AI Disinformation Study: How to Read the Numbers Behind the Headlines

A report from the fact-checking charity Full Fact and the University of Westminster, posted in October 2026, analyzed 112 pieces of AI-generated or AI-altered false content that circulated between January 2025 and March 2026. Its headline numbers are easy to repeat: 83.9% were substantively false or misleading, 41.1% had potential to contribute to specific harms, and 58.9% had no or very limited potential for consequences. How should a careful reader treat percentages like these?

Start with the sample

The 112 items are not a random sample of all AI-generated falsehoods. They are fact checks that Full Fact chose to publish, and the report itself says this limits how far the findings can be generalized. So 58.9% describes this set, not "AI fakes in general". A statistical interval can still show the uncertainty if we imagine the 112 as a sample from a larger pool of similar checks: 66 of 112 gives a 95% interval of roughly 49.8% to 68.0%, which the confidence interval calculator for a proportion reproduces. Because the selection was not random, the true uncertainty is larger than that interval suggests.

Reach is not harm

Of the 112 items, 22 (19.6%) reached a million views or more, and the report found that more than half of those widely viewed items had no or limited potential for consequences. More than half of 22 means at least 12, or at least 54.5%. A large audience does not guarantee harm, and a small one does not guarantee safety: views measure exposure, not effect. To test whether the two classifications are associated, you would need the full cross-table of both. With those counts, the chi-square test of independence would show whether reach and harm potential go together.

Percentages of what?

The report notes that confirmed or suspected AI claims made up 24% of Full Fact's published checks in the study period, against 8% a year earlier. That is a threefold rise and a 16-point increase in share, but it is a share of one organization's output, not of all misinformation, and it depends on what that organization chooses to check and how well it can detect AI. In only 33 of the 112 checks (29.5%) was the role of AI established unequivocally; in many others the language was weaker. Always ask what the denominator is. The scientific calculator makes such ratios quick to check.

A percentage without its denominator and its sampling rule is a number without a meaning.

Judgments are not measurements

The harm categories come from a model that asks whether a claim is substantively false, whether enough people believe it, and whether they have the capacity and motivation to act. These are judgments, and an independent evaluation cited in the report noted that such judgments can be inconsistent, which is why the authors added a review step. When people classify items, the reliability of the classification (how often two reviewers agree) matters as much as the totals. Percentages built on subjective categories deserve the same caution as percentages built on measured ones.

Questions for any study of this kind

How was the sample chosen? What is each percentage a share of? Is the outcome measured or judged? Do the numbers describe exposure or effect? Do the authors state their own limits? This report is open about several of them, which is a good sign. The same checklist works for studies of scams, health claims, elections and more. For planning your own study, see our guide on choosing a sample size.