AI safety tests are heavily flawed, new study finds — here’s why that could be a huge problem

Benchmarks test AI models on a range of factors, ranging from their logical skills to how well they deal with emotion, but ...