Back to Explore
NewsVibeFix editorialUpdated Nov 20, 2026

The Census Bureau renamed AI 'Super Intelligence': how much of big firms' 40% adoption rate is just a changed question

Simon Willison noted on November 2 that the Census Bureau's business AI survey changed its question twice in a year — and renamed AI 'Super Intelligence' in October 2026. Large firms' 40.7% is real, but it was not measured with the same ruler as the 12% of a year ago.

Simon Willison's line chart of U.S. business AI use by firm size, marked 'Question broadened Nov 2025' and 'Series not comparable', with 250+ firms at 40.7%.

On November 2, developer Simon Willison updated a chart on his blog: a line chart of AI adoption among American businesses. The line for the largest companies — those with 250 or more employees — had climbed all the way to 40.7%. Read on its own, the story writes itself: AI is taking over American business at astonishing speed. But the real reason Willison updated the chart is the grey vertical bar in the middle of it, and the small print sitting on top of it: "Question broadened. Series not comparable." What this chart is really about is not how hot AI is, but how the most-cited official AI numbers in the United States had their survey question changed twice in a year — and renamed once — and are still being read as if they were one continuous trend line.

This deserves to be taken seriously, because these numbers do not come from a consultancy's sample survey. They come from the U.S. Census Bureau's Business Trends and Outlook Survey (BTOS). Every two weeks, 200,000 businesses, drawn from a rotating panel of 1.2 million, with each individual business surveyed only about once every 12 weeks. New data is published on alternating Thursdays at 10am. It is one of the most authoritative official sources on whether businesses actually use AI, cited by media, banks, and researchers everywhere. And now, the question it asks has changed twice.

Simon Willison's line chart of U.S. business AI use by firm size, marked 'Question broadened Nov 2025' and 'Series not comparable', with 250+ firms at 40.7%.

1. What the chart says: how the 40.7% is calculated

First, the numbers themselves, so that everything after this is not built on a misunderstanding. The 40.7% in Willison's chart is the October 2026 figure for businesses with 250 or more employees, defined as "used AI in any business function in the last two weeks." But it is not the raw result of any single survey. It is a six-survey trailing average that Willison calculated himself — the arithmetic mean of six consecutive biweekly survey estimates, computed separately for each question wording, never bridging the break. A trailing average smooths out volatility, but it is Willison's own construction, not the Census Bureau's methodology, and it should be read as such.

On the same basis, adoption by firm size in October 2026 was: 250+ employees 40.7%, 100–249 employees 33.5%, 50–99 employees 27.9%, 20–49 employees 23.8%, 1–4 employees 23.3%, 10–19 employees 22.1%, 5–9 employees 21.1%. The bigger the business, the more likely it is to use AI — the gradient holds at every size band. Counting all American employer businesses together, the latest reading (the two weeks ending October 4, 2026) was 24.6%, while the share expecting to use AI in the next six months was 28.6%. Both figures come from Idea Link's compilation of Census Bureau data, with the same BTOS underneath.

It is also worth noting the difference between a trailing average and a single reading. For those same 250+ businesses, the single latest reading (September 21 to October 4, 2026) was 43.6%, while the previous two-week period was 36.5% — a seven-point swing between consecutive surveys is not unusual. So 40.7% is a solid trend reading, but it describes a subgroup — large businesses — not all American businesses, and certainly not all American workers. Calling 40.7% "the U.S. business AI adoption rate" mistakes a subgroup figure for the whole population, and it is the easiest way to misread this chart.

The chart's history is worth a look too. In September 2025, Apollo Global Management drew a chart from the same Census data that seemed to show large-firm AI adoption topping out and falling back: businesses with 250+ employees peaked at around 14% in July 2025, then slipped back to about 12% by August. Plenty of people read that as "corporate AI adoption is slowing down." Willison had used that chart himself a year earlier, twice — once to demonstrate GPT-5 data analysis, once to try out Claude's new Code Interpreter. A year later, the "topping out" chart and today's "surging to 40%" chart are not the same line at all. The question changed in between.

2. The first change: November 2025, "producing goods or services" became "any business function"

The first change took effect on November 17, 2025, and the Bureau formally announced it on December 3, 2025. Before the change, the question read: "In the last two weeks, did this business use Artificial Intelligence (AI) in producing goods or services?" The examples were machine learning, natural language processing, virtual agents, voice recognition, and so on. After the change, it read: "In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions?" The examples did not change by a word.

Only half a sentence changed, but what was being measured changed completely. "Producing goods or services" asked whether AI had reached the core output of the business; "any of its business functions" asks whether AI is used anywhere in the business at all — finance using AI to reconcile accounts, marketing using AI to write copy, customer service using AI to answer email, all of it counts. Willison's own words: the answers before and after November 17 are "not directly comparable." On his chart he separated the two periods with a grey bar, labelled "Question broadened Nov 2025" and "Series not comparable," and drew a separate trailing average inside each period, with no line connecting them. That is the honest way to draw it — and the easiest part to lose when the chart gets reposted.

How big is the break? The all-business figures show it most clearly. Under the old question, adoption was 3.7% at the first reading in September 2023, and 10.0% at the last old-question reading in September 2025. Under the new question, the first reading in November 2025 jumped straight to 17.3%, reaching 24.6% by early October 2026. In other words, merely broadening the question lifted the reading from 10% to 17.3% — more than seven points — and nobody can separate from this data how much of that is businesses genuinely starting to use AI, and how much is existing use that the old question simply never captured. The climb in Willison's chart for large businesses, from 12% to 30% to 40%, crosses exactly that break: about 12% in September 2025 under the old wording, about 30% by February 2026 under the new wording, about 40% by October 2026. The segment before the break and the segment after it are two different lines. Joined together, they create the illusion of "tripling in a year."

