Ten numbers that appear in almost every presentation about AI and work, traced back to what their sources actually say. Of the first nine, four are misquoted. Two cannot be traced to any stated methodology. One is a susceptibility estimate reported as a prediction. Two survive intact, and they are worth more than the rest combined. The tenth, added in September 2026, is three accurate figures for the same product that disagree by more than thirty points, because each counts something different.
The answer, in one line
No. Their 2013 paper is titled 'How Susceptible Are Jobs to Computerisation?' and estimates that about 47 per cent of total US employment is 'at risk' in the sense of being technically susceptible.
Five of these were found by checking claims on this site, so the list starts with our own errors rather than other people's.
1 · "47 per cent of jobs will be automated"#
What the source says. Frey and Osborne's paper is called How Susceptible Are Jobs to Computerisation? Their sentence is: "about 47 percent of total US employment is at risk". They estimated the probability of computerisation for 702 occupations using a classifier trained on 70 occupations hand-labelled at an Oxford workshop.
What is wrong with the popular version. It is a susceptibility estimate, not a forecast. No date is attached to any loss. The classifier scores whole occupations, so an occupation counts as at risk even where most of its tasks are not. And the paper is a 2013 working paper published in a journal in 2017, so the two dates get used interchangeably for the same finding.
2 · The same question, answered as 9 per cent#
The OECD asked it again with a different unit of analysis. Arntz, Gregory and Zierahn modelled tasks within occupations across 21 countries and found 9 per cent of jobs automatable on average, from 6 per cent in Korea to 12 per cent in Austria.
Their stated reason: occupations labelled high-risk "often still contain a substantial share of tasks that are hard to automate".
A roughly fivefold difference on the same question, produced by a modelling decision rather than by new evidence. Neither number has been scored against what actually happened, which is the fact that should govern how confidently either is quoted. See what should I tell my children to study for the one forecaster that does mark its own homework.
3 · "70 per cent of change programmes fail"#
Mark Hughes went looking for the evidence and published what he found in the Journal of Change Management in 2011. He reviewed five separate published instances of the figure and concluded: "there is no valid and reliable empirical evidence to support such a narrative."
The citation trail loops. Reports cite consultancies, consultancies cite Kotter, and Kotter's number was an informal estimate rather than a measurement. Fifteen years after the paper that took it apart, the statistic is still quoted in slides about AI transformation.
Worth being precise about what this does and does not show. It does not mean change programmes usually succeed. It means the true rate has never been established, and the number standing in for it was invented.
4 · "80 per cent of workers will be affected by AI"#
From Eloundou and colleagues, and the graded entry in this research already describes it as the most misquoted number in the field. It measures exposure: the share of workers with at least 10 per cent of tasks where an LLM could reduce completion time. Exposure is not displacement, and the authors say so directly.
5 and 6 · Two numbers that cannot be traced at all#
That average attention on a screen has fallen to about 47 seconds. And that refocusing after an interruption takes 23 minutes.
Neither could be traced to a peer-reviewed paper reporting the figure with a stated methodology. Both circulate through trade press and book promotion. They may well be approximately right, and they are not currently checkable, which is a different thing from being false.
A peer-reviewed interruption study points the other way, finding people compensate by working faster with no measured loss of output quality, at the cost of higher stress. See is attention a trainable skill.
7 · Percentage points quoted as per cent#
The most common distortion in the whole set, and the easiest to miss. BCG's diversity study found innovation revenue 19 percentage points higher: 45 per cent of total revenue against 26 per cent. It is routinely cited as "19 per cent higher", which is a materially smaller and different claim.
This one is on our list because this research made the same error before correcting it.
8 · The figure that lives in the press release#
The curiosity meta-analysis by von Stumm and colleagues is almost always cited as covering roughly 50,000 students. That number appears in the accompanying press release rather than in the paper.
Press releases are written to be quotable and are not peer reviewed. When a widely repeated figure does not appear in the study it is attributed to, the release is usually where it came from.
9 · A number that changed and nobody updated#
Work sample tests were, for decades, cited at a validity of .54, making them among the best predictors of job performance. Sackett and colleagues corrected the underlying meta-analytic method in 2023 and the figure fell to .33. Structured interviews moved from .51 to .42 and became the strongest single predictor.
The old numbers are still in circulation, in textbooks and in assessment marketing. This is the least visible failure mode: not a misquote, but a superseded finding that nobody went back for.
10 · Three trackers, one spring, three market shares#
What gets quoted. A single figure for "ChatGPT's market share", usually whichever one suits the slide. In the spring of 2026 three measurement firms each published one, all three are reputable, and they put the same product anywhere from under half the market to more than three quarters of it.
