Does naming Grok as fact-checker change how people update beliefs?

Renault, Mosleh and Rand's preregistered survey experiment (N = 1,592 U.S. participants, recruited roughly evenly by party, December 2025) finds that "although participants across all subgroups meaningfully updated their beliefs in the direction of the LLM fact-checks," disclosing which bot produced the fact-check changed the size of that update by partisanship: "revealing that a fact-check was generated by Grok reduced updating among Democrats while increasing updating among Republicans," whereas "responses to Perplexity are not" polarized by disclosure. A parallel asymmetry appears in the field data on who gets checked at all: Republican-authored posts were targeted for fact-checking "97.3% more" than Democrat-authored posts across all requests, and more likely to be "rated as inaccurate by both LLMs."

The authors frame this as polarization of the AI brand itself, not just its verdicts: "users requesting fact-checks from Grok are much more likely to be Republican than Democratic, while the opposite is true for fact-check requests from Perplexity," so the choice of which bot to ask already sorts by party before any verdict is issued. Disclosing "Grok" then functions as a partisan cue that moves belief independent of the verdict's content. The authors describe this as a preview of a possible future in which LLMs become as polarized as legacy media, including "companies designing LLMs to appeal to right- or left-leaning audiences by reinforcing each group's preconceptions."

This sits alongside Where does AI's persuasive power actually come from?, which locates persuasive power in training and prompting choices tested across political issues: this paper shows that even holding a fact-check's content fixed, disclosing which AI brand issued it moves belief in opposite directions for the two parties. It also complicates Does AI fact-checking actually help people spot misinformation?'s account of users treating AI verdicts as a uniform replacement signal for their own judgment — here the replacement signal is read through a partisan lens attached to the model's identity, not received identically by every user.

The disclosure-polarization result comes from one preregistered survey run in one month, with one pair of AI brands already carrying particular ownership associations and reputations in U.S. politics; the excerpt does not show whether the effect would hold for a less politically branded model, or whether it reflects prior attitudes toward Grok's owner rather than anything about the fact-checking mechanism itself. The field-level targeting asymmetry is observational, and the authors note it "may partly reflect network structures" rather than deliberate partisan targeting. The implication the evidence does support is that once an AI fact-checker is publicly branded, disclosing that brand is not a neutral transparency move — it changes how much the same verdict shifts belief, depending on the reader's party.