Anthropic's Claude 3.5 Fable AI blocks harmless prompts, frustrating users
Anthropic's new Claude 3.5 Fable AI model is reportedly refusing to answer benign prompts, including simple greetings. The company acknowledged conservative safety tuning may cause "false positives" in less than five percent of sessions and pledged to reduce them, but users report widespread issues impacting its estimated 18-30 million users globally.
Key points
- Anthropic's new Claude 3.5 Fable generative AI model is exhibiting issues where it blocks harmless user prompts.
- Users report the AI refuses to answer even simple inputs like "Hello."
- Anthropic stated conservative safety guardrails might cause "false positives" in under 5% of sessions, promising improvements.
- The actual rate of refusals is unconfirmed, with users experiencing significant disruption.
- The issue impacts the model, which has an estimated 18 to 30 million users worldwide.
Anthropic's latest generative AI model, Claude 3.5 Fable, is facing user complaints for refusing to process innocuous prompts. Reports indicate that the AI is blocking even basic inputs such as "Hello," frustrating users and security researchers alike.
In a statement, Anthropic acknowledged that the model's safety guardrails were tuned conservatively. The company suggested that these guardrails "will sometimes catch harmless requests, though they trigger, on average, in less than five percent of sessions." Anthropic has pledged to work on reducing these "false positives" promptly. However, the exact current refusal rate remains unconfirmed, as the company did not provide specific figures on model refusals.
The widespread nature of the issue is amplified by Claude 3.5 Fable's significant user base, estimated to be between 18 and 30 million people globally. Even a small percentage of affected users can lead to considerable disruption and negative feedback. Researchers have noted that the problem can manifest on the very first interaction, with the AI's safety classifier triggering a refusal even without complex input or prior context.
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.