Morgan Wack, a postdoctoral scholar in the Department of Communication & Media Research at the University of Zurich, joined Verasight to talk about "Fact-check? True!: Tracking Verification Requests and the Absence of Truth Verdicts across Correction Systems," a new working paper with Patrick Warren of Clemson University. The paper includes a question in Verasight's APSA omnibus survey, which was assigned to 1,000 U.S. adults and fielded from September 25-30, 2026. In this article, we highlight the key points from the interview and discuss next steps for this area of research.
People question true and false information almost equally
"I think the interesting thing, and I think matching somewhat of my co-author and I's priors, are that people in general really question both true and false information almost exactly equally. I think we were really surprised by how close this actually was to kind of 50/50 split. I think we were expecting maybe split in one direction or another. But what we found was actually it was about 51% of the posts that people asked Grok to fact-check ended up being true, and we did this with both LLM-based fact-checks and kind of human validation. But across basically every topic we looked at, every sort of research focus we looked into, it was about a 50/50 split of true and false information."
The study treats requests to Grok, the AI assistant on X, as a record of what people doubt. The authors collected a 10% random sample of these queries from January and February 2026; the final sample contains 36,675 queries on 29,294 posts. Among the 19,042 queried posts on which Grok gave a true or false verdict, Grok judged 51.4% true. Across the ten named topics in the paper, the share of posts judged accurate ranged from 49.4% to 58.3%.
Fact-checking organizations rarely rate claims as true
"We see only about an 8% rate of... true fact-checks in the data that we look at."
Table 2 of the paper compares the share of each verification system's verdicts on true claims. Grok's share is 51.4%. Of the Community Notes written on the queried posts, 27.3% verify the post rather than flag it, and none of those 2,981 notes was displayed, since the scoring code excludes notes with this classification from display. Among professional fact-checkers, drawn from the FACTors dataset of 39 organizations, the share is 8.0%, or 11% when verdicts of Mostly True are included. The paper notes that experimental work finds debunking effective at correcting both forms of misperception.
Two forms of misperception: incorrect acceptance and incorrect rejection
"And so what we wanted to do was basically look at two different types of misperceptions. Essentially, if you imagine misperception as some sort of incorrect belief, we see it as occurring in kind of two different directions. You can either incorrectly believe something or you can incorrectly reject something, and so we call this either incorrect acceptance or incorrect rejection. And so a lot of the kind of focus of the fact-checking literature and the correction literature more broadly is about incorrect acceptance. So it's about people believing things that are untrue. This is what people mostly refer to when they talk about kind of both mis- and disinformation. But there also is a substantial portion of misperceptions that fall into this other bucket of incorrect rejection, where things are actually true and people doubt them and don't believe this to be the case."
Figure 1 of the paper, "The four types of accuracy perception," crosses the underlying truth of a post with what a user believes. Incorrect acceptance is believing a post that is inaccurate; incorrect rejection is doubting a post that is accurate. Notes that flag a post as misleading address incorrect acceptance, and notes that verify a post address incorrect rejection. The authors write that the evidence on corrections, and the infrastructure it supports, "often focuses on the correction of false information."
Next steps
The paper names a practical next step for platforms that use crowdsourced systems like Community Notes: to "evaluate the public display of notes verifying posts." Asked what comes next for the research, Morgan Wack described current work on large language models and persuasion:
"We're doing a lot of work now on kind of LLMs and how they might shift people's opinions on political topics, but also everyday topics, people interacting with this new type of tool in new ways. So that's sort of what we're focused on at the moment, and yeah, if people are interested, feel free to reach out."
About the data
The study draws on a 10% random sample of queries addressed to Grok on X between January 1 and February 28, 2026 (47,899 queries in the export; a final sample of 36,675 queries on 29,294 posts), publicly released Community Notes data joined to those posts, and the FACTors dataset of professional fact-checks from 39 organizations. It also includes one question in Verasight's APSA omnibus survey, which was assigned to 1,000 U.S. adults and fielded from September 25-30, 2026; the authors use Verasight's supplied population weights. The question read: "On average, for all the fact-check questions Grok is asked to judge, for what fraction do you think Grok responds that the claim was 'true'?" The finding that the median estimate fell in the band from 45% to 54% comes from the Verasight survey. The authors note that the findings rest on queries directed to one bot on one platform over a two-month period.
