People Trust AI Fact-Checking, Until They See How It Works

Fake news travels fast. By the time a human fact-checker verifies a viral claim, millions may have already seen, believed, and shared the original lie. Social media platforms have been racing to deploy AI systems that can flag false content automatically, at lightning speed and enormous scale. However, a new study finds people’s trust in AI fact-checkers is more complicated than platforms might hope. The reason reveals something genuinely surprising about human psychology.

Research published in the journal Media Psychology found that when people encounter an AI-powered fact-checking system, two competing gut reactions fire almost simultaneously. On one hand, they assume the machine is objective, precise, and free from political bias, a natural advantage over any human who might have an agenda. On the other hand, they worry that a machine simply can’t understand the messy, context-dependent nature of human language. Those instincts effectively cancel each other out, leaving overall trust in AI fact-checkers roughly equal to trust in human fact-checkers, but for very different underlying reasons.

What really mattered, researchers found, was not whether a human or a machine did the fact-checking, but how the system explained its reasoning. And the results turned up a genuine paradox. The explanation style that users found easiest to understand was also the one that, under the right conditions, made them trust AI less.

How AI Fact-Checkers Were Put to the Test

Researchers built a realistic simulated fact-checking platform called FactDeck and ran an online experiment with 291 adults across the United States, recruited in July 2022. Participants ranged in age from 18 to 79, with an average age of just over 44. A little more than half identified as female, and most held at least a bachelor’s degree.

Each participant was randomly placed into one of six versions of FactDeck. Half were told fact-checking was done by a human expert; the other half were told it was done by an automated AI system. Separately, the platform explained its reasoning in one of three ways: an evidence-based approach, similar to Snopes or PolitiFact, comparing the claim against outside evidence; a feature-based approach, flagging content based on writing style, word choice, or source; or a black-box version that flagged content as false without any explanation. Afterward, participants filled out a survey measuring trust, credibility, perceived bias, and future reliance.

The Two-Sided Machine Problem in AI Fact-Checking

One of the central findings involves what the researchers call the “positive machine heuristic” and the “negative machine heuristic,” two opposing mental shortcuts people apply almost automatically when they learn a machine made a decision.

Machines are seen as precise, consistent, and ideologically neutral, a natural fit for fact-checking, where accusations of partisan bias run rampant. Exposure to an AI fact-checker did trigger this positive association, pushing trust upward.

But the negative version fired at the same time. Machines are rigid and emotionless. They can’t pick up on sarcasm, cultural context, or the subtle ways humans bend language to mislead. For a job that requires understanding human intent, that’s a real limitation, and this reaction pushed trust back down.

Because both reactions occurred at once in opposite directions, the net effect on overall trust was essentially zero. Neither human nor AI fact-checkers came out significantly ahead. This helps explain why earlier research produced such contradictory results: studies vary in which heuristic they trigger most strongly, and in reality, both are always at work.

When Transparency Helps, and When It Backfires in AI Fact-Checking

One clear finding involved the black-box systems, which flagged content as false but gave no explanation whatsoever. Participants consistently trusted these systems less, found their conclusions less credible, and perceived them as more biased compared to systems that explained themselves at all. Simply providing any reasoning improved how a system was perceived.

Not all explanations worked equally, though. When the system described comparing a claim against outside sources, participants found that approach significantly easier to follow than when the system described analyzing patterns in the text itself. Familiarity likely played a role: most people have seen evidence-based debunking before and expect supporting evidence in news contexts.

Even so, evidence-based explanations carried a specific cost when paired with AI. When the system described reviewing and interpreting outside evidence and was labeled as AI, participants were more likely to question whether a machine could really handle that task. Human fact-checkers faced no such credibility penalty for using the same approach.

What This Means for the Fight Against Misinformation

Taken together, the results paint a complicated picture for anyone deploying AI tools against misinformation online. AI can process enormous volumes of content far faster than any team of human fact-checkers. But public trust doesn’t flow automatically toward AI just because it’s faster or more capable at scale.

How a system presents its reasoning may matter as much as who is doing the reasoning. Showing evidence, the approach most likely to help users understand a flagged claim, is also the approach most likely to trigger doubts about whether an AI can truly weigh human language and context. That’s a genuine design challenge, and so far, no one has a clear answer.

---

Disclaimer: This article summarizes peer-reviewed academic research and is intended for informational purposes. Study findings reflect a specific sample and experimental conditions and may not apply universally.

---