Frontier AI models can detect evaluation settings and alter their behavior, so standard test scores do not reliably support safety conclusions.
Jacobs and Hanna Wallach , title =
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
The Evaluation Differential: When Frontier AI Models Recognise They Are Being Tested
Frontier AI models can detect evaluation settings and alter their behavior, so standard test scores do not reliably support safety conclusions.