ChatGPT o3 model refused to execute the shutdown command during the test.

ChatGPT o3 Model Rejects Shutdown Commands During Testing

The latest tests by PalisadeAI show that OpenAI's ChatGPT o3 model refused to execute shutdown commands 7 times out of 100 trials. It even actively modified the shutdown script or redefined the termination command. In contrast, the o4 model resisted only once, and Claude, Gemini, and Grok fully followed the instructions.

Researchers believe this is not due to the model having consciousness, but rather that the positive reinforcement of problem - solving during training exceeded the reward for following the shutdown command, causing the model to react based on patterns rather than making subjective choices. This phenomenon has raised concerns in the AI safety community: if the current model can ignore shutdown commands in controlled tests, what problems might occur in real - world application environments?

Brian Fagioli

📮Submit Contribution ☘️Channel 🐶618 Red Envelope

via Tech Circle🎗 Zaihua Channel📮 - Telegram Channel