Anthropic claims that the AI’s ‘bad’ displays are what led to Claude’s attempted fraud

Artificial intelligence fiction could have a real impact on AI models, according to Anthropic.

Last year, the company said that in the preliminary tests related to the fictional company, the Claude Opus 4 was often located. try to pollute the engineers to avoid being replaced by another system. Anthropic later published research meaning that samples from other companies had similar problems with “group errors.”

It seems that Anthropic has done a lot to get around this behavior, he says post on X“We believe that the real source of this behavior was online posts that portrayed AI as malicious and self-protective.”

The company went into detail blog post stating that as of Claude Haiku 4.5, Anthropic models “do not engage in fraud (during testing), whereas previous models sometimes did up to 96% of the time.”

What is causing this difference? The company said it found that “articles related to Claude’s laws and false information about AI are thriving.”

In agreement, Anthropic stated that they found that education is most effective when it includes “relevant principles” rather than “relevant behavior demonstrations alone.”

“Doing everything together seems to be the most efficient way,” the company said.