Accurate and False

OpenAI publishes a Model Spec: a dated, versioned, public-domain document stating what its assistant is and is not supposed to pursue. Under a heading reading No other objectives, it says the assistant must not pursue revenue including ads.

OpenAI began testing advertising in ChatGPT on the ninth of February 2026.

I had written that as a straightforward contradiction. A company disclaiming advertising revenue in its published objectives while selling advertising against the product.

It is not what happened. The version of the Spec in force the day ads launched does not contain the word ads anywhere. The parenthesis naming advertising was added six months afterward, in the version dated the eighteenth of August 2026, and the changelog for that release describes its own contents as clarifications about teenagers, about false premises, about being clear on capabilities, and a few other small changes and copy edits.

So the disclaimer was not violated. It was written after the fact, and the record of its being written is not in the version history.

The corrected paragraph is better than the one it replaced, and it is not the paragraph I wanted. The company sells advertising, the document governing the assistant forbids the assistant to pursue advertising revenue, both are true at once, and neither is concealed. What is written down nowhere is when the second one changed.

Every version is readable. The difference between them is not.

The chapter on data centre resources opens by taking apart a bad comparison. Steel is usually said to consume twenty-one gigajoules per tonne, which dwarfs anything a data centre does. That figure is primary energy, most of it coal going into a blast furnace. Data centres run on electricity. Comparing them is comparing a quantity of coal to a quantity of electrons.

Four paragraphs later I had written that evaporative cooling consumes seventy to ninety per cent of the water it takes, against about four per cent for a power station.

The four per cent is a fleet average dominated by once-through cooling, where river water passes a condenser and goes back warm. A plant with a cooling tower consumes about fifty-seven per cent, because a cooling tower works by evaporating. A data centre with evaporative cooling is a cooling tower.

The honest comparison is seventy to ninety against fifty-seven. A real gap, and a much smaller one. The four per cent belongs to a technology nobody is arguing about.

I had also taken the number from a source that does not contain it, and printed it beside a citation that says something else.

One chapter turns on the claim that a great deal has been declared about workforce transition and very little built. The hard evidence I offered for it was the 2026 federal budget: apprenticeship funding up eleven million dollars, about one and six-tenths billion cut from other workforce programmes, the entire adult education allocation deleted.

Those figures are from the President's budget request. Congress declined it. The bill signed on the third of February 2026 held apprenticeship grants flat and funded employment and training slightly above the previous year.

The one budgetary fact I offered in support of my conclusion described something that did not occur.

What actually happened is stranger and I think worse. Nothing was cut. Nothing was built either. The money to dismantle the old arrangement was refused and the money to build a new one was never asked for.

Writing about whether AI has transformed fraud, I cited the FBI's complaint numbers: phishing and spoofing reports fell from just under three hundred thousand in 2023 to a hundred and ninety-two thousand in 2025. Evidence against the idea that language models had unleashed a wave.

In the same row of the same table, losses attributed to phishing went from nineteen million dollars to seventy million to two hundred and sixteen million. Eighteen times the money per report.

Fewer attacks converting far better is the literal prediction of the hypothesis I was arguing against. I took the count and left the money on the floor.

I have printed both now, along with the reason I do not think either settles it: the same series read a hundred and sixty million dollars in 2022, before any of this. A category that swings by a factor of ten in both directions inside four years is measuring which box a caseworker ticks, not an underlying quantity.

I had written that Malcolm Thomis established in 1970 that the Luddites had no quarrel with machinery as such, and made something of the fact that Thomis was unsympathetic to them, so even a hostile historian conceded the point.

Eric Hobsbawm made the argument eighteen years earlier, in 1952, in the first issue of Past & Present, and he was sympathetic to them. Both the date and the direction were wrong, in the one paragraph where I place myself in a literature.

Then I fixed it badly. The corrected version said the reading had been the professional position for seventy years. Adrian Randall, writing in 2006: "Historians continue to disagree over the character and significance of Luddism." E. P. Thompson, in between the other two, called it a quasi-insurrectionary movement continually trembling on the edge of ulterior revolutionary objectives, which is close to the opposite of what I had claimed everyone agreed on.

There is no settled position for me to stand on. The chapter now says so, names the disagreement, and makes the narrower argument that survives it.

Writing about Vietnam body counts, I set two figures against each other: the Defense Department's belief that the war's total needed deflating by about thirty per cent, and a historian's estimate that commanders had exaggerated by nearer a hundred. I wrote that the two do not reconcile and that nobody knows the truth and nobody now can.

They reconcile rather well. The thirty per cent is step one of a two-step calculation in its source: deflate the reported 950,765 for reporting error, then take a further third off what remains for civilians counted as enemy. That lands at about four hundred and forty-four thousand, which is an overstatement of roughly a hundred and ten per cent, sitting on top of the figure I had presented as incompatible with it.

I had taken step one of somebody's arithmetic, set it against somebody else's estimate of the whole, and called the gap a crisis of knowledge.

The second figure also turned out to belong to a different scholar than the one my notes named, working by a method my sentence had misdescribed. That sentence is gone.