Corrections: what we got wrong

What we said: We said Virginia's data center tax exemption returns 48 cents per dollar given up, 'nearly three times the 17-cent average of the state's other sales tax exemptions.' The audit's own table compares it with 41 cents for the average Virginia incentive, and the report rates the exemption 'moderate' and 'similar to' that average. The figure 17 cents appears nowhere in it.

What it says now: 48 cents against 41 cents, with the auditors' own characterisation. The exemption does beat the average, but not by three times, and it still returns less than it gives up.

How it was caught: Our own audit. A first check appeared to confirm the 17-cent figure from an opinion column that cited the audit; reading the audit gave the opposite answer.

What we said: Ten sources added during a renumbering were never linked to the claims they supported, while the pages needing them cited unrelated sources. Clicking a citation on the data center noise question produced a story about county property taxes, and the questions about the AI race, which quote Stanford's AI Index, cited two articles about water scarcity in Oregon and Arizona.

What it says now: Each of those pages cites the sources that actually support it. The nightly review now flags any source that no page cites, which is the cheap half of the same signal.

How it was caught: Our own audit, prompted by Jon asking us to look for more errors of the kinds already found.

What we said: We said the industry's commissioned study modelled 'over $200 billion' in annual tax contributions.

What it says now: The study reports $162.7 billion in federal, state and local tax contributions in 2023, its most recent year.

How it was caught: Our own audit.

What we said: We gave 700,000 litres as the freshwater cost of training one AI model without saying that this is the water evaporated at the building only.

What it says now: Both figures, labelled: about 700,000 litres on site, and about 5.4 million litres once the water used to generate the electricity is counted. They differ by roughly eight times, and the smaller one is the one that travels.

How it was caught: Our own audit.

What we said: We described Virginia's new rules for large customers as requiring them to pay for '60% of generating capacity reserved, used or not', which reads as a block of power plants set aside for data centers.

What it says now: The 60% figure is the share of minimum charges a customer must guarantee with collateral, and only when its credit is not strong enough on its own. It is a credit requirement, not reserved generation.

How it was caught: Our own audit.

What we said: We said a PJM capacity auction price went from $28.92 to $329.17, 'more than ten times higher.' That pairs the starting point of one auction cycle with the result of a different one.

What it says now: The actual sequence is $28.92, then $269.92, then $329.17, then the $333.44 ceiling. The first jump is more than ninefold, which is dramatic enough without combining two auctions.

How it was caught: Our own audit.

What we said: We repeated that Microsoft's newest data center designs 'use no water at all' for cooling in normal operation, and that Google has committed to the same by 2030, citing an article that says neither.

What it says now: The site states the reduction Microsoft does report, and says plainly that the stronger versions of the claim are widely repeated but are not in that source.

How it was caught: Our own audit.

What we said: We said an electricians' union local worked 28 million hours, 'double what they worked a decade earlier', with '600 new apprentices in a single year'. The cited newspaper article contains neither of the latter two figures.

What it says now: Only what the article supports. The other two are left out until they can be sourced.

How it was caught: Our own audit.

What we said: We gave Ohio's tax exemption cost as '$1.6 billion in 2025, roughly $330 per Ohio household' in a way that implied both figures came from the organisation cited. The per-household number was our own arithmetic.

What it says now: The page says the per-household figure is this site's own division of the state total, and is a way to feel the size of it rather than a bill anyone receives.

How it was caught: Our own audit.

What we said: We credited a flexible-load pilot with '9 sites, 49 participants' including Meta and Microsoft. The launch announcement names nine organisations and neither of those two, and gives no participant count.

What it says now: Nine sites, and the nine organisations actually named, with a note that the roster changes.

How it was caught: Our own audit.

What we said: We labelled $64 billion of blocked and delayed projects as a 2025 figure. The tracker's own window runs from 2024 into early 2026. We also stated flatly that the tracker's clients are AI companies.

