Anthropic's Fable 5 adjusts guardrails after user backlash

Anthropic has faced considerable backlash over its Fable 5 model, which features a controversial invisible guardrail designed to prevent users from utilizing the model for training other AI systems. This measure, intended to safeguard against potential misuse, instead resulted in user outrage, with many AI researchers expressing their frustration publicly. The invisible nature of this guardrail was perceived as a breach of trust, prompting Anthropic to reconsider its approach.

In a statement to Wired, Anthropic confirmed that it would modify Fable 5’s safeguards to make them visible to users. The company acknowledged that the decision to keep these guardrails hidden was a miscalculation. Previously, the model employed techniques such as prompt modification and parameter-efficient fine-tuning (PEFT) to obscure limitations, which users found unacceptable. By making these safeguards transparent, Anthropic aims to restore trust and provide users with a clearer understanding of the model’s capabilities and restrictions, particularly in frontier LLM development.