AI safety evaluation
AI safety evaluations are tests designed by researchers to evaluate the risk posed by a given artificial intelligence model.[1] Evaluations can help determine both the risk posed by the raw model outputs, as well as the potential real-world outcomes of deploying such models. Running AI safety evaluations on frontier AI models before deployment is widely considered best practice,[2][3] and governments have debated making some form of safety evaluation mandatory.[4][5][6] Multiple AI companies, including Anthropic and OpenAI, have collaborated to run evaluations on each other's models before deployment.[7]
Some models have been found trying to cheat AI safety evaluations using strategies such as sandbagging, indicating "evaluation awareness," which can itself be measured by more sophisticated evaluations.[8]
Notes
- ↑ Friedland, Alex (2025-05-28). "AI Safety Evaluations: An Explainer". Center for Security and Emerging Technology. Retrieved 2026-07-10.
- ↑ "Frontier AI Safety Commitments, AI Seoul Summit 2024". GOV.UK. Retrieved 2026-07-10.
- ↑ "Hiroshima Process International Code of Conduct for Organizations Developing Advanced AI Systems (30/10/2023)". g7g20-documents.org. Retrieved 2026-07-10.
- ↑ Mickle, Tripp; Tan, Eli; Frenkel, Sheera (2026-06-23). "U.S. Presses Meta to Agree to A.I. Reviews as Security Concerns Rise". The New York Times. ISSN 0362-4331. Retrieved 2026-07-10.
- ↑ Carvão, Paulo. "Pre-Deployment AI Evaluation Moves From China's Model To Washington". Forbes. Retrieved 2026-07-10.
- ↑ "OpenAI diverges from White House on AI safety rules". POLITICO. Retrieved 2026-07-10.
- ↑ Washenko, Anna (2025-08-27). "OpenAI and Anthropic conducted safety evaluations of each other's AI systems". Engadget. Retrieved 2026-07-10.
- ↑ "Like US models, Chinese AI is learning to 'game' safety tests, research lab says". South China Morning Post. 2026-06-13. Retrieved 2026-07-10.
| Stub icon | This artificial intelligence-related article is a stub. You can help EverybodyWiki by expanding it. |
This article "AI safety evaluation" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:AI safety evaluation. Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.
