Anthropic will now visibly tell users when its most powerful AI model downgrades requests for security reasons, after criticism that it had been silently restricting certain uses.

What happened: Anthropic released Fable 5, its most capable model to date, this week. The company's safety document revealed the model would silently downgrade requests related to advanced AI development—for example, if a researcher used it to build their own AI system. After backlash from AI researchers, Anthropic said on Wednesday it would make these restrictions visible: flagged requests will now visibly fall back to a less capable model (Opus 4.8), and on the API, refusals will include a reason.

Why it matters: Anthropic had buried the restriction in a 319-page safety document without alerting users when it was applied. Critics, including Jeremy Howard of Fast.ai, argued that silently downgrading access would slow AI development. The company's initial approach touched a nerve because it obscured how its own safeguards worked—a transparency issue that comes as AI safety measures are becoming part of national security policy.

What to watch: The company said it will continue downgrading some requests, citing both its terms of service (which prohibit use to build competing AI systems) and national security concerns (preventing foreign adversaries from improving their AI capabilities). Anthropic emphasized the restrictions do not affect the vast majority of coding and ML work. The company also filed confidentially for an IPO earlier this month and is currently in an unresolved federal court battle with the Department of War over its designation as a national security supply-chain risk.