Anthropic has reversed a plan to secretly throttle Claude Fable 5's performance for AI researchers training competing models, acknowledging it made "the wrong tradeoff." The reversal follows sharp criticism that the covert approach was hostile to the broader AI research community. However, a separate issue persists: Fable 5 requires data retention for up to 30 days (or two years if violations are flagged), causing Microsoft to restrict the model internally while other Claude versions operate with zero data retention.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Anthropic had planned to invisibly degrade Claude Fable 5's performance for users trying to train competing AI models, but after backlash from the research community, the company is backing down and apologizing. Going forward, any protective measures will be visible to users.
Why it matters
The covert approach drew criticism from figures including former White House AI advisor Dean Ball, who called it "shockingly hostile." The incident highlights tensions around how AI companies balance competitive interests with openness to the broader research community. However, a separate data-retention requirement for Fable 5 remains controversial—the model stores prompts and outputs for up to 30 days (up to two years if policy violations are flagged), prompting Microsoft to restrict it internally while other Claude models operate under zero-data-retention rules.
What to watch
Microsoft is not including Fable 5 in the internal model picker for its GitHub Copilot version due to the data-retention requirement, signaling that data privacy concerns may limit adoption among large enterprises.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
