Anthropic has withdrawn its plan to secretly degrade the performance of Claude Fable 5 for users attempting to train competing AI models. The company acknowledged that its initial approach was flawed and expressed regret for not balancing its protective measures correctly. 'We made the wrong tradeoff and we apologize for not getting the balance right,' Anthropic stated to WIRED. The decision follows significant criticism from the research community, which viewed the policy as overly restrictive and potentially harmful to innovation.
The policy would have allowed Anthropic to invisibly reduce the model's performance for users engaged in AI research, raising concerns about transparency and fairness. Dean Ball, a former White House AI advisor, described the covert approach as 'shockingly hostile.' Will Brown from open-source startup Prime Intellect criticized the policy as a way for Anthropic to assert dominance over the AI research landscape, stating, '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.''
In addition to the throttling policy, another controversy surrounds the data retention requirements for Claude Fable 5. The model stores prompts and outputs for up to 30 days, and up to two years if policy violations are flagged. This has become a dealbreaker for some organizations, including Microsoft, which is restricting the model internally due to these data retention rules. Source: thedecoder
