Experts find flaws in hundreds of tests that check AI safety and effectiveness

Scientists say almost all have weaknesses in at least one area that can ‘undermine validity of resulting claims’