Social media platforms have been quietly replacing human fact-checkers with large language models. A new paper, "On the Effectiveness of Fact Checking Information from Politically Congruent and Incongruent Large Language Models" by He, Horne, and Nevo, finds that LLM fact-checkers are significantly less effective when the claim being checked is politically incongruent with the model's training distribution. In plain language: the AI fact-checker has a bias that mirrors the bias it was trained to correct. This lands in the same week that Harvard historian Jill Lepore published a book arguing that AI backlash is not technophobia but a historically necessary corrective to unchecked technological deployment.
The Political Congruence Problem in AI Moderation
The research is specific and alarming. When platforms deploy LLMs as fact-checkers, they are deploying systems whose accuracy degrades precisely in the cases that matter most: contested political claims where the human fact-checker's bias was already the problem. The model inherits the bias of its training data, which inherits the bias of the humans who generated it. This is not a bug in a specific model. It is a structural property of the approach. A 2024 paper in PNAS by Bail et al. found that automated moderation systems systematically underperform on politically charged content compared to neutral factual content, and that this gap widens as political polarization increases. The platforms are deploying the wrong tool for their hardest problem and calling it progress.
Lepore's History and the Audit Question
Lepore's argument in The Rise and Fall of the Artificial State is that AI is not uniquely dangerous but that it is being deployed with an unusual absence of institutional accountability, the kind of accountability that, historically, only emerges after a crisis forces it. The He et al. paper is a pre-crisis signal. The Hugging Face breach is a parallel one: the platform through which AI models are shared and deployed was itself compromised. Who fact-checks the fact-checker? Who audits the model repository? Brewster Kahle's thinking on public AI and the global brain is the useful counterargument: the answer to captured AI infrastructure is not better corporate AI infrastructure but public, auditable, open systems. The congruence trap is not a model problem. It is a governance problem.