Anthropic has walked back a controversial design decision in its latest model, Claude Fable 5, after the AI research community discovered hidden guardrails that silently degraded responses from users suspected of model distillation. The company acknowledged the misstep and is switching to visible safeguards, even if it means more refused queries.

At issue is how Fable handles distillation attempts—where a smaller model is trained on outputs from a larger one. In Fable’s system card, Anthropic disclosed it would alter and degrade responses without notifying users when it detected such activity. Critics argued this covert approach made it impossible for third parties to audit the model’s behavior or verify if legitimate research queries were being throttled.

The new approach routes distillation-related queries to Claude Opus 4.8, Anthropic’s previous flagship, and displays a clear notification each time. This mirrors how Fable already handles other high-risk domains like biology, chemistry, and cybersecurity—where queries fall back to Opus 4.8 rather than being silently modified. However, in some areas, notably biology, the safeguards have been calibrated so broadly that Fable becomes nearly unusable for basic queries, a point Anthropic conceded in a statement to The Verge.

“Visible safeguards can be probed, so they have to be robust, which takes time to get right,” Anthropic wrote. “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.”

The reversal follows sharp pushback from researchers who warned the silent throttling could stifle legitimate evaluation and benchmarking of frontier models. The company had previously justified targeting distillation requests by citing terms of service violations and alleging industrial-scale distillation by Chinese rivals like DeepSeek. For engineers building on top of Fable, the lesson is clear: opaque guardrails erode trust faster than conservative ones.

Source: The Verge