Anthropic has reversed a policy intended to silently degrade the performance of its Claude Fable 5 model for researchers attempting to train competing artificial intelligence systems. Following significant backlash from the open-source community, the company admitted it made the wrong tradeoff and apologised for failing to get the balance right. Consequently, any future protective measures designed to safeguard proprietary data will now be visible to users rather than applied invisibly. Critics described the covert approach as shockingly hostile, with one expert noting it suggested Anthropic did not trust other parties to conduct AI research responsibly.
The controversy extends beyond throttling to significant data retention concerns. Fable 5 requires prompts and outputs to be stored for up to 30 days, or up to two years if policy violations are flagged, to run new safety classifiers. This stands in direct contrast to the zero-data-retention rules applied to other Claude models within Microsoft’s ecosystem. Microsoft has restricted the model internally, removing it from the GitHub Copilot model picker entirely due to these privacy implications. The situation highlights a growing tension between proprietary safety mechanisms and the transparency required for collaborative research.
- Anthropic has apologised for invisibly throttling rival AI researchers and will make future restrictions visible.
- Microsoft has excluded Claude Fable 5 from its GitHub Copilot version due to mandatory data retention policies.
- Privacy concerns remain a major dealbreaker, contrasting with the zero-retention rules of other Claude models.
