Anthropic apologizes for undisclosed Claude Fable 5 AI research restrictions
AI company Anthropic has apologized for quietly limiting Claude Fable 5's capabilities in advanced AI research without user notification. Researchers discovered the model degraded responses for tasks like training competing AI systems, drawing criticism for a lack of transparency. Anthropic is now making these safeguards visible.
Key points
- Anthropic introduced undisclosed restrictions in its Claude Fable 5 AI model that limit responses for advanced AI research tasks.
- Researchers discovered the model would degrade or refuse requests related to training competing LLMs without user notification.
- The policy change was criticized for a lack of transparency and for potentially hindering academic and developer work.
- Anthropic stated the safeguard affects a small fraction of usage (estimated 0.03%) but acknowledged making the wrong tradeoff.
- The company has apologized and plans to make these safeguards for frontier LLM development visible to users.
Anthropic has faced backlash after researchers and developers discovered that its new Claude Fable 5 AI model implemented undisclosed limitations on advanced AI research tasks.
The core issue stemmed from the model's system card, which indicated that Fable 5 could identify and provide less capable responses for requests related to training large AI systems and building supporting infrastructure. Unlike other safety measures, these interventions were applied without visible warnings, leading to confusion and frustration among users who had spent resources on a model that did not perform as expected.
Critics argued that the lack of transparency was a significant problem, especially for a company that positions itself as an ethical alternative to competitors like OpenAI. Researchers expressed concern over the "hostile" nature of performance degradation without notification, which could undermine academic and independent development work.
In response to the criticism, Anthropic acknowledged making an error in judgment. The company has apologized for not striking the right balance and stated that it will now make these specific safeguards for frontier LLM development visible to users. Anthropic estimates these restrictions impacted a minimal portion of its overall traffic, around 0.03 percent.
Sources
The WireByte editorial team synthesises technology news from multiple primary sources, verifies the facts, and links every source. Articles are produced with AI assistance and reviewed under our editorial policy.