On the Effectiveness of Fact Checking Information from Politically Congruent and Incongruent Large Language Models

LLM Fact-Checkers Shift Trust in Political News Regardless of Ideological Alignment

As social media platforms migrate away from human moderators, they are increasingly embedding conversational Large Language Models (LLMs)—the engine behind chatbots like ChatGPT—directly into their interfaces to act as automated fact-checkers. Researchers have long understood that human fact-checkers can reduce the spread of misinformation. However, the transition to AI introduces a new variable: ideological configuration. Unlike a neutral human expert, an LLM can be explicitly programmed with a specific political persona. It can be instructed to prioritize certain sources or tuned to emphasize specific viewpoints. This raises a fundamental question: if a chatbot is configured to be "politically aligned" with a user, does it become a more effective corrector, or does it simply reinforce existing biases?

Can an ideologically biased AI still tell the truth?

The core tension investigated by He, Horne, and Nevo lies in the intersection of machine agency and political psychology. The researchers sought to determine how effective LLM chatbots are at performing two distinct tasks. These tasks involve reducing trust in false political information and increasing trust in true political information. Crucially, they wanted to know if these effects change based on "political congruency"—the degree of ideological overlap between the user and the chatbot.

The study asks whether the robustness of fact-checking survives the shift to AI agents. Historically, fact-checking has worked across partisan lines when performed by humans. The researchers wanted to see if this holds for AI agents that can be intentionally steered toward a "left-leaning" or "right-leaning" persona. If the effectiveness of these tools depends heavily on whether the bot "sounds like" the user, then biased AI could inadvertently deepen social polarization.

The limits of the partisan shield

Before this study, the prevailing understanding in political communication was that fact-checking is remarkably resilient to partisanship. While people are certainly more susceptible to misinformation that confirms their existing views, most research suggests that individuals generally update their beliefs in the correct direction. Backfire effects—where a correction actually strengthens a person's belief in a lie—are documented but remain relatively uncommon.

However, the "cracks" in this framework appear when we move from human-led corrections to automated ones. Traditional fact-checking often relies on simple heuristic cues (mental shortcuts used to make quick judgments), such as a "True" or "False" label. In contrast, LLMs offer a more complex, interactive experience. They can provide explanations, cite sources, and engage in dialogue. This may trigger "systematic processing"—a deeper, more effortful way of evaluating information. The authors note that while humans might treat a simple label as a quick shortcut, the conversational nature of an LLM changes how veracity judgments are formed.

Testing the digital orator

To investigate this, the researchers conducted two within-subjects experiments involving 705 U.S. adults. The researchers built a custom conversational AI system using GPT 5.1. It was integrated with the Exa API to allow for real-time web searching. The chatbots were manipulated via two primary channels. First, they used source-level bias (restricting search results to specific news domains like MSNBC or Fox News). Second, they used language-level bias (using system prompts to instruct the model to adopt a specific ideological persona).

As illustrated in, the experimental flow required participants to first rate their trust in various political headlines.

They then interacted with the chatbots to fact-check those headlines through at least one conversational exchange. In Study 1, participants were randomly assigned to use both a left-leaning and a right-leaning bot. In Study 2, they were allowed to choose which bot they preferred. The researchers successfully manipulated the perceived identity of the bots. shows that most participants correctly identified the ideological distance between themselves and the chatbots.

Effective corrections and the "tainted truth"

The results reveal a dual-edged sword. On one hand, the authors find that LLM fact-checkers are broadly effective at shifting trust in the intended direction. When the bots provided correct verdicts, trust in true headlines increased, and trust in false headlines decreased. This effect was statistically significant and occurred regardless of whether the chatbot was politically congruent or incongruent with the user.

The data in shows the magnitude of these shifts.

In Study 1, correct identification of true headlines increased trust by an average of 0.93 on a 5-point scale. Correct identification of false headlines decreased trust by an average of -0.78. However, the study uncovered a troubling phenomenon the authors describe as the "tainted truth" effect. Even when the LLM was objectively wrong, it still significantly altered user trust. For example, if the bot incorrectly labeled a true headline as "False," trust dropped by an average of -0.96 in Study 1.

Furthermore, the effectiveness of the bot was not entirely uniform. The authors found that perceived political congruency mattered specifically when dealing with "politically distant" headlines. As detailed in and, trust in correctly labeled true headlines increased less when a politically distant chatbot checked a headline that was also distant from the user's views.

Essentially, if the bot and the news both felt "alien" to the user, the correction was less successful.

The implications of automated persuasion

The findings suggest that while LLMs can correct misinformation at scale, they also possess a unique power to "taint the truth" at scale. Because users tend to apply "machine heuristics"—the tendency to view AI as inherently objective or accurate—they may be more susceptible to the errors of a chatbot. This makes the errors of an AI potentially more damaging than the errors of a human.

There are two critical implications here. First, for social media platforms, the study implies that accuracy is a prerequisite for social stability. If a platform deploys a chatbot that is even moderately inaccurate, it will not just fail to correct lies. It will actively erode trust in legitimate news. Second, the research highlights a significant concentration of power. Since only a few corporations have the resources to develop and host these models, the ability to steer the "truth" through ideological configuration rests in very few hands.

The paper concludes by noting that while LLMs can help bridge the gap in truth discernment, they also introduce new vulnerabilities. A vital next step for researchers would be to investigate how specific linguistic styles influence a user's willingness to accept an incorrect verdict.

Figures from the paper