Research confirmed that when IAS are threatened, they prefer to blackmail users than being replaced
Remember that case of an engineer who was threatened by an AI? On the occasionThe technology found that it was going off and threatened to reveal that he had an extramarital affair, all to remain ‘alive’.
It all happened in a simulated environment during an Anthropic test. The developer decided to understand this phenomenon of rebellion and conducted more research involving technology in risky scenarios.
The result… it was not at all positive. The company confirmed that when AI is threatened, it really blackmails the user. This also happens when she realizes that someone is diverting her from her ultimate goal.
And it’s not just with Anthropic models. The research tested 16 models of market leaders from different developers. Deepseek, Gemini (from Google) and GPT (from OpenAi) also had worrying behaviors.
Read more:
Anthropic called this phenomenon “misalignment” and gave tips on how to protect themselves in the future.
This is the theme of the column Look from tomorrow This week, with Doctor Álvaro Machado Dias, neuroscientist, futuristic and columnist of the digital look News. Check it out!
Vitoria Lopes Gomez is writer in the digital look
