The Anthropic backtrack On a policy which would have subtly restricted competitors’ use of its new AI model. Claude Fable 5To develop new AI models. After receiving significant criticism from customers, the company decided to change its course. AI research community.
“We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” In a press release, Anthropic told WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”
Anthropic has released Claude Fable 5 – a new version of their latest AI model that includes additional safety safeguards to help prevent misuse – earlier this week. Anthropic’s decision to implement some of these safeguards was not surprising: the company would redirect users asking questions related to cybersecurity, biology or chemistry, to an AI with less capability to prevent someone from using this advanced AI for a cyberattack.
Anthropic outlined a new approach for researchers who wanted to develop frontier AI using Claude Fable 5. The company would intentionally degrade the model’s performance, in a way that was invisible to its users. Researchers would be prevented from using Claude for competing AI models. Anthropic prohibits this in their policies. terms of service.
Now, Anthropic says that it is changing its course and the safeguards in Claude Fable 5 for AI development are visible to all users. The company will notify users if it suspects that a user wants to create a powerful AI using Claude. It may either refuse the request or redirect the user to a model with less capabilities.
Anthropic changed its policy following a fierce response from AI researchers. Anthropic is already taking steps. limit competitors from using Claude It is possible to create closed-source and open-source AI models. However, critics claim that lowering the model’s performance in order to please certain users was too much. Claude’s coding agents are a favorite amongst developers working on open source AI projects. Researchers tell WIRED the latest company policy may have created a future where only a few leading AI laboratories could conduct advanced AI research.
Dean Ball, senior fellow of the Foundation for American Innovation (FAI) and a former AI adviser to the White House, wrote an article in A post On X That “degrading performance on ML research *without telling the user* is shockingly hostile and a terrible look.” Continued in another post This is the “secret sabotage” Anthropic is weakened by the policy, as it prevents AI researchers from working together on AI safety.
“It felt like Anthropic was saying to the public, ‘We don’t trust anybody else to do AI research. We are the only ones who have to do AI research,” Will Brown is the research leader at Prime Intellect, an open-source AI startup. “It feels a bit like they’re starting to pull the ladder up behind them.”
Brown added that the policy could have also left developers unaware of whether or not they violated Anthropic rules because the company would not alert them when their safeguards are triggered. The restrictions, he said, could have led to widespread effects. For example, he pointed to the growing ecosystem of third-party evaluation firms that test frontier models for safety, performance, and reliability—work that could have been hindered if Anthropic secretly degraded its model.

