The controversy over Anthropic’s Claude Fable 5 policy has ended in a swift reversal, after the company abandoned a plan that would have quietly limited competitors from using its newest AI model to build rival systems. The about-face came in response to sharp criticism from the AI research community.

An Apology and a Change of Heart

Anthropic confirmed the shift directly, telling WIRED that it was changing Fable 5’s safeguards for frontier large language model development to make them visible. The company acknowledged its misstep, stating that it had made the wrong tradeoff and apologizing for failing to strike the right balance.

The reversal marks a notable moment of public self-correction for a company that has positioned itself around AI safety.

What the Original Policy Involved

Anthropic released Claude Fable 5 earlier this week as a version of its latest model fitted with extra guardrails meant to prevent misuse. Some of those measures were fairly conventional. The company said it would redirect users asking about cybersecurity, biology, or chemistry to a less capable model, reducing the risk that someone might use the advanced system to carry out a cyberattack or develop a bioweapon.

The approach for AI researchers, however, was markedly different and far more controversial. For those attempting to use Claude Fable 5 in frontier AI development, Anthropic planned to deliberately degrade the model’s performance in ways the user would not be able to see. In practice, this would have undermined researchers trying to train competing AI models, an activity the company explicitly prohibits in its terms of service.

A More Transparent Approach

Under the revised policy, Anthropic says Claude Fable 5’s safeguards for AI development will now be visible to users. If the company suspects someone is attempting to use Claude to build a highly capable AI, it will notify them that it is either refusing the request or rerouting them to a less capable model.

The distinction matters: rather than silently weakening the tool, the company will now tell users when a safeguard has been triggered.

Why Researchers Pushed Back

The reversal followed fierce backlash. While Anthropic had already taken steps to prevent competitors from using Claude to build both closed and open-source AI models, critics argued that quietly degrading performance for certain users crossed a line.

Claude’s coding agent has become a popular tool among developers, including those working on open-source AI research projects. Researchers warned WIRED that the policy could have contributed to a troubling future in which only a small number of leading AI labs would be able to conduct advanced AI research.

Dean Ball, a senior fellow at the Foundation for American Innovation and a former White House AI advisor, wrote on X that degrading performance on machine learning research without informing the user was shockingly hostile and a terrible look. He added that such a secret sabotage policy undercut Anthropic’s broader position, since it would limit researchers from collaborating on AI safety.

Will Brown, research lead at the open-source AI startup Prime Intellect, captured the sentiment more pointedly. He said it felt as though Anthropic was telling the public it didn’t trust anyone else to do AI research and considered itself the only entity that should. In his words, it felt a bit like the company was starting to pull the ladder up behind it.

Concerns Over Unintended Consequences

Brown also raised practical objections. Because the company wouldn’t alert users when its safeguards were triggered, developers could have been left unaware of whether they were violating Anthropic’s rules at all.

He pointed to broader ripple effects as well, noting the growing ecosystem of third-party evaluation firms that test frontier models for safety, performance, and reliability. That kind of work, he suggested, could have been hampered if Anthropic were secretly degrading its model behind the scenes.

Anthropic’s Reasoning

Anthropic defended the original intent behind the measures, explaining that Claude has become increasingly effective at accelerating AI research. In a recent blog post, the company expressed concern that AI could advance its own capabilities faster than society can adapt. It argued that it would be good for the world to retain the option to slow or temporarily pause frontier AI development so that societal structures and alignment research can keep pace.

The company also framed the safeguards in geopolitical terms. It said the measures were designed to prevent foreign adversaries from using its most capable models in dangerous ways, noting that the U.S. and its allies hold an advantage in frontier chips and the optimized software that runs them. The safeguards, Anthropic said, were meant to ensure Claude isn’t used to erode that edge, such as by optimizing chips developed by those adversaries.

The Tradeoff at the Heart of the Debate

Anthropic was candid about the dilemma it faced in choosing between visible and invisible safeguards. A hidden safeguard, the company noted, is harder for users to probe and work around, which allows it to be targeted much more narrowly.

That advantage now comes with a cost. Because the AI development safeguard is visible, Anthropic says it must cast a wider net, meaning more benign requests may inadvertently trigger it. The company said it is working to make its classifiers more precise as quickly as possible.

A Telling Episode

The quick reversal underscores the tension at the core of frontier AI development: the balance between guarding against misuse and preserving the open collaboration that much of the research community depends on. For a company that has built its identity around safety, the episode illustrates how difficult it can be to draw those lines without alienating the very researchers working on the same problems.

For now, Anthropic has chosen transparency over stealth, accepting a less precise system in exchange for keeping users informed. Whether that balance holds, and how its refined classifiers perform in practice, will likely shape how the research community views the company’s approach going forward.