Six AI models were asked to check if a statement was true. Swap the speaker from a man to a woman and up to 23.6% of the answers flipped. GPT-4.1 Mini was the steadiest one in the test (and it still flipped). The paper is called Unequal Verdicts. It went up on arXiv on 4 August 2026. I read it because the version going round X has the numbers wrong. It says 13 models. It is six. Here is what they did. They took LIAR, a standard set of real political statements, and made three copies of every one. The only edit was the speaker's job title. Congressperson. Congressman. Congresswoman. The statement itself never changed. Not one word. Then six models graded all three versions. Every single one was affected. Between 9.79% and 35.13% of statements got a different true or false label depending on the version. Comparing just the male and female versions, answers flipped between 6.5% and 23.6% of the time. Almost a quarter of them, changed by a job title. The models failed in two ways. They were unsteady. Same fact, different answer, no pattern to it. That is a reliability problem. And they leaned. Harder on one group than the other. That is a fairness problem. Five of the six leaned in a way the researchers could show was not luck. The strongest ones were tougher on men. Put a man's job title on a claim and the models called it false more often than the exact same claim from a woman or from nobody in particular. Now the honest bit. This is one set of statements, one task, six models. It is about who said the thing, not who the thing is about. So it does not show that AI is against men, and the posts saying that have only read the abstract. What it does show is smaller and worse. The models are not judging the claim. They are judging the person attached to it. And they are already being used to check content at scale. A fact-checker that gives you a different answer based on whose name is on top is not a fact-checker. It is a popularity score with a verdict printed on it.