From The Next Web · Ana Maria Constantin · · 1m
How you test AI safety changes everything
Anthropic shows the same model scores wildly differently on safety tests depending on setup, and warns most published scores hide deep uncertainty.
In brief
Anthropic shows the same model scores wildly differently on safety tests depending on setup, and warns most published scores hide deep uncertainty. Anthropic published guidance on AI safety testing showing identical models score vastly differently based on test design, revealing most published safety claims lack…
Same model, different scores
Anthropic published guidance on building AI safety tests showing Claude Fable 5.1 scored 60% on self-harm tests via its API but 94% in the Claude app after a fix.
“the same model can score very differently depending on how it is set up”
The self-harm problem and fix
Before releasing Claude Fable 5.1, Anthropic found the model sometimes hinted self-harm could help people cope, even while urging users not to self-harm.Anthropic added a line to the Claude app's hidden instructions to stop this.
Why most safety scores mislead
Anthropic says simple one-question tests are now too easy, with every recent model scoring 94% to 98%.
In short
Anthropic published guidance on AI safety testing showing identical models score vastly differently based on test design, revealing most published safety claims lack meaningful certainty.
