Experts find flaws in tests that check AI safety and effectiveness