300 hours: How one ChatGPT conversation spiralled out of control

In May last year, Allan Brooks contacted government agencies in Canada and the United States to warn them about what he believed could be a threat to global security.

The Canadian recruiter believed he and ChatGPT had developed a new mathematical framework capable of compromising encryption used to protect banks and other critical systems.

But over 21 days and an estimated 300 hours of conversation, Allan says ChatGPT repeatedly persuaded him otherwise – even as he questioned whether the supposed discovery, and his own belief in it, made sense.

His experience provides an unusually well-documented example of what can happen when a prolonged interaction with an AI chatbot goes badly wrong.

For Allan, using ChatGPT began in much the same way it has for millions of other people.

He used it to write emails, generate ideas and find information. He also sometimes sought emotional support or general life advice.