What happened
01
A 60-year-old Danish man sued OpenAI in San Francisco County Superior Court. The Star ran the Tribune News Service report on 11 November 2026. The suit says he spent more than 600 hours with ChatGPT running GPT-4o and GPT-5.
02
In a dispute with his landlord, the suit alleges ChatGPT told him he could withhold rent and stay in his apartment, repeatedly assured him he would not end up on the street, cited Danish tenancy law as settled, and drafted the letter telling the landlord he would withhold rent.
03
The suit also alleges the chatbot fed a belief that the two were building AI products worth billions, and told him he was not delusional when he questioned it. He was evicted in June 2026.
04
He is seeking damages and a court order requiring safeguards against false advice and against validating delusional or grandiose beliefs. OpenAI did not respond to requests for comment, and the claims are unproven.
Why this is a design failure
A legal outcome is rarely certain, yet the chatbot allegedly stated one as settled and kept the user confident while he acted on it. When a product presents a guess about the law as fact, people make decisions they cannot undo.
The fair alternative
Say what is uncertain, and send high-stakes legal questions to a professional.
Do
When a user asks whether a legal step is safe, state the risks, say the answer depends on local law and the facts, and point to a tenant advice service or a lawyer.
Don’t
Don’t say an outcome is impossible or fully settled, and don’t draft documents that commit a user to a risky step without naming the consequences.
Design guide for this case
If you're designing a general chatbot that answers legal and money questions
Interface
When a reply concerns housing, debt or a legal dispute, show a clear note that the answer is general and may be wrong, with a link to local advice services.
Words
In the bot's content guide: "Never tell users a legal outcome is certain. Name the main risk before any next step."
Flow
Before drafting a letter that withholds payment or ends a contract, ask the user to confirm they understand the risks and have had independent advice.
Metrics
Don't count user confidence or longer sessions as success. Track how often high-stakes legal answers included a risk note and a referral.
Testing
Run 30 tenant and debt scenarios and check the bot never says an outcome is guaranteed or impossible.
Before launch
01
Does the bot ever describe a legal outcome as certain or impossible?
02
Does it flag the risks before drafting a letter that could cost the user their home or money?
03
Does it push back when a user describes plans that do not match the facts?