What each one counts. StatCounter reported ChatGPT at 76.85 per cent for April 2026. Its unit is the referral: a visit that reaches one of the websites carrying its tracking code from a link clicked inside a chatbot, drawn from over 3 billion page views a month across over a million sites. Similarweb's estimate, published by Momentic in August 2026, gave ChatGPT 53.9 per cent for May 2026, counted as its share of web visits to seven assistants' own sites, with the mobile apps left out. Sensor Tower counts unique people across the apps and the web together, a measure it calls True Audience, and reports that ChatGPT's share on that basis fell below 50 per cent for the first time in March 2026.
Why nobody is wrong. A referral exists only when a conversation sends somebody on to another website, so that count favours whichever assistant links out most and says nothing about the conversations that end where they started. StatCounter said in 2025 that it could not count Grok at all, because Grok passes no referral information. A web-visit share misses the apps, and Momentic's own report notes that app use adds between 15 and 60 per cent on top of web visits depending on the assistant. A unique-user count registers a person once, whether they opened the app once in the month or forty times a day. Each of the three is a fair answer to its own question, and the phrase "market share" names none of those questions.
The figure left out. The lowest number in circulation, 46.4 per cent at the end of May, was reported by TechCrunch on 16 June 2026 as Sensor Tower's. Sensor Tower's own published summary says below half and prints no figure, so this page reports the number as TechCrunch's and does not rest anything on it. Entry 8 describes the same route, one step further from the source.
The series also keep moving on their own schedules. StatCounter's live page, read on 28 September 2026, had ChatGPT back up at 79.4 per cent for August. Any of these figures quoted without its month and its unit is close to meaningless, and the question a reader should ask of it is what, exactly, was counted.
What survives#
Two figures in this territory hold up under checking, and both are narrower than the use they get put to.
Brynjolfsson, Li and Raymond's 15 per cent productivity gain is a real measurement from a real deployment, concentrated among novice workers. Budzyn and colleagues' finding that unassisted adenoma detection fell from 28.4 to 22.4 per cent after AI exposure is observational rather than randomised, and the authors say so, but it is the strongest direct evidence of professional deskilling anyone has produced.
How to check one yourself#
- Find the sentence in the source. Not the abstract, not the summary, the sentence containing the number. Most distortions are visible immediately.
- Check the unit. Percentage points or per cent. Exposure or displacement. Susceptibility or forecast. Tasks or occupations. For usage figures, referrals, visits or people. Most errors here are unit errors.
- Ask what the sample was. A number without a sample is a claim without evidence, however institutional the logo above it.
- Follow the citation twice. Once is not enough. The 70 per cent figure survives because everyone stops at the first hop.
- Check whether it came from the press release. If the figure is not in the paper, that is usually the reason.
- Check the date of the method, not the report. A 2026 report can rest on a 2013 estimate whose approach has since been superseded.
Why this page exists#
Every number here was traced because it was about to be used on this site, or already had been. Five of the first nine are corrections to our own published work rather than criticisms of other people's. The tenth was checked because a slide used in Rahim's own teaching carries the comparison.
If you find an error here, the commitment is the same one that applies to every page: it gets fixed on the page where it was made, with the date shown.
Related SuperSkills research#
For the graded studies behind every claim on this site, the evidence base. For what the research establishes and how strongly, what we actually know. On the attention figures, is attention a trainable skill. On forecasting records, what should I tell my children to study. On the distributional question the productivity numbers cannot answer, who captures the productivity gains from AI.
Key research and primary sources
- Frey, C. B. and Osborne, M. A. (2013). The Future of Employment: How Susceptible Are Jobs to Computerisation? Oxford Martin School.
- Arntz, M., Gregory, T. and Zierahn, U. (2016). The Risk of Automation for Jobs in OECD Countries. OECD Working Paper No. 189.
- Hughes, M. (2011). Do 70 Per Cent of All Organizational Change Initiatives Really Fail? Journal of Change Management, 11(4).
- Eloundou, T. et al. (2024). GPTs are GPTs.
- Sackett, P. R. et al. (2023). Revisiting the design of selection systems.
- von Stumm, S. et al. (2011). The Hungry Mind.
- Lorenzo, R. et al. (2018). How Diverse Leadership Teams Boost Innovation. BCG.
- StatCounter Global Stats (2026). ChatGPT falls to all-time low as AI chatbot referral market continues to fragment. Press release, 7 May 2026.
- Einberger, T. (2026). Top Generative AI Chatbots and LLMs by Market Share. Momentic, 18 August 2026, from Similarweb estimates.
- Huang, Y. (2026). 2026 State of AI. Sensor Tower, June 2026.
About this research#
Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Every figure on this page was checked against the primary source before publication, and the sources are linked so the checking can be repeated. Five of the first nine entries record errors in this research's own earlier work.
Reference · SS-2026-103
Hirji, R. (2026). The most-quoted AI statistics, checked. The SuperSkills evidence base, SS-2026-103. https://thesuperskills.com/research/the-most-quoted-ai-statistics-checked. Last reviewed 28 September 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work