What it says now: The label gives the real two-year window, and the funding note reports what the tracker says about its own independence, marked as its position rather than an audited fact.

How it was caught: Our own audit.

What we said: Two different projections of 2030 data center electricity use were presented as one range, mixing a sensitivity analysis with a compounded-growth scenario.

What it says now: Both are shown as separate rows with their methods named: 578 to 782 TWh, and 521 to 843 TWh.

How it was caught: Our own audit.

What we said: The size of the Prince William Digital Gateway project was given as 20 million square feet in one place and 23 million in another.

What it says now: About 23 million square feet throughout, matching the approved application.

How it was caught: A check for numbers that contradict themselves across pages.

What we said: Several figures cited documents that do not contain them: two electricity-sales forecasts cited a press release rather than the outlook table they come from, and a PJM demand-growth figure cited a news article about auction prices rather than PJM's own load forecast.

What it says now: Each cites the document the number is actually in, including a newly added citation for PJM's load forecast.

How it was caught: Our own audit.

What we said: Four specific figures could not be found in any cited source: an opposition-group count available only in a paid report, an analyst estimate that 30 to 50% of the 2026 pipeline was at risk, a date attached to a Georgia rate estimate, and four source links that had gone dead.

What it says now: The unsupported figures are removed or attributed to whoever claims them. The dead links are replaced or the sources dropped.

How it was caught: A link check across every source, plus our own audit.

What we said: We removed a true claim by mistake. Finding that the figure 'the US holds about three-quarters of world AI compute against China's 15%' was absent from the source cited for it, we deleted it from four places. The claim is real and was already on this site under a different source, which states exactly that.

What it says now: Restored, cited to the research group that actually measured it, and narrowed to what they measured: share of global GPU cluster performance, by where the hardware physically sits. A neighbouring forecast of US capacity growing to 152 gigawatts by 2030 really is unsupported and stayed out.

How it was caught: Re-checking our own correction before publishing it. Deleting a supported claim is the same kind of error as publishing an unsupported one.

What we said: A published link had a fragment of HTML stuck to the end of it and did not work, and two different documents were stored under the same report identifier.

What it says now: Both fixed, and the daily updater now sanitises URLs and rejects duplicate identifiers rather than trusting that a string starting with 'http' is a usable link.

How it was caught: The first run of the nightly review job.

What we said: Fifteen footnotes on the two argument pages linked to the wrong sources. The handbook and the site had each been numbering sources independently from 31 onward, so a point citing note 32 sent readers to an unrelated source 32 on the sources page.

What it says now: The handbook and the site share one numbering system. A note number in the handbook is now the same source number on the site, and the nightly review checks that they still agree.

How it was caught: Jon asked where the site delivers its promise to say who funded each source, which exposed the fork.

What we said: A funding note had been copied from another source and not rewritten. It said the consultancy E3 had also produced a study paid for by the Data Center Coalition, extending that conflict of interest to a Virginia legislative report E3 never worked on. The citation also pointed at a Department of Taxation document while claiming to be a legislative audit.

What it says now: The note describes only what that document's own evidence supports, and the citation points at the audit it names. Implying a conflict of interest that does not exist is the worst error this site can make.

How it was caught: Our own audit of every funding label.

What we said: The most-cited number on the pro side was wrong. We said the industry's own commissioned study found 5.5 million supported jobs and 4.5 additional jobs for every direct job.

What it says now: The study reports 4.7 million supported jobs and an employment multiplier of 7.5, meaning 6.5 additional jobs per direct job. We overstated the total by 800,000 jobs while understating the multiplier.

How it was caught: Reading the report itself instead of our own summary of it.

What we said: Counts that were typed by hand had drifted: the site advertised a 39-page handbook when the PDF was 42 pages, said 'Seven arguments' when there were eight, and 'Fourteen arguments' when there were sixteen.

What it says now: Every count is derived from the handbook at build time, so it cannot drift again.

How it was caught: A question from Jon about whether the PDF was being kept current.