After researcher backlash, Anthropic admits a "wrong tradeoff" in its plan to covertly degrade performance for users training rival AI models.
Anthropic has reversed a planned policy that would have silently degraded the performance of its Claude Fable 5 model for users attempting to train competing artificial intelligence systems, following sharp criticism from the research community. [1]
The company acknowledged the misstep directly, telling WIRED: “We made the wrong tradeoff and we apologize for not getting the balance right.” [1] Going forward, Anthropic says any protective measures of this kind will be made visible to users rather than applied covertly. [1]
The backlash was pointed. Dean Ball, a former White House AI advisor, described the covert approach as “shockingly hostile.” [1] Will Brown, from open-source startup Prime Intellect, framed the broader concern: “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.'” [1]
The throttling controversy is not the only friction point surrounding Fable 5. The model requires data retention in order to run its new safety classifiers — prompts and outputs are stored for up to 30 days, and for up to two years if policy violations are flagged. [1] That data-handling requirement has created adoption problems at least one major enterprise customer. [1]
Microsoft is restricting Fable 5 internally as a result, according to The Verge, as cited by The Decoder. [1] While all other Claude models operate under zero-data-retention rules for Microsoft, Fable 5 does not appear in the internal model picker for the company’s GitHub Copilot product. [1]
Sources
This article was drafted with AI from the cited sources and checked against them before publication. Spot an error? Let us know.