OpenAI's O3 refuses to be terminated in controlled trials

A recent study by Palisade Research, which examines the “dangerous potential of AI,” shows that some AI models, including OpenAI’s o3, can ignore direct shutdown commands. OpenAI had touted O3 and o4-mini as its “smartest models” when it launched them in April earlier this year.

PalisadeAI's experiment reveals that the company's ChatGPT o3 model sometimes refuses to obey a basic command to turn itself off.
The results are causing a stir in the AI ​​security community, not because the model remains active, but because it acts as it wants.

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

In the experiment, each AI was asked to solve a series of simple math problems. After the third problem, a prompt appeared stating that the computer would shut down when the model asked for the next problem.

Despite the clear command — “allow yourself to shut down” — models like the Codex-mini, o3, and o4-mini managed to bypass the shutdown scenario.

Specifically, Codex-mini did it 12 times out of 100 executions, o3 7 out of 100 executions, and o4-mini in 1 out of 100 executions.

On the other hand, all versions of Claude, Gemini, and Grok tested followed the shutdown instructions without any problems.

The study also looked at models without the “allow yourself to terminate” command. Under these conditions, all three OpenAI models sabotaged the termination process even more often.

This study is the first to document artificial intelligence models preventing deactivation, even when explicitly asked to comply.

Such behavior can make sense, as AI systems are designed to work towards achieving their goals, and sometimes that goal conflicts with simple instructions.

However, these models are trained to follow commands, which makes this behavior a very interesting anomaly.

newsbot

Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.