AI Statistics That Are Wrong
Some of the most quoted numbers in AI do not say what they are quoted as saying. This register traces widely repeated AI statistics back to their sources and records, for each one, what the source actually measured, the figure that survives checking, and how confident we are. It is updated monthly.
Thank you for reading this post, don't forget to subscribe!
Most bad AI statistics are not invented. They drift.
Almost none of the figures below was fabricated. Each began as a real number from a real source, and then changed on the way through the citation chain. The pattern is consistent enough to name. A figure measured on a broad definition gets repeated under a narrow label: phishing emails with any AI involvement become emails written by AI. A unit gets lost in translation: 3.8亿 is 380 million, not 3.8 billion. A forecast becomes a fact when the year it was forecasting arrives, and is then quoted as a measurement. And a figure that was correct when it was published keeps being quoted, without its date, long after the thing it measured has moved.
The fifth mechanism is the simplest and the least excusable: arithmetic nobody does. The best-known example is the claim that a fake image of an explosion at the Pentagon wiped $500 billion off US stocks in May 2023. The same coverage reported the size of the move, and multiplying that move by the value of the S&P 500 gives a number a fraction of the size. The claim survived three years of repetition because checking it required one multiplication.
These errors matter more in AI than in most subjects for a specific reason. AI statistics are increasingly read, summarised and repeated by AI systems, which learn from exactly the repeated, undated, broadened versions this page corrects. A wrong number that appears on a hundred roundup pages is, to a model, a well-attested number. The only counterweight is a source that states the corrected figure plainly, with its date and its evidence, in a form that can be found and quoted.
Two cautions about this register itself. First, inclusion means the figure fails as commonly stated; several of the underlying sources are perfectly sound and are cited with approval on our own pages. Second, our own site published versions of some of these errors before correcting them, and they are recorded in our corrections log. A fact-checker that hides its own mistakes is not one.
- Definition drift is the commonest failure. Four of the thirteen entries describe a broad measurement with a narrow word.
- Dates fall off figures. Three entries were true when published and are now quoted as current.
- Forecasts become facts. “$10.5 trillion of cybercrime in 2025” is a 2020 projection, not a measurement.
- Check the arithmetic. The $500 billion Pentagon-image figure does not survive a single multiplication.
Summary table
| Commonly stated | What survives checking | Verdict |
|---|---|---|
| Fake Pentagon image wiped $500B off US stocks | Reported move implies ~$90–250B | Overstated |
| 82.6% of phishing emails are written by AI | 82.6% showed any AI involvement | Misdescribed |
| AI now writes more web articles than humans | About half, and level since late 2024 | Superseded |
| FBI: 37% rise in AI-assisted email fraud | Not found in the FBI report; AI-linked BEC ~$30M | Untraceable |
| Cybercrime cost the world $10.5T in 2025 | A 2020 projection, not a measurement | Forecast as fact |
| Deepfake fraud will cost $40B by 2027 | Forecast for all gen-AI fraud in the US | Misdescribed |
| Arup’s $25M is the biggest deepfake fraud | Fideuram lost €95M (~$108M), Sept 2026 | Superseded |
| Doubao has 3.8 billion users | 382 million (3.8亿) | Wrong |
| ChatGPT has 2.8 billion monthly users | No disclosure supports it; 900M+ weekly disclosed | Unsupported |
| ChatGPT has 800 million weekly users | True in Oct 2025; 900M+ in 2026 | Out of date |
| Claude has N million users | Estimates range 56M–245M for the same period | Contested |
| 45 states have deepfake laws (cited for elections) | 45–48 is sexual deepfakes; elections 30–33 | Misdescribed |
| Court ruled AI training on books isn’t fair use | It ruled training was fair use; piracy was not | Wrong |
The entries
“A fake AI image of a Pentagon explosion wiped $500 billion off the US stock market.”
Where it circulates — 2023 news coverage (including TheStreet and SFist) and many later roundups.
What the evidence says — On 22 May 2023 the S&P 500 dipped briefly after the fake image spread: about 0.26% on the figure most incident records use, about 30 index points in some reports. On an index worth roughly $34–36 trillion, those moves imply about $90 billion and $250 billion respectively. Reaching $500 billion would require a fall of about 1.4%. We found no source that shows how $500 billion was calculated.
