How Many People Use AI? What the Numbers Can and Cannot Tell You

There is no authoritative count of AI users, and the figures that circulate are answers to at least four different questions. This page does not give you a percentage. It gives you the publishers who produce real ones, the frame each of them can support, and a way to test whether a number somebody quoted at you is even physically possible.

There is no single number, and here is why

Counting users of a technology requires either a census of the providers or a probability sample of the population. Nobody has the first: the providers are private companies under no obligation to report, several of the largest models are open weights running on hardware nobody tracks, and a great deal of AI use now happens inside products that never say the word to the user — search summaries, autocomplete, spam filters, photo tools.

The second exists but is narrower than people assume. National statistics offices and a handful of research organisations run properly sampled surveys, and they ask about self-reported use of named products in a named country over a named period. That is a real number. It is not “how many people use AI”.

So the figures in circulation come from three places: company disclosures, which are marketing statements rather than audited filings; modelled estimates from traffic and app-store analytics firms, which are extrapolations from panels; and surveys, which measure what people say. All three can be useful. None of them is a count.

Four different questions in one phrase

Before comparing two usage figures, check that they answer the same question. Most disagreements between headline numbers dissolve here.

The denominator matters as much as the numerator. A percentage of adults in one country, a percentage of internet users worldwide, and a percentage of knowledge workers are three different bases, and the same underlying fact produces wildly different headlines depending on which one is used.

Who actually publishes a figure

These are recurring, methodologically documented sources. Go to them rather than to an aggregator article, because the aggregator always drops the frame.

• Pew Research Center. Runs probability-based panel surveys of US adults and publishes its questionnaire, its field dates and its margin of error alongside every result. If you need a defensible statement about what American adults say they do with AI, this is where it comes from.
• Eurostat. Publishes the annual ICT usage surveys for households and for enterprises across EU member states, with the question text and the national sample sizes. The enterprise survey has carried an AI-use question in recent waves, which makes it the closest thing to an official European adoption series.
• The US Census Bureau’s Business Trends and Outlook Survey. A probability sample of businesses, run at a high frequency, that asks firms directly whether they used AI to produce goods or services in the last two weeks. Because it samples the business register rather than a panel of volunteers, its answers are usually far lower than survey figures from consultancies — a gap covered in the adoption-statistics page.
• National statistics offices and regulators. The UK’s ONS and Ofcom, and their equivalents elsewhere, publish technology-use series with documented sampling. These are slow, dull and correct, which is exactly what a citable figure needs to be.
• The Stanford HAI AI Index. An annual report that aggregates other people’s data and, unusually for an aggregator, cites each series to its original publisher. Useful as an index into primary sources rather than as a source itself.
• Company statements. Weekly-active figures announced in blog posts and at developer events are real information, but they are unaudited, the definition is chosen by the company, and the comparison base changes silently between announcements. Quote them as “X said in month Y”, never as a measurement.

What question wording does to the answer

Question wording moves survey results more than almost anything else about the survey, and AI questions are unusually sensitive to it because the term is unstable. Four wordings that will each produce a different answer from the same person:

• “Have you used artificial intelligence?” Many respondents read this as a claim about robots and say no while holding a phone that does on-device translation.
• “Have you used ChatGPT or a similar chatbot?” Naming a product raises the number sharply, and biases it towards whichever product is named.
• “In the last week, did you use a tool that generates text or images?” Behavioural, time-bounded, and much harder to over-claim on. Lower and more reliable.
• “Does your employer use AI?” A question about somebody else’s behaviour, answered by a person who may not know. This wording produces the widest spread between surveys.

Two more effects worth naming. Self-report over-states socially desirable behaviour and under-states invisible behaviour, and AI is currently both desirable in some populations and invisible in most products. And an opt-in web panel recruited through advertising is disproportionately made of people interested in technology, which inflates every technology-use question by an amount nobody can quantify after the fact.

A plausibility check you can run yourself

A claimed user count implies a quantity of inference, and inference implies hardware. That chain is arithmetic, so you can test whether a headline is even possible. This will not tell you a claim is true, but it catches the ones that are impossible, which is more than most fact-checking of these numbers achieves.

Claim under test: N weekly users of a chat product.

Assumptions, all of which you must state when you use this:
  messages per user per week   m = 10
  output tokens per answer     t = 300
  seconds in a week            s = 604,800

Implied sustained output rate:
  tokens_per_second = N * m * t / s

Worked at N = 500,000,000:
  500e6 * 10 * 300        = 1.5e12 output tokens per week
  1.5e12 / 604800         = 2.48e6 output tokens per second, sustained

Now convert to hardware. Aggregate serving throughput per accelerator
depends enormously on model size and batching; take a band rather than
a point:

  optimistic  4,000 output tokens/s per accelerator (small model, deep batching)
  pessimistic   400 output tokens/s per accelerator (large model, tight latency SLO)

  2.48e6 / 4000 =    620 accelerators, busy 24/7
  2.48e6 /  400 =  6,200 accelerators, busy 24/7

Then divide by average utilisation. Traffic is diurnal; a service sized
for peak at 30% average utilisation needs roughly 3x the above.

The result is the interesting part. Half a billion weekly users, at these assumptions, is a fleet in the low thousands to low tens of thousands of accelerators. That is entirely achievable for a large provider, so the claim passes the check. Run the same arithmetic on a claim of a billion daily users each having long conversations with a frontier model, and it stops passing.

The per-accelerator throughput band is the load-bearing assumption, and you can derive your own rather than borrowing this one — the roofline calculation gives the ceiling from memory bandwidth and parameter count, and the FLOPs-per-answer derivation gives the arithmetic side.

Before you quote a usage figure

1. Find the primary publisher. If the trail ends at a blog post citing another blog post, you have no figure. The technique for tracing one back is in the page on AI statistics nobody can source.
2. Read the field dates, not the publication date. A report published in March may be reporting fieldwork from the previous autumn, which in this field is a different era.
3. Identify the population sampled and the denominator used. “US adults” and “global internet users” are not interchangeable.
4. Find the question text. If the publisher does not release it, treat the result as directional at best.
5. Check whether the sample is a probability sample or an opt-in panel. Only the first supports a margin of error, and a margin of error quoted on an opt-in panel is decorative.
6. Quote it with all of that attached: publisher, field dates, population, question. A figure that cannot carry its frame in one sentence is a figure you should not be quoting.