Anthropic has retuned the biology safeguards in its Claude Fable 5 model, sharply reducing how often the AI wrongly declines to answer legitimate questions about biology.
According to Moneycontrol, the update cuts false positives — cases where the model treats a harmless question as dangerous and refuses it — by 85%, while keeping safeguards in place for requests that could genuinely cause harm. A report carried on MSN describes the same change, saying the update reduces biology-related fallbacks by about 85% and lets the model handle more health, education and laboratory-related queries. NDTV Profit similarly reports that Anthropic has updated Claude Fable 5's biology safeguards and expanded access to health and lab questions.
The underlying problem is a familiar one for anyone who has hit a wall with a chatbot. Biology is one of the touchiest domains for AI safety teams, because the same general knowledge that helps a nurse, a student or a lab technician overlaps at the edges with information that could be misused. Models tuned to err on the side of caution end up refusing a great deal of ordinary, useful work — a biology homework question, a protocol lookup, a question about a medication or a disease.
Anthropic's framing, as reflected across the three reports, is that precision improved rather than protection loosened: fewer wrongly blocked questions, same blocks on the requests that matter.
It matters because over-refusal is the quiet tax AI safety puts on real users, and this is a concrete claim that the tax can be cut without dropping the guardrails.