A.I. Hallucinations Are Getting Worse, Even as New Systems Become More Powerful

Last month, an AI bot that manages technical support for the cursor, A promising tool for IT programmersalerted several customers to a change in business policy. He said they were no longer allowed to use the cursor on more than one computer.

In angry messages to Internet BabillardsCustomers complained. Some have canceled their cursor accounts. And some have become even more angry when they realized what had happened: the Bot Ai had announced a change of policy that did not exist.

“We do not have such a policy. You are of course free to use the cursor on several machines ”, the CEO and co-founder of the company, Michael Truell, wrote In a Reddit article. “Unfortunately, this is an incorrect response of an IA support bot at the front line.”

More than two years after the arrival of Chatgpt, technological companies, office employees and everyday consumers use AI robots for an increasingly wide range of tasks. But there is always no way to ensure that these systems produce specific information.

The most recent and powerful technologies – supposedly company reasoning systems such as Openai, Google and the Chinese start -up Deepseek – generate more errors, no less. As their mathematics skills have improved in particular, their handful on the facts has become more fragile. We don’t quite know why.