A new study gave five frontier AI models 1,000 real-world claims to fact-check. They disagreed on 67% of them.