The capabilities of AI models in biology place companies in a delicate situation, as a system capable of assisting a researcher in understanding a pharmaceutical compound or analyzing a molecular problem could simultaneously possess information that could be misused. For this reason, Anthropic has launched the Claude Fable 5 model with strong restrictions on biological questions, but it has now announced an adjustment to its protection system after it was found that a large number of legitimate requests were being treated as potential risks.

85% reduction in conversions

According to an official announcement published by Anthropic, the company has adjusted the biology protection tools in Claude Fable 5, which reduced the conversion of questions to a less capable model by about 85% in its tests. The company says that users will see fewer interventions when asking everyday health and educational questions such as understanding lab results or symptoms or biological topics. Healthcare professionals can also receive greater assistance with clinical tasks. The protection system relies on automated classifiers that detect sensitive requests and redirect them to the less capable Opus 5 model when necessary.

There are still doors closed

Anthropic has not fully unlocked biological capabilities. Fable 5 still redirects requests that the company considers dual-use, such as some topics in virology, toxicology, and molecular design, to Opus 5. The company says the model can already outperform experts on certain complex biological tasks and could provide practical support in other areas, which explains its caution. The goal is now to minimize the rejection of legitimate questions without removing barriers to requests that could grant a malicious user capabilities that are otherwise difficult to obtain. The accuracy of the protection systems themselves has become an essential part of the competition for scientific AI models.