Six recurring sources account for most of the numbers on this site. This page names the feed or document behind each, how to re-query it, and what to expect when your figure differs from ours.
- Take the edition, not the coverage. Annual reports describe the previous year, so the headline number and the year it covers are usually one year apart.
- Leaderboards and star counts move weekly. Expect a different figure when you re-query, and do not mix benchmark generations.
- One study is one study. The detector false-positive result is strong evidence against accusing anyone on a score, not a universal constant.
- A projection is not a measurement. Read an energy forecast's assumptions, and go to the regulation itself rather than a summary of it.
- If your number differs, check the date first. Then the edition, then whether it is a projection; if it still differs, send it in as a correction.
Six sources account for most of the recurring numbers on this site, and every one can be re-queried by you. This page names which feed, which document, and what to expect when your figure differs from ours.
Six recurring sources — the ones that appear across multiple pages and carry most of the weight. One-off figures cited on a single page are sourced inline on that page instead.
This is not a claim that every number on the site is reproducible. Some figures come from reports without published underlying data, and where that is true the page saying so is the honest signal.
Citation asks you to trust us. Reproducibility removes the need to. A citation says where a number came from; this page tells you how to get it again without us.
If you recompute a figure and get a different answer, that is a correction and we want it. Contact is on the press page; anything confirmed is corrected on the page it affects.
The sources, and how to re-query each
Automated traffic share — Imperva Bad Bot Report
Used for: the 51% and 53% automated-traffic figures, and the “machines outnumber people” claim.Imperva, 2026 Bad Bot Report launch post, read at source 23 Sep 2026: “accounting for more than 53% of all web traffic in 2025, up from 51% the year before”.
How to check it: each annual edition is a free download from Imperva’s site, behind a contact form, and restates the prior calendar year; the launch blog post carries the headline figures without the form. Take the edition, not the press coverage — the headline number and the year it describes are usually one year apart, which is where most misquotes come from.
Expect disagreement between editions. Imperva's own methodology changes; the direction is comparable across years, the decimals are not. Any page here quoting these figures says which edition it used.
Agent benchmark scores — OSWorld and OSWorld 2.0
Used for: the 12% → 66.3% → 20.6% sequence that this site treats as its central example. Re-queried 10 Sep 2026: the OSWorld 2.0 board now tops out at 77.9%, a score Anthropic itself labels partial credit (41.7% strict) — the sequence moved again within a month, which is the point.OSWorld 2.0 leaderboard, last updated 4 Sep 2026, read at source 10 Sep 2026: “Claude Fable 5.1 77.9%, Simular Sai 73.0%, GPT-6 Astra 72.6%”. The board warns that “rows mix benchmark-author runs, vendor self-reports, and independent co-author runs”.
How to check it: both are public leaderboards. Re-query and you will likely get a different top score — leaderboards move weekly, and that is exactly the point being made.
What to hold constant when you check: the benchmark generation. A score from OSWorld and a score from OSWorld 2.0 are not comparable, and treating them as one series is the error the whole page exists to describe.
Detector false-positive rates — Liang et al., Patterns
Used for: the 61.3% figure for non-native English writers.Liang et al., Patterns 4(7):100779, July 2023, read at source 23 Sep 2026: “(average false-positive rate: 61.3%)”.
How to check it: the paper is open access — Patterns 4(7):100779, July 2023. The seven detectors, the 91 TOEFL essays and the per-detector breakdown are all in it.
What this figure does not say: it is one study, one essay set, one point in time. It is strong evidence that detector output is unsafe to accuse anyone with. It is not a universal constant, and this site does not present it as one.
GitHub repository metrics
Used for: star counts and licence claims in the repo analysis.
How to check it: api.github.com/repos/OWNER/NAME returns stargazers_count and a license object. No key needed for public repositories.
Expect our number to be low. Star counts move weekly; anything quoted here is a snapshot with a date attached, and a stale one within a month.
Energy projections — IEA Energy and AI
Used for: the 945 TWh by 2030 figure, and the updated 950 TWh.
How to check it: free report, April 2025. iea.org serves a bot-verification challenge to automated and headless access, but the report PDF opens directly. The 950 TWh seen elsewhere is not a rounding of 945: it is the agency’s own updated projection, in its April 2026 Key Questions on Energy and AI. Quote whichever edition you used, and say which.Both read at source 17 Sep 2026. IEA, Energy and AI, 2025: “Data centre electricity consumption is set to more than double to around 945 TWh by 2030.” IEA, Key Questions on Energy and AI, 2026: “Our updated projections see electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, accounting for around 3% of global electricity demand by that date.” Read the assumptions section, not the headline — it is a projection built on stated growth rates, and the report says so.
A projection is not a measurement, and this site labels it that way wherever it appears.
Regulatory dates — EU AI Act
Used for: every date and fine ceiling on the rules page.
How to check it: Regulation (EU) 2024/1689 and its amendments are on EUR-Lex; the Commission publishes its guidelines and enforcement material directly. Go to the regulation for obligations and to the guidelines for interpretation — law-firm summaries are useful but they compress, and the compression is where errors enter.
If your number differs from ours
- Check the date first. Most disagreements are two people holding figures from different years.
- Check the edition or generation. Imperva editions and OSWorld generations are not interchangeable.
- Check whether it is a projection. Several widely quoted numbers are forecasts that were later repeated as measurements.
- Then tell us. A figure that cannot be reproduced should not be on this site, and once confirmed it is corrected in place.
No original measurement. Nothing here is measured by this site. Every figure is someone else's work, named and dated, and the value added is the collation and the caveats — not the data.
That is also why the method page matters: it states which sources qualify and what is excluded, so you can judge the selection as well as the figures.