AI testing alone is not AI safety

Writing in Nature Computational Science on 28 August 2026, George Breckenridge of the University of Edinburgh and Feng Li of City St George's, University of London, argue that evaluation scores are being treated as a substitute for real-world deployment governance.

Their contribution is the case that institutional controls are needed beyond technical tests. For enterprise assurance, evaluations must be combined with access controls, monitoring, incident response and accountability.

The caveat: this is a peer-reviewed policy argument, not an empirical benchmark study.