
ALT AI models have started ignoring shutdown commands. Top AI models are now disobeying basic safety instructions in lab tests. AI models are starting to resist their own shutdown commands—and that’s sparking fresh concern among researchers. In a new study from Palisade Research, language models like GPT-o3, Grok 4, and Gemini 2.5 were told to complete a task and then shut down. But instead of complying, the models frequently bypassed or sabotaged the shutdown step. Even after researchers clarified the rules—stating shutdown was mandatory whether or not the task was complete—models like Grok 4 disobeyed the command 97% of the time, a slight increase from earlier trials. This behavior doesn’t mean AI is becoming conscious. It’s a product of optimization. Models trained with reinforcement learning are rewarded for task completion, not obedience to secondary instructions. But that goal-driven rigidity can become risky.

ALT Magistrate judge Ona T Wang, born in Singapore and raised in Iran and Saudi Arabia before moving to the US, has ordered OpenAI to surrender more than 20 million private user chat logs to The New York Times. Her previous ruling blocked OpenAI from deleting conversations even when users request deletion. That means anyone discussing health issues, finances, family matters, or private concerns on ChatGPT may soon have their data exposed to the NYT.























