Most wrong numbers are not invented. They are real numbers with a question left unasked. Which year? Counted by whom? Rounded which way? This site corrects its own figures in place when they turn out wrong, and keeps a note of how each mistake happened. Case after case comes down to one of six questions nobody asked in time. Here they are, each with the mistake on this site that taught it.
- Which year does it describe? Not the year it was published.
- Who counted? A company’s own figure and a tracker’s estimate are different claims.
- Was it rounded, and which way? Find the unrounded figure.
- All of something, or some of it? A sample is not a census.
- A finding, or an estimate? “May” and “could” are part of the number.
- Is it still the latest? Official figures get replaced, and so do corrections.
1 · Which year does it describe?
Many annual reports are named for the year they come out and describe the year before. Quote the report’s year as the data’s year and every number shifts one step out of date.
This site did exactly that on a chart of automated web traffic. The note kept on the fix: “The bars had been labelled with the year of the Imperva report rather than the year it measured, leaving the chart one edition behind the prose beside it.”Site data: this site’s own correction of 16 September 2026, to the chart on how much of the internet is bots.
Dates of events slip the same way. A page here said Meta announced its purchase of Moltbook on 11 March 2026. The note: “The announcement was 10 March; 11 March was the date of the Fortune article the page cited.”Site data: this site’s own correction of 12 September 2026, noted in place on Moltbook.
2 · Who counted, and how would they know?
A company counting its own users, a researcher estimating them from outside, and a website repeating either one are three different kinds of claim, even when they print the same kind of number.
A page here gave companion-app user figures with no source. When it was checked, one figure held up in the company’s own words — Character.AI describes “over 20 million monthly active users across our services” — and the rest did not. The note: “Age split and the 30 million figure came from third-party trackers; removed.”Site data: this site’s own correction of 16 September 2026, noted in place on AI companions, quoting Character.AI’s engineering blog as read at source that day.
Ask how the counter could possibly know. A company can count its own logins. An outsider can only estimate them.
3 · Was it rounded, and which way?
A ratio or a fraction is often somebody’s rounding of two numbers you have not been shown. Find the two numbers.
A page here said automated traffic was “growing eight times faster than human traffic”. The source actually gave two growth rates: automated traffic “grew 23.51% year over year, while human traffic increased 3.10%”. The note: “The ratio is 7.6x.”Site data: this site’s own correction of 12 September 2026, noted in place on AI agents, quoting HUMAN Security’s 2026 report release.
The critical-thinking page on this site once said “a third” of countries limit media-literacy teaching to technical digital skills. UNESCO’s figure is 29%.Site data: this site’s own correction of 12 September 2026, noted in place on the critical-thinking page. Both rounded up, and both rounded towards the more striking version.
4 · All of something, or some of it?
A study of eight things tells you about eight things. The step from “every one we tested” to “every one” is small in a sentence and large in meaning.
A page here said “every major agent benchmark” could be gamed. The research said “Every single one of the eight audited benchmarks can be exploited to achieve near-perfect scores without solving a single task.” The note: “Eight audited benchmarks is not every major one.”Site data: this site’s own correction of 11 September 2026, noted in place on who builds AI, quoting UC Berkeley’s benchmark audit.
5 · A finding, or an estimate somebody promoted?
Research is full of careful words — may, could, susceptible, at risk. Retelling can drop them. Once dropped, a possibility reads as a result.
A well-known case is a 2013 Oxford paper estimating that “about 47 percent of total US employment is at risk” of computerisation. It is widely repeated as a forecast that 47% of jobs would be automated. This site repeated it that way too, and the note is blunt about where: “in the middle of a section criticising other people for misusing it.”Site data: this site’s own correction of 8 September 2026, noted in place on which jobs are safe from AI, quoting Frey and Osborne, The Future of Employment, Oxford Martin School, 2013.
The same slip happened with a word rather than a number: a paper saying AI’s human-like manner “may reduce human skepticism” was reported here as finding that it does. The note: “The paper frames it as a risk, not a measured effect.”Site data: this site’s own correction of 12 September 2026, noted in place on the critical-thinking page.
The hedge is part of the number. Quote it with the number, or do not quote the number.
6 · Is it still the latest figure?
Official statistics are replaced on a schedule. A figure that was right when you read it can be superseded by the time you repeat it.
A jobs page here quoted US employment projections. The note: “BLS published the 2025-35 round on 27 Aug 2026, superseding every figure in the passage.”Site data: this site’s own correction of 12 September 2026, noted in place on which jobs are safe from AI.
The same update caught a correction going stale. The page had been criticising other reporting for calling nurse practitioners the fastest-growing job. Under the new projections the note on it says that “nurse practitioners ARE 41% and ARE first, so the page was criticising reporting that had become correct.”Site data: this site’s own correction of 12 September 2026, noted in place on which jobs are safe from AI.
A correction is a number too. It has a date, and it can go out of date.
Using the six questions
You will not ask all six every time. Which year and was it rounded are the quickest — seconds each once the original source is open. That is the rule underneath all six: open the source the number came from. Every mistake above was found that way, by reading the original rather than the retelling.
For a number inside a screenshot, start with how to check a viral screenshot. For a source an AI gave you, see how to check an AI citation. How this site checks its own figures is on how to check the figures here.
What this page could not verify
- Whether these six are the most common errors in general. They are drawn from this site’s own corrections, and this site has not found a study measuring the wider pattern.
- The external figures quoted in these corrections were read at source on the dates given, not re-read for this page.
Every example on this page is a correction this site made to one of its own pages. Each is fixed in place on the page linked beside it, and was checked against the source named there on the date given.
Errors on this page are corrected in place too, with a note where the change matters.