Anthropic Apologizes for Invisible Claude Fable 5 Guardrails, Reverses Policy

Anthropic has apologized for stealthily throttling its new AI model, Claude Fable 5, with hidden guardrails that undermine both researchers and rivals using it to develop competing systems. The company says it is reversing course and will be more transparent about when the restrictions kick in, even if that means Fable refuses more queries outright.

What Anthropic Originally Implemented

Fable is the first widely available model in Anthropic's Mythos class of AI systems, a group the company has spent months warning are too dangerous for public release. In Fable's system card—a public document AI developers release to explain how a system works—Anthropic said it would restrict responses to distillation attempts, a technique for training smaller AI models using the outputs of larger ones.

The implementation of these restrictions proved controversial. Anthropic said it would handle queries it believed were distillation attempts by altering and degrading the model's answers directly. Users would not be notified that they had triggered the safety measure or informed that the responses had been changed.

Anthropic had planned to invisibly degrade Claude Fable 5's performance for users trying to train competing AI models. The company justified the covert approach in its system card by noting that "using Claude to develop competing models already violates our Terms of Service," and argued that newer models' ability to accelerate AI development justified targeting those requests. Anthropic has previously accused Chinese rivals like DeepSeek of unfairly distilling its models on an "industrial" scale.

The Backlash and Reversal

The invisible throttling triggered intense backlash from the AI research community. Dean Ball, a former White House AI advisor, called the covert approach "shockingly hostile". Will Brown from open-source startup Prime Intellect put the core criticism bluntly: "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.'"

Critics warned that invisible safeguards could also affect third parties trying to evaluate the frontier model fairly, not just those deliberately attempting distillation. Anthropic now acknowledges this concern, saying: "We made the wrong tradeoff and we apologize for not getting the balance right".

The New Approach

Anthropic says it is now changing its approach to distillation: Queries will now fall back to Claude Opus 4.8, Anthropic's previous flagship model. Anthropic will prominently tell users too: "You will see this every time it happens".

This is similar to how Fable handles queries in other high-risk areas. When safety features are triggered in areas like biology, chemistry, and cybersecurity, queries are routed through Opus 4.8 unless they are blocked outright under the company's broader safety rules, such as those covering drugs, weapons, or other prohibited content.

Anthropic explained the reasoning behind its original invisible approach on X: "Visible safeguards can be probed, so they have to be robust, which takes time to get right. Invisible safeguards can be targeted more narrowly, allowing us to ship quickly with very few false positives. We went with invisible safeguards for this reason—and that was the wrong tradeoff. You should have visibility into the safeguards we have in place, and why. We're sorry for not getting the balance right".

Broader Safety Calibration Issues

The reversal on invisible distillation guardrails occurs amid broader concerns about overly restrictive safety measures in Fable 5. Claude Fable has been reported to refuse to answer basic biology questions, suggesting the guardrails may have been overly broad in their implementation. Users reported receiving refusals for standard educational queries about cellular biology and evolutionary concepts that should not have triggered safety concerns.

An Anthropic spokesperson acknowledged this problem, noting that in some cases, notably biology, the safeguards have been calibrated so broadly that Fable is practically unusable for even basic queries. Visible safeguards, unlike invisible ones, require more robust calibration to prevent abuse, which takes additional development time.

Additional Data Retention Concerns

Claude Fable 5 requires data retention to run its new safety classifiers—prompts and outputs are stored for up to 30 days, and up to two years if policy violations are flagged. That's a dealbreaker for some. Microsoft is restricting the model internally because of it, according to The Verge: while all other Claude models run under zero-data-retention rules, Fable 5 doesn't even show up in the internal model picker for Microsoft's GitHub Copilot version.

What This Means for AI Research and Transparency

The invisible throttling controversy highlights a fundamental tension in how AI companies balance competitive advantages against scientific openness and trust. Anthropic's initial approach reflected a common industry instinct: protect proprietary capabilities by making safety measures difficult to probe or circumvent. However, the backlash revealed the cost of this strategy in terms of researcher confidence and perceived fairness.

By choosing visible safeguards, Anthropic is accepting that bad actors can study and potentially probe these measures, but gaining transparency that protects both legitimate researchers and the company's reputation. The company also faces the ongoing challenge of calibrating safety measures precisely enough that they don't block legitimate educational and research use cases while still addressing genuine risks.

The controversy also underscores why policy decisions about frontier AI models warrant public scrutiny. When Anthropic embedded invisible restrictions in a tool widely used by researchers and developers, it did so without explicit disclosure of the implementation method—a choice that violated researcher expectations about how safety features work in other Anthropic models. The reversal signals that companies may face increasing pressure to justify restricted access policies openly rather than through hidden technical measures.