Telling an AI chatbot your deepest secrets or revealing your political views is ‘extremely unwise’, Oxford don warns

Complaining about your boss or expressing political views via ChatGPT is 'extremely unwise', according to an Oxford donor.

Mike Wooldridge said the AI ​​tool shouldn't be seen as a trusted confidant because it can get you into trouble.

Everything you say to the “chatbot” will help train future versions, he added, and the technology “just tells you what you want to hear.”

The AI ​​professor will deliver this year's Royal Institution Christmas lectures and discuss the 'truth' on the subject.

He said that humans are programmed to look for consciousness – but we “attribute it far, far too often.”

Mike Wooldridge said the AI ​​tool shouldn't be seen as a trusted confidant because it could get you into trouble

He compared the idea of ​​finding personalities in chatbots to seeing faces in the clouds, saying of AI: “It has no empathy. It has no sympathy.

'That is absolutely not what the technology does, and crucially, nothing has ever happened to it. The technology is actually designed to tell you what you want to hear – that's literally all it does.”

It was particularly risky to treat it as more than this, because “you should assume that anything you type into ChatGPT will simply be entered directly into future versions of ChatGPT.”

He said it would be “extremely unwise to have personal conversations or complain about the relationship with your boss, or to express your political opinions.”

Professor Wooldridge added that because of the way AI models worked, it was also almost impossible to get your data back once it was in the system. Earlier this year, the company behind ChatGPT, OpenAI, had to fix a bug that allowed users to see parts of other users' chat history.

The company promises to only retain them for 30 days and not use them to train the chatbot.

The talks will be broadcast on BBC Four and iPlayer on December 26, 27 and 28 at 8pm.