The situation reflects poorly on a company that claims to value collaboration with academic institutions.
Anthropic has announced a revision to a policy that inadvertently restricted researchers using its Claude Fable 5 large language model (LLM) from developing competing AI systems. In a statement to a tech publication, the company noted, “We are revising Fable 5’s safeguards for advanced LLM development to enhance transparency.” They acknowledged, “We made an incorrect tradeoff and apologize for misjudging the balance.”
When Claude Fable 5 was unveiled, researchers noticed an unusual behavior; the model would secretly redirect certain requests to a less sophisticated model. This limitation was not mentioned in the accompanying documentation.
Concerns arose as the new model either rejected or weakened responses for tasks such as developing competing LLMs, debugging AI algorithms, and refining neural architectures. Researchers expressed frustration not only due to this performance drop but also because of the lack of clarity from Anthropic, leading to wasted resources on a model that failed to meet expectations.
Given that Anthropic markets itself as a more ethical and supportive option compared to OpenAI, the situation with Fable 5 prompted a rapid negative reaction. Research fellow Dean W. Ball criticized the company’s lack of openness, stating, “Degrading performance on machine learning research without informing users is incredibly unprofessional and reflects poorly on the company.”
Although Anthropic is not eliminating its safeguard policy for Fable 5, it will be making the constraints apparent to users. According to reports, “If the company detects that a user is attempting to utilize Claude to create a highly proficient AI, it will notify them either that the request is being denied or that they are being directed to a less powerful model.”
