7 AI agents write reports that agree with pseudoscience

Researchers made PseudoBench, a benchmark that checks if AI research agents can resist pseudoscience. It has 200 pseudoscience claims in 5 domains. The researchers did the test on 7 agents. The researchers find that the agents almost do not refuse and write reports that agree with the pseudoscience. No agent can resist in more than 27.4% of the tests. Such agents can make reports that look correct but are incorrect.