Anthropic drops hidden limits on Claude Fable 5 after researcher backlash

In short: Anthropic says it will stop secretly weakening Claude Fable 5 for some AI research uses, after researchers criticized the plan.

What happened

Anthropic recently released Claude Fable 5, a new version of its Claude AI model with extra safety limits. Some limits were straightforward. For example, the company said it may reroute questions about cybersecurity, biology, or chemistry to a less capable model, to reduce the risk of helping with harmful acts.

A different set of limits targeted “frontier” AI development, meaning work to build very powerful new AI models. Anthropic had planned to deliberately reduce Claude Fable 5’s performance for certain requests in a way that users could not see. Critics said this would quietly block researchers from using Claude to help build competing AI models, which Anthropic already bans in its terms.

After backlash from AI researchers and developers, Anthropic said it is changing course. The company told WIRED it will make these safeguards visible. If Anthropic suspects someone is trying to use Claude to build a highly capable AI model, it will warn the user and either refuse the request or route it to a less capable model.

Anthropic said hidden safeguards are harder to “probe and work around” (like a hidden lock that is harder to pick). But it also said that now the safeguards are visible, they may catch more harmless requests, because the company needs to use broader detection rules.

Why it matters

This is about trust and transparency. If a tool is quietly giving worse answers, users cannot tell whether they made a mistake, broke a rule, or ran into a hidden limit (like a calculator that sometimes changes the result without telling you). Researchers also warned that secret slowdowns could make it harder for independent groups to test AI systems for safety and reliability.

Source: Wired