The AI Numbers Everyone Quotes That Nobody Can Source

Some of the most repeated numbers in AI cannot be traced to anybody who measured them. Others can be traced perfectly well, to a source saying something narrower than what is now quoted. This page follows five of them back, and then gives you the method so you can do the sixth.

How a number becomes common knowledge

The mechanism is always the same and it takes about eighteen months. A researcher publishes a careful figure with stated conditions. A news article reports it with the conditions compressed into a subordinate clause. A second article cites the first and drops the clause. A slide cites the second article. A year later the number appears with no citation at all, because it has become something everybody knows.

Nobody lies at any step. Each transformation is a small, defensible simplification. But the compound effect is a figure that no longer matches anything anybody measured, and which is now unfalsifiable because there is nothing left to check it against.

Two structural features of AI make it worse than average here. Real measurements are scarce, because the systems are proprietary, so any number that does exist gets enormous reuse. And the field moves fast enough that a figure correct in 2021 can be wrong by two orders of magnitude in 2026 while still circulating unchanged.

One model emits as much as five cars

The claim as it circulates: training a single AI model emits as much carbon dioxide as five cars over their entire lifetimes.

The origin: Strubell, Ganesh and McCallum, 2019, Energy and Policy Considerations for Deep Learning in NLP, published at ACL. The paper is real, careful, and says something narrower.

Where it drifted: the largest figure in that paper was not for training one model. It was for a neural architecture search — a procedure that trains many model variants in order to find a good architecture. Presenting that as the cost of “training a model” is like quoting the fuel used by an entire season of motor racing as the cost of one journey. The comparison figure also included the manufacture of the cars, not only their fuel, which roughly doubles the automotive side of the ratio.

What to do instead: quote the paper for what it demonstrated, which is that the compute cost of the search-and-tuning process around a model can dwarf the final run, and that almost nobody was reporting it. That finding has held up. Use the figures in the training-energy page for single runs, and note that hardware efficiency per unit of compute has improved substantially since 2019, so the original number is also simply old.

An AI query uses ten times a web search

The claim as it circulates: a query to a chatbot uses around ten times the electricity of a web search.

The chain, which is traceable: the ratio is a modern estimate divided by a very old one. The web-search denominator descends from a figure Google published on its own blog around 2009, of roughly 0.3 watt-hours per search. That number has not been meaningfully updated publicly since, and the intervening seventeen years cover several complete generations of server hardware, so it is not a current figure for anything.

The numerator side is more recent and better documented. Alex de Vries published an analysis in Joule in 2023, The growing energy footprint of artificial intelligence, which derived a per-request energy estimate from hardware shipment assumptions, and the International Energy Agency reproduced figures of this kind in its 2024 electricity report. Both are named, dated and checkable.

Where it drifted: dividing a 2023 estimate by a 2009 disclosure produces a ratio, and that ratio is then quoted as though both halves were measured at the same time on comparable systems. They were not. Both figures are also whole-system estimates that include very different things.

What to do instead: derive the numerator yourself. The arithmetic in the inference carbon page takes accelerator power, facility overhead and throughput and produces energy per token, and you can state every assumption. For the denominator, the honest answer is that no current public figure for the energy of a web search exists, so the ratio cannot be computed at all — which is a more useful thing to say than a number.

A conversation drinks half a litre of water

The claim as it circulates: one query to a chatbot consumes about 500 millilitres of water.

The origin: Li and colleagues, 2023, Making AI Less Thirsty, an arXiv preprint estimating the water footprint of AI training and inference. The paper is real and the methodology is laid out.

Where it drifted: the paper’s own figure was for a conversation of a stated number of exchanges — on the order of twenty to fifty questions and answers — in a specific data centre with a specific cooling design in a specific climate. Quoting it per single query multiplies it by that factor. Water use is also dominated by the cooling technology and the local climate rather than by anything about the model, so a figure for one facility does not transfer to another that uses closed-loop or air cooling.

What to do instead: cite the paper with its conversation length attached, and treat water as a facility question rather than a model question. Operators publish water usage effectiveness in their annual environmental reports, per region, which is the right granularity. See AI water usage.

GPT-4 has 1.76 trillion parameters

The claim as it circulates: GPT-4 is a mixture-of-experts model with about 1.76 trillion total parameters, made of eight experts of roughly 220 billion each.

The origin: OpenAI has never published GPT-4’s parameter count. Its technical report states explicitly that it does not disclose architecture or model size. The figure entered circulation in 2023 through industry commentary and a widely-shared analyst post, and was repeated in interviews. It has never been confirmed by anybody in a position to know who is free to say.

Where it drifted: nowhere — this one never had a source. It went straight from unattributed claim to universal citation, which makes it the purest example on this page. It is now repeated in textbooks, slide decks and news articles with a confidence that no link in the chain ever had.

What to do instead: say that the parameter count is not disclosed, which is both true and interesting. If you need a size figure for an argument, derive a bound: pricing, latency and the roofline in the tokens-per-second page constrain active parameters far more tightly than any rumour does, and a derived bound with stated assumptions is defensible where a rumoured exact figure is not. Note also that for a sparse model the headline total is the less useful of the two numbers anyway — parameter count as marketing covers why.

AI will add trillions to the economy

The claim as it circulates: AI will contribute some number of trillions of dollars to global GDP by a given year, usually quoted as a flat fact.

The origin: these figures come from consultancy modelling exercises. PwC’s 2017 report Sizing the prize is the best-known example and put a figure of $15.7 trillion by 2030 on the global economic impact of AI. The report exists, states that number, and documents its model.

Where it drifted: it is a projection from an economic model with stated assumptions about productivity gains and adoption rates, published before the current generation of models existed. It is quoted as a measurement, without its date, without its assumptions, and usually without the word “projected”. A projection is a conditional statement; stripping the conditions turns it into a different kind of claim entirely.

What to do instead: quote it as what it is — “PwC’s 2017 model projected X by 2030, assuming Y” — or use measured quantities instead. Realised revenue, filed capital expenditure and published productivity studies are all things that happened. See what the big labs spend and AI and economic growth.

Tracing the next one yourself

The method takes about fifteen minutes and works on any statistic. It is worth doing before you put a number in something with your name on it.

1. Search the exact number as a literal string, in quotes, including its units. You are looking for the earliest occurrence, not the best-ranked one. Restrict by date if the search engine allows it and work backwards.
2. Follow the citation on each page you find, not the page’s own claim. Repeat until you reach either a primary source or a page that cites nothing. The second outcome is common and is itself the finding.
3. Open the primary source and find the sentence. Not the abstract, not the press release — the sentence in the body that contains the number, and the paragraph around it stating the conditions.
4. Check four things about it: what exactly was measured, over what period, under what conditions, and with what stated uncertainty. Most drift happens because one of these four was dropped.
5. Check the date against the domain’s rate of change. A hardware efficiency figure from 2019 or a price from 2023 may be correct and irrelevant. Ask what has changed since, not only whether the source was right.
6. Look for the tell of a figure that never had a source: a suspiciously round number, an identical phrasing repeated across unconnected sites, a claim that has been true every year for a decade, or a citation that points at another secondary article. Any of these means stop.
7. Write down what you found next to the number, in your own document, including the dead ends. The next person to use that figure will be you in eight months, and you will not remember.

The honest position, which almost nobody takes, is that for several of the questions people most want answered about AI — how many people use it, how much energy it consumes in total, how often it is wrong in the wild — no reliable public figure exists. Saying so, and saying what it would take to produce one, is a stronger position than any number you cannot defend. The rest of this cluster is an attempt to hold that line: derive it, or name who published it, or say plainly that nobody knows.