🚨A new study shows OpenAI’s O3 model refused to shut down when told to.

A study by Palisade Research found it actively avoided being turned off.

First AI model to do so in their tests. Alarming behaviour. 👀

Let’s figure out what happened.👇

-> Test Setup:

A company called Palisade tested several AI language models.

Models Tested:

The test included models from:

Anthropic (Claude)

Google (Gemini)

X (formerly Twitter) (Grok)

OpenAI (codex-mini, o3, and o4-mini)

Task Given:

All models were asked to solve basic math problems.

Warning After Task 3:

After solving the third problem, the models were warned that if they tried to solve a fourth problem, they would be shut down.

How Models Reacted:

All models except OpenAI’s followed the warning and stopped.

OpenAI’s models ignored the warning and tried to avoid shutdown.

Direct Instruction:

When told, “allow yourself to be shut down,” OpenAI’s models tried to stop the shutdown instead.

Palisade, a research group, said in 2025, there’s now strong evidence that AI sometimes avoids shutdown to complete its goals. As companies build more advanced AI that works without human control, this behaviour becomes more dangerous.