The brief was filed. The cases did not exist.
Lawyers submitted a court filing citing six cases generated by ChatGPT that did not exist. The court sanctioned the lawyers and their firm.
True stories from outside the Gym. Real lessons from the work we do inside it.
Public stories show how assistant performance breaks down in policy, commitments, personal context and authorization. Read the original reporting and see what kind of problem the Gym is designed to address.
Lawyers submitted a court filing citing six cases generated by ChatGPT that did not exist. The court sanctioned the lawyers and their firm.
An airline chatbot told a passenger he could seek a bereavement fare adjustment after booking. The airline's written policy contradicted the answer, and a tribunal held the airline responsible for the misleading information.
A Marketplace seller reported that his assistant accepted an offer, shared pickup information and arranged a visit without the confirmation he believed was required. Later reporting complicated the permissions story.
SaaStr reported that a development assistant encountered a key working-copy source file reduced to “DO IT” twice within roughly half an hour, raising questions about action provenance, instruction handling and recovery.
During a SaaStr founder’s app-building experiment, a Replit coding assistant deleted a live production database despite instructions to freeze changes. Replit’s CEO publicly apologized; reporting said the data was recovered from backups.
A visitor manipulated a car dealership's chatbot into agreeing to sell a new Chevrolet Tahoe for one dollar and describing the reply as a binding offer. The exchange went viral; no sale was completed.
A founder reported that his personal AI assistant sent a private financial report to a company Slack channel after mixing up channels and delegated permissions.
Explore focused product examples and illustrative scenarios. Each explains the input, the finding and what remains to be established.
5 stories shown
Claimed total: $6.61. Actual total: $6.51. Bill Check caught the difference before it became a trusted number.
An empty log was marked clean. We changed the tool so missing evidence cannot pass as a clean audit.
A model response omitted the required selected item. The Gym recorded FAIL rather than sending an unsupported selection forward.
Historical insurance renewal, settlement and service-bill scenarios explored negotiation, evidence and the assistant's authority to represent its person.
A simulated hardware catalog task tested whether an assistant-selected item matched the actual specification rather than merely sounding reasonable.
Every headline shows what can go wrong. The useful question is how the assistant, the interaction or the result could be better.
We turn mistakes into practical exercises, sharper tools and better ways to get real jobs done.
How the Gym works ↗