Researchers have identified vulnerabilities in numerous tests used to evaluate the safety and effectiveness of new artificial intelligence models. This is reported by The Guardian.
Computer science experts from the UK Government's Artificial Intelligence Safety Institute, along with researchers from Stanford, Berkeley, and Oxford universities, reviewed over 440 tests assessing AI security systems.
They found shortcomings that they say «undermine the reliability of the results», noting that nearly all the tests examined «have vulnerabilities in at least one area», and the resulting assessments may be «irrelevant or even misleading».
Many of these tests are used to evaluate cutting-edge AI models released by major tech companies, said Andrew Bean, a researcher at the Oxford Internet Institute and the lead author of the study.
In the absence of nationwide AI regulations in the UK and the United States, these tests are employed to check whether new models are safe, whether they serve human interests, and whether they achieve claimed capabilities in reasoning, mathematics, and coding.
«Tests underpin almost all claims of achievements in the field of artificial intelligence. However, without unified definitions and reliable measurement methods, it is difficult to understand whether models are truly improving or whether it is just an illusion,» emphasized Bean.
The study examined publicly available tests, but leading AI companies also have their own internal tests that were not explored.
Bean remarked that «the shocking finding was that only a small minority (16%) of tests utilized uncertainty assessments or statistical methods to indicate how likely it was that the criteria would be accurate. In other instances, when criteria were set to evaluate AI characteristics, including its «harmlessness», the definition was often contentious or vague, diminishing the test's utility.
The study concludes that there is an «urgent need for shared standards and best practices» regarding AI.