Flaws in AI Safety Testing Revealed

Experts have identified critical weaknesses in hundreds of AI safety tests used to evaluate the security and efficacy of new artificial intelligence models. According to The Guardian, computer science specialists from the UK Government's AI Security Institute, along with researchers from Stanford, Berkeley, and Oxford, examined over 440 tests assessing AI safety systems.
They found shortcomings that, as they stated, «undermine the reliability of the obtained results», noting that nearly all reviewed tests «have vulnerabilities in at least one area», making the results «potentially irrelevant or even misleading».
Many of these tests are used to assess cutting-edge AI models released by major tech companies, highlighted researcher Andrew Bean from the Oxford Internet Institute.
In the absence of national AI regulations in the UK and the US, these tests are employed to determine whether new models are safe, align with human interests, and achieve claimed capabilities in reasoning, mathematics, and coding.
«Tests form the basis of almost all claims about advancements in artificial intelligence. However, without unified definitions and reliable measurement methods, it is challenging to understand if models are genuinely improving or if it's merely an illusion», Bean emphasized.
The study focused on publicly available tests, while leading companies in AI also possess proprietary internal tests that were not examined.
Bean noted that «a shocking finding was that only a small minority (16%) of tests used uncertainty ratings or statistical methods to show how likely the criteria would be accurate. In other cases, when criteria were established to evaluate AI characteristics, including its ‘harmlessness’, the definitions were often vague or contentious, reducing the utility of the test.
The research concludes that there is an «urgent need for common standards and best practices» regarding AI.