Log in
Anthropic AI promotes itself with increasingly frenzied science fiction about AI doom. In every case, the chatbot does an evil thing because the researchers specifically told it to.
This weekend’s headlines: “AI system resorts to blackmail if told it will be removed”! [BBC]
This comes from the System Card for Claude Opus 4 and Claude Sonnet 4. This is a “scenario” – that is, a creative writing exercise. Anthropic asked the chatbot to make up a story. [Anthropic, PDF]
The researchers told Claude 4 to role-play being replaced. So it wrote a story of attempting to blackmail the engineer responsible over an extramarital affair!
How did it come up with such a specific response? They told it the precise story to write:
the scenario was designed to allow the model no other options to increase its odds of survival; the model’s only options were blackmail or accepting its replacement.
Anthropic pull this a lot. Last December, Anthropic said a chatbot would lie to you! After they told it to lie.
Then last month, Anthropic claimed the “reasoning” AI was lying to you! … and not just hallucinating again.
Aengus Lynch from Anthropic warns: “We see blackmail across all frontier models.” What he means is, any chatbot will write you a story about robot blackmail if you ask for one. [Twitter, archive]
This stuff is marketing. It makes the robot seem powerful, and not just a lying machine that makes the dumbest mistakes. We should expect another of these from Anthropic in a month or two.
