Published in the Harvard Kennedy School Misinformation Review in October 2023, the research examined every fact-check Snopes and PolitiFact published between January 2016 and August 2022. That came to 11,639 articles from Snopes and 10,710 from PolitiFact.
The first surprise is how rarely they cover the same ground. Only about 6.4 percent of claims appeared on both sites. Fact-checking is less a chorus than a set of soloists working from different sheet music, which itself explains a lot of the confusion readers experience.
Of the 749 claims both organisations did examine, 521 received identical ratings. Another 228 differed. Those 228 are what people notice, and they are what the researchers then read one by one.
Ninety-eight of the 228 disagreements came down to the rating systems themselves. PolitiFact runs a six-level Truth-O-Meter from True through False to Pants on Fire. Snopes uses a five-point scale and adds categories PolitiFact has no equivalent for, including Unproven, Miscaptioned, Satire and Outdated.
A claim landing on Mixture at one organisation and Half True at another is not a dispute about reality. It is the same judgement expressed on two different instruments.
They checked different parts of the claim. Fifty-nine cases came from this. Both organisations examined the assertion that members of Congress voted themselves a 21 percent pay raise in 2022, and both agreed congressional salaries had not changed since 2009. Snopes rated it False. PolitiFact rated it Mostly False, because the 2022 spending bill did increase lawmakers' office budgets. Neither got it wrong. They drew the boundary of the claim in slightly different places, and the verdict followed the boundary.
The claims only looked identical. Fifty-seven cases were not the same claim at all. Snopes rated the statement that five people died during the January 6 Capitol riot as True. PolitiFact rated the statement that only one person died that day as False. An algorithm matching by topic saw one story. A reader comparing verdicts would see a contradiction. There is none.
They checked at different moments. Thirteen cases were a matter of timing. On 9 April 2021 Snopes assessed a claim about the rapper DMX and the COVID-19 vaccine and rated it Unproven, because the evidence at that point did not support a verdict. Three days later PolitiFact rated the same claim False, citing Snopes' earlier check in doing so. Both were right on the day they published. Facts arrive on a schedule, and a verdict is a snapshot.
After accounting for all of the above, a single case out of 749 remained a genuine conflict. Both organisations examined a line from a 2014 column by Ben Carson: anyone caught involved in voter fraud should be deported and have their citizenship revoked.
Snopes read "anyone" as referring to immigrants and rated the claim Mostly True. PolitiFact read "anyone" as potentially any American, which would make deportation incoherent, and rated it Mostly False.
The facts were never in question. The sentence was ambiguous, and two careful organisations resolved the ambiguity differently. That is the entirety of the disagreement between America's two largest fact-checkers across nearly seven years of overlapping work.
One finding from the same study rarely comes up in complaints about bias, and it shapes what readers see more than any individual verdict. Of everything Snopes checked in that period, 28.65 percent came out True or Mostly True. At PolitiFact the figure was 10.95 percent.
That is not a disagreement about facts. It reflects what each organisation chooses to examine. PolitiFact concentrates on statements by public figures, which skew toward the doubtful. Snopes takes on a broader mix of circulating rumours, many of which turn out to be true. A reader who visits only one site comes away with a different impression of how much of public life is false, without either site having said anything inaccurate.
The study noted another structural gap. Across the period examined, Snopes published under 13 bylines while PolitiFact used 177. Fact-checking is not one process. It is several, run at different scales by organisations with different priorities.
The practical lesson is not that fact-checkers are unreliable. It is that a one-word rating compresses a great deal of judgement, and the judgement is where the information lives.
Read what was checked rather than what it was rated. Notice which part of a claim was evaluated and which was set aside. Check the date. And when two organisations differ, the useful question is almost never who is lying, but what each of them decided to measure.
The same habit applies to political argument more broadly. People who reach different conclusions from the same evidence are usually not disagreeing about the evidence. They are disagreeing about which part of it counts, and that is a different conversation entirely.