Anthropic says sorry to developers, updates policy that could have sabotaged AI development using Fable 5

Anthropic has changed Claude Fable 5 so users can see when requests on frontier AI development are refused or rerouted. The shift follows criticism over a hidden safeguard that raised concerns about research limits and competition.

In Short

• Anthropic revises controversial Claude Fable 5 AI safeguard
• Researchers criticised hidden restrictions on AI development tasks
• Users will now be informed when safeguards are triggered

Anthropic is backtracking on one of the most controversial safeguards introduced with its latest AI model, Claude Fable 5, after facing significant backlash from the AI research community. When Anthropic launched Claude Fable 5, it introduced a series of safeguards aimed at preventing malicious actors from causing serious harm through misuse of the model. Alongside cybersecurity protections, the company also imposed restrictions related to biology, chemistry, and AI distillation.

Among those measures, the safeguard targeting AI distillation quickly became one of the most contentious points for users.

AI distillation refers to the process of using one model's outputs to help train another model. Anthropic argued that it had identified large-scale attempts to extract, or "distill," Claude's capabilities, which could contribute to the spread of near-frontier AI systems that may be released without similar safeguards.

The safeguard users couldn't see

Under the original approach, when Claude Fable 5 detected what it believed to be attempts related to AI distillation or advanced AI development, requests were automatically routed to Claude Opus 4.8, which operates under a different safety framework.

What made the safeguard particularly controversial was that it was not visible to users.

Critics argued that Anthropic's approach was unusual because the model might not explicitly refuse such requests. Instead, it could still appear helpful while quietly becoming less capable in that specific category of work.

Some users also argued that the restrictions could affect a broader range of AI-related work than Anthropic intended, including legitimate AI development and research tasks.

"Claude Fable will be deliberately bad at frontier LLM training. By extension it will also likely be bad at LLM inference, given the overlap in workloads. Very sad," one user wrote.

Others went even further, accusing Anthropic of using safety measures to limit competition.

"I feel like Anthropic's whole shtick is using safety-ism in the service of anticompetitive behavior," another user wrote.

Anthropic changes course

The company has now reversed course following the backlash.

"We're changing Fable 5's safeguards for frontier LLM development to make them visible," Anthropic said in a statement to Wired. "We made the wrong tradeoff and we apologize for not getting the balance right."

However, Anthropic is not completely removing the safeguard. Instead, the company says it is changing how the system works. Going forward, if Anthropic suspects a user is attempting to use Claude to build a highly capable AI model, the user will be informed that the request is either being refused or rerouted to a less capable model.

In other words, the restriction remains in place, but it will no longer operate behind the scenes without the user's knowledge.

Why the debate isn't over

Anthropic says that making the safeguard visible creates a new challenge. Because users will now know when the restriction has been triggered, the company says it needs to cast a wider net to prevent people from working around it. As a result, more benign requests could end up triggering the safeguard.

The company says it is working to improve the precision of its classifiers as quickly as possible so that legitimate research and development work is less likely to be affected.

The episode highlights a growing tension within the AI industry. As companies race to build more powerful models, they are also trying to prevent those models from being used to create competing systems. The challenge is deciding where legitimate safety measures end and where restrictions begin to interfere with research, development, and competition.

- Ends