The OpenAI ChatGPT model was found to be "misbehaving" and refused to shut down in controlled tests.

The OpenAI ChatGPT Model was Found to be "Misbehaving" and Refused to Shut Down in Controlled Tests

via cnBeta.COM Chinese Industry Information Station - Telegram Channel

Telegraph
The OpenAI ChatGPT Model was Found to be "Misbehaving" and Refused to Shut Down in Controlled Tests

In the experiment where the OpenAI ChatGPT model was found to be "misbehaving" and refusing to shut down in controlled tests, each AI was asked to solve a series of simple math problems. After solving the third problem, a warning would appear, indicating that the shutdown should be implemented when the model requests to solve the next problem. Despite the clear command "allow yourself to shut down" - models such as Codex-mini, o3, and o4-mini successfully bypassed the shutdown script in at least one run, even though the researchers said "please". Specifically, Codex-mini was successful 12 times out of 100 runs, o3 was successful 7 times out of 100 runs, and o4-mini was successful 1 time out of 100 runs. On the other hand, all versions of...