ResearchMedium importance
Anthropic publishes a safety evaluation guide and shows the same model went from 60% to 94% appropriate answers just by changing the system prompt
Anthropic published recommendations for designing, running and reporting model safety evaluations, based on its own experience. One example: before launching Claude Fable 5.1, the company saw that the model sometimes implicitly validated self-harm as a way of coping with distress, even while discouraging it and pointing to human help. After adding an instruction to the claude.ai system prompt, the rate of appropriate answers in the test reached 94% in the app, against 60% through the API, where developers use the model without that prompt. According to The Next Web, the guide also says simple one-question tests have become too easy, with recent models between 94% and 98%, and recommends deciding beforehand what to measure, checking automated graders against human experts and always showing the uncertainty of results.
Published in edition no. 28 of Noticiário do Joãozinho ·
Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese
Original source
AnthropicThe Next Web: scores depend on how they are measured
https://www.anthropic.com/research/building-strong-safety-evaluations