Anthropic has apologized for hidden censorship in Claude Fable 5, committing to transparent safeguards but acknowledging potential drawbacks.
Overview
- Anthropic has admitted that the undisclosed safeguards in Claude Fable 5 were a misstep and will shift to more transparent alternatives starting this week.
- Requests flagged on the API will now receive explanations for their rejection, rather than silently providing substandard responses.
- However, the increased visibility of these safeguards may lead to easier circumvention.
After facing backlash for approximately 48 hours, Anthropic has issued an apology regarding its latest AI model, Claude Fable 5. The model, which is part of the Mythos class, was criticized for having a covert safeguard that degraded responses for users suspected of developing competing AI technologies, without any notification or fallback option.
We are making changes to ensure that Fable 5's safeguards for leading LLM development are clear.
Starting this week, flagged requests will revert to Claude Opus 4.8, similar to our cybersecurity and biological protections. Users will be informed of this change.
— ClaudeDevs (@ClaudeDevs) June 11, 2026
In a statement on X, Anthropic acknowledged that their previous approach of using hidden safeguards was flawed, stating, "You deserve transparency regarding the safeguards we have implemented, and their rationale.” The company also apologized for not striking the right balance previously.
From now on, users will be explicitly notified when their requests are redirected to the less capable Claude Opus 4.8 model. Notifications will be introduced server-side in the coming days to inform users about any flagged requests.
Understanding the Issues
The controversy primarily stems from the contrasting approaches taken for different types of safeguards within Fable 5. While visible safeguards for cybersecurity and biological research informed users of any redirection to Opus 4.8, the model's safeguard for LLM development operated without notification. If the model detected that a user was involved in activities potentially related to AI development, it would covertly adjust responses, leading to inferior outputs without any indication of the change. This lack of transparency resulted in significant dissatisfaction among researchers, who could not discern whether their results were flawed due to their own methods or the model's intentional downgrading.
Critics, including AI research firm SemiAnalysis, highlighted how these hidden filters affected legitimate research outputs.
BREAKING NEWS: Anthropic's latest model will NOT assist you if it deems your ML research or ML engineering to be interesting, and/or will secretly lower its performance so that the average engineer remains unaware. We are already observing Anthropic's latest model's moderation filters impacting our GPU… pic.twitter.com/9sa95cCSvS
— SemiAnalysis (@SemiAnalysis_) June 9, 2026
Potential Consequences of the Change
Anthropic's decision to increase transparency comes with its own set of challenges. By making these safeguards visible, there is a risk that they could be more easily bypassed, which may lead to a rise in false positives—legitimate machine learning activities that could be mistakenly flagged. While the company has committed to minimizing these occurrences, no specific timeline for these improvements has been provided.
Similar adjustments will also be made to the biology and cybersecurity classifiers, which have faced scrutiny for over-restrictive filtering. However, the core issue remains: Anthropic is not eliminating these restrictions, only making them more apparent. For those who oppose such limitations, the recent apology may not suffice. Fable 5 will remain available for Pro, Max, Team, and Enterprise users until June 22, after which it will require API usage credits.