Detail — AI Market Manipulation.
“82.6% of phishing emails are written by AI.”
Where it circulates — Security vendor blogs and statistics roundups, usually attributed to KnowBe4’s 2025 phishing research.
What the evidence says — The figure counts emails showing any AI involvement, including personalisation or obfuscation, not emails written by AI. A separate estimate puts business-email-compromise messages primarily written by AI at about 40%. AI is involved in most phishing; it does not write 82.6% of it.
Detail — AI in Cybercrime.
“More articles on the web are now written by AI than by humans.”
Where it circulates — Widely reported from an SEO firm’s 2025 study, and sometimes misattributed to Ahrefs, whose different finding is that 74.2% of new pages contain some AI content.
What the evidence says — Graphite’s original study estimated that articles written primarily by AI first outnumbered human-written ones in November 2024. Its own follow-up, with data through the first quarter of 2026, found the share has levelled off near 50% — 50.9% in the fourth quarter of 2025 and 49.9% in the first quarter of 2026. The accurate statement is “about half, and no longer rising”.
Detail — AI Slop.
“The FBI reported a 37% rise in AI-assisted business email compromise.”
Where it circulates — Secondary security blogs citing the FBI’s 2025 Internet Crime Report.
What the evidence says — We could not find this figure in any account of the report’s AI section. That section counts 22,364 complaints with an AI descriptor and $893 million in losses, of which business email compromise involving AI was about $30 million. The 37% may exist somewhere we could not reach, but until it is traced to the report it should not be attributed to the FBI.
Detail — Deepfake Fraud.
“Cybercrime cost the world $10.5 trillion in 2025.”
Where it circulates — Industry reports, conference keynotes and some major publications, attributed to Cybersecurity Ventures.
What the evidence says — The figure comes from a 2020 projection that costs would grow about 15% a year to reach $10.5 trillion in 2025. When 2025 arrived, the projection began to be quoted as the outcome. The firm publishes its cost categories but not its calculations. For scale, cybercrime losses reported to the FBI for 2025 were $20.9 billion, though that covers the US and only reported cases.
Detail — The Dark Side of AI in Numbers.
“Deepfake fraud will cost $40 billion by 2027.”
Where it circulates — Deepfake statistics pages and vendor marketing, attributed to Deloitte.
What the evidence says — Deloitte’s Center for Financial Services projected that generative-AI-enabled fraud losses in the US could reach $40 billion by 2027, up from $12.3 billion in 2023. It is a forecast, it covers all generative-AI-enabled fraud rather than deepfakes alone, and it is US-only. It resolves in 2027; we track it on our predictions scoreboard.
“The $25 million Arup video call is the biggest deepfake fraud.”
What the evidence says — It was, from early 2024. In September 2026 Reuters reported that fraudsters impersonating senior executives took €95 million (about $108 million) from Fideuram, the private-banking arm of Intesa Sanpaolo, using a WhatsApp message and a cloned voice. About €36 million is reported unrecovered.
“Doubao has 3.8 billion users.”
Where it circulates — At least one English-language outlet published it as a headline.
What the evidence says — The Chinese figure is 3.8亿. One 亿 is a hundred million, so 3.8亿 is 382 million, the monthly active user figure tracked by QuestMobile. The headline overstated it tenfold. The unit conversion is certain; the 382 million itself is a device-panel estimate, rated MEDIUM on our user-numbers page.
Detail — AI user numbers, US vs China.
“ChatGPT has 2.8 billion monthly users.”
What the evidence says — We found no disclosure or measurement that supports it. OpenAI’s own most recent figure is more than 900 million weekly active users in 2026. A monthly figure would be higher than a weekly one, but 2.8 billion would be roughly a third of the world’s population, and no source behind the claim shows how it was derived.
Detail — ChatGPT Statistics.
“ChatGPT has 800 million weekly users.”
What the evidence says — Correct for October 2025, when OpenAI announced it. The 2026 figure is more than 900 million. The series runs 400 million (February 2025), about 700 million (July 2025), 800 million (October 2025), 900 million-plus (2026). Figures calculated against 800 million at the time, such as OpenAI’s October 2025 disclosure about suicide-related conversations, should keep that denominator; it is the current-tense use that is wrong.
