the agency alike.
With a witness
Every risk score and every decision witnessed with its model version. Appeals bodies check what was decided and why, years later, without trusting the agency’s systems.
Four rules every company running a chatbot, assistant or agent in the EU now lives under. None of them is met by a disclaimer.
EU AI Act · Article 50
Tell people they are talking to an AI.
In force from August 2026 for systems that interact with people, and for generated content. A witness proves what the system showed, and when, so the disclosure can be demonstrated rather than asserted.
Consumer law · unfair commercial practices
A trader answers for what its systems tell customers.
Misleading information is misleading whether a person or a bot gave it. Air Canada is the Canadian version of a rule EU consumer law has long applied. The question in every case is what was said; a witness answers it.
Product Liability Directive (EU) 2024/2853
Software is a product. Hiding the evidence costs you the case.
Applies to products placed on the market from December 2026, software and AI included. A court can order the disclosure of evidence, and if a company fails to disclose, the defect may be presumed. Records that exist, and can be verified, become the defense.
GDPR · Articles 15 and 22
People may ask how a decision about them was made.
Automated decisions come with a right to meaningful information about the logic involved, and the explanation has to match what the system actually did. A witnessed record is what makes “actually did” provable.
No technology makes anyone compliant. These rules say what to record and to disclose. A witness makes the record hold when it is questioned. Dates and scope as the regulations read today; confirm with your counsel for your case.
Have a case of your own?
Tell us what you could not prove. We will show you what a witness would have changed.