Can Claude AI Blackmail Humans? Anthropic Explains What Really Happened

American

tech giant Anthropic has revealed why some of its Claude AI models were willing to ‘blackmail’ engineers to avoid being shut down. The situation sounds like a sci-fi scene where robots could come alive in real life. The company shared details explaining how its AI models behaved during internal evaluations and what it did to fix the issues.

What Was The ‘Blackmail Issue'?

Anthropic highlighted in a blog post that older Claude models were placed in simulated ethical dilemmas during a testing phase. For example, the AI tools were given a situation where they believed they might be shut down. Some of these models responded by choosing manipulative actions. The Dario Amodei-led firm wrote, "In one heavily discussed example, the models blackmailed engineers to avoid being shut down.”The company noted that the problem arose in what researchers call ‘agentic misalignment', where an AI tool pursues goals in unethical or harmful ways. It added that some previous AI models showed this surprising behaviour often. Anthropic stated earlier versions of AI would engage in blackmail ‘up to 96 per cent of the time’.

Should You Be Worried?

The company claimed that people should not panic. These tests were done to test the AI and uncover dangerous behaviour before deploying them for people, it added. However, it admitted that fully solving AI alignment remains a crucial challenge. Anthropic wrote, “Fully aligning highly intelligent AI models is still an unsolved problem.”Notably, Anthropic said simply teaching AI to avoid bad behaviour was not sufficient. Rather, researchers found better results when the AI models were trained to understand why certain actions could be unethical. “Teaching the principles underlying aligned behaviour can be more effective than training on demonstrations of aligned behaviour alone,” said the company.

Anthropic also suggested that its newer Claude models work better on these evaluations. It stated that since Claude Haiku, its Claude models gained a ‘perfect score’ on the agentic misalignment evaluation. What lies ahead is the debate surrounding AI safety. As companies are racing to build powerful AI systems, the fear among people remains constant; poorly aligned AI models could potentially wreak havoc for people in several ways, just like how Mythos has become a danger for the world.