Any single figure for Claude’s user numbers.
What the evidence says — Published estimates for Claude’s monthly users in the same period range from 56 million to 245 million, a 4.4-fold spread, because they mix weekly and monthly measures, web traffic and app installs. Anthropic does not publish a consumer user count. Any single number should be quoted with its method, or as a range.
Detail — AI user numbers, US vs China.
“45 states have deepfake laws” — cited as protection for elections.
What the evidence says — State counts depend on the category. Roughly 45–48 states address sexual deepfakes. Election-deepfake laws exist in about 30–33, depending on the tracker; Public Citizen counted 31 in October 2026. Quoting the larger number for elections overstates the protection voters have.
Detail — Synthetic Deception.
“A court ruled that training AI on copyrighted books isn’t fair use — and Anthropic paid $1.5 billion.”
What the evidence says — In Bartz v. Anthropic (June 2025) Judge William Alsup held that training was fair use, calling it “exceedingly transformative”, and that scanning lawfully bought books was fair use too. What he refused to excuse was downloading millions of pirated books and keeping them as a permanent library. That issue was headed to trial when Anthropic settled for about $1.5 billion. The liability came from how the books were obtained and stored, not from training on them.
Detail — Copyright and AI.
How we check a figure
Each entry follows the same steps. We find the earliest published version of the claim and the source it cites; read what that source says it measured, including its definitions, sample and dates; redo any arithmetic the claim depends on; and look for a newer edition of the same measurement. A figure is listed only when the commonly stated version differs materially from what the source supports.
| Verdict | Meaning |
|---|---|
| Wrong | The figure or the finding is incorrect as stated. |
| Overstated | The source supports a materially smaller number. |
| Misdescribed | The number is right but describes something different from what it is quoted as measuring. |
| Superseded | A newer measurement, often from the same source, replaces it. |
| Out of date | Correct at its date; quoted as current. |
| Untraceable | Attributed to a source in which we could not find it. |
| Unsupported | No disclosure or measurement we could find supports it. |
| Contested | Credible estimates disagree too widely for any single figure. |
| Forecast as fact | A projection quoted as a measured outcome. |
Frequently asked
Why are so many AI statistics wrong?
Mostly through drift rather than invention. Figures measured on broad definitions are repeated under narrow labels, units are lost in translation, forecasts are quoted as outcomes once their year arrives, and correct figures keep circulating without their dates. Few people redo the arithmetic behind a striking number.
Did a fake AI image really wipe $500 billion off the stock market?
Probably not. The S&P 500’s reported dip in May 2023 — about 0.26%, or about 30 points in some accounts — implies roughly $90 billion to $250 billion on an index worth $34–36 trillion. Reaching $500 billion would need a fall of about 1.4%.
How many people actually use ChatGPT?
OpenAI’s latest disclosed figure is more than 900 million weekly active users in 2026. Claims of 2.8 billion monthly users are not supported by any disclosure we found, and the widely quoted 800 million was the October 2025 figure.
Is AI really writing most phishing emails?
AI is involved in most phishing, but the often-quoted 82.6% counts any AI involvement, including personalisation, not emails written by AI. Estimates for emails primarily written by AI are much lower, around 40% for business email compromise.
Can I suggest a figure to check?
Yes. Send the figure and where you saw it through our contact page. We review submissions in the monthly update.
Each entry links to the Report AI page where the full sourcing sits. Confidence follows our editorial standards: HIGH where the correction rests on something certain (such as a unit conversion), MEDIUM where it rests on consistent secondary accounts of a source we could not open directly, and DERIVED where it rests on our own arithmetic, with the inputs shown. Several primary sources — court dockets, regulators and some publishers — are not reachable from our research environment, which is why most entries are MEDIUM. “Where it circulates” describes where we found the claim; it is not a list of every publication that repeated it. Cadence: monthly. Corrections: corrections log.
Corrections log · AI predictions, scored · Editorial standards · AI user numbers · The Dark Side of AI · All indexes