None of this means AI adoption is not growing. Within the new wording, the rise from 17.3% in November 2025 to 24.6% in October 2026 for all businesses is a real trend, and so is the rise from 30% to 40% for large businesses. The problem is only with comparisons that cross the break. Any chart that draws a single line from 12% in September 2025 to 40% in October 2026 and titles it "AI adoption tripled in a year" is passing off a change in the measuring instrument as a change in behavior.

3. The second change: from October 19, 2026, AI became "Super Intelligence"

If the first change was a technical adjustment to a statistical definition, the second one is absurd. When Willison went looking for the wording of the latest survey, he found the word itself had changed. Starting with the survey collection of October 19, 2026, the question reads: "In the last two weeks, did this business use Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI) in any of its business functions?" The examples changed with it: "SI-powered data analysis; generative SI for marketing, product design, or coding; virtual SI agents for customer service; SI-powered forecasting or pricing; SI-driven computer vision for quality control, etc."

The rename follows Executive Order 14434. The Census Bureau's BTOS about page states that, beginning with the October 19, 2026 collection, the instrument wording was updated due to Executive Order 14434. When Willison wrote the post on November 2, the latest published data was still the October 22 release, using the old AI wording; the first data collected under the "Super Intelligence" wording would appear in the November 5 release, covering responses from October 19 to November 1. Checking the Census Bureau's BTOS data page on November 8, the page already showed "Data as of November 01, 2026," and the relevant sections were relabelled "SI (formerly AI) Use." The rename is no longer just wording in a document — it is the live data series itself.

There is an open question that the next release will settle: the businesses filling in this survey last saw the question twelve weeks ago, when it was still called AI, and now it suddenly reads "Super Intelligence (formerly referred to as AI)." The parenthetical probably makes it clear to most respondents that it means the same thing, but whether the weight of the phrase "Super Intelligence" changes anyone's answer, nobody knows yet. Willison is cautiously optimistic — he hopes the new language will not confuse respondents — but he cannot help laughing: "I can't help but laugh at the absurdity of 'SI-powered data analysis; generative SI for marketing, product design, or coding; virtual SI agents for customer service' though." An ordinary employee doing data analysis now has to answer an official survey about whether they use "Super Intelligence." The word changed; whether what is being measured changed will only become visible in the data after November 5.

It is worth adding that this rename is not an isolated aftershock, but the second cut on the same timeline. The first change broadened where AI is used; the second change replaced the word "AI" itself. Between today's data and the 3.7% reading of September 2023, there now stand not one but two breaks — one broadening of scope, one replacement of terminology. Any long-term trend reading that crosses both breaks should first ask: are you really comparing the same thing?

4. Why this is more than a joke: change the wording, and every headline moves with it

This story is easy to tell as a joke — the government renaming AI to "Super Intelligence" by executive order sounds like satire. But the practical impact is real: this dataset is the most-cited official source for business AI adoption in the world, and when its wording changes, every headline, research note, and investment memo that cites it moves too — not because business behavior changed, but because the question changed.

The same flaw is not unique to the United States. Eurostat's data has a similar break: EU business AI adoption was 20.0% in 2025, up from 13.5% in 2024 and 8.1% in 2023 — but the 2025 questionnaire added a technology type for generating pictures, video, or sound, and broadened "generating written or spoken language" to include programming code, so the years are not directly comparable either. The UK's Office for National Statistics reported 30% in September 2026 (38% for businesses with 10 or more employees), again under a different definition. Norway reported 48% for 2026, Lithuania 34.7%, each with its own breaks and classification changes. Official AI adoption statistics were never one ruler applied end to end; they are a pile of differently-defined numbers, each with its own breaks.

One more detail deserves to be singled out: the "adoption topping out" chart of a year ago and today's "surging to 40%" chart use the same survey from the same agency, yet tell two completely opposite stories. In September 2025, the old-wording reading for large businesses slipped from a peak of 14% to 12%, and plenty of commentary concluded that "corporate AI adoption is slowing." A year later, the same businesses read 40% under the new wording, and the headlines flipped to "AI is taking over business." Neither headline invented a number, and both told only half the story. In hindsight, that 12% plateau under the old wording looks more like a signal that the measuring tool had hit its ceiling — the old question only asked about "producing goods or services," which could only ever capture a limited set of uses, so the reading stalled around 12% not because businesses stopped using AI, but because the old question could not see any more uses. The moment the new question widened the scope to "any business function," the reading jumped a level. That is the real lesson of this chart: whenever the measuring instrument itself is changing, every turn in the trend line should first be asked whether the behavior changed, or the ruler did.

A bigger confusion comes from the difference between counting firms and counting workers. The 24.6% figure for all businesses is an average by firm count — the vast majority of American businesses are small, small businesses adopt at lower rates, and so the overall figure is pulled down. Weighted by employment, the Atlanta Fed's survey found that 78% of the U.S. labor force works at firms that have adopted AI; the Census Bureau's own AI supplement was 32% employment-weighted against 18% firm-weighted. 78%, 32%, 24.6%, 18% — all four numbers are correct, and they measure four different things. Lifting any one of them out on its own under the headline "U.S. business AI adoption rate" is presenting a part as the whole.

So the real problem is not that the numbers are fabricated, but that the definition matters more than the number. That is also where Willison's post earns its value: he did not declare whether AI adoption is high or low, he took a widely-cited trend line apart and showed that it breaks twice in the middle — once when the question was broadened, once when it was renamed. The easiest mistake a reader of official statistics can make is to look at the number and not the questionnaire; and the questionnaire is the definition of the number. Change the definition, and the meaning of the number changes, while the headlines usually keep writing to the old meaning.

5. For people who build with AI, how should these numbers be read

For VibeFix readers — independent developers and small teams who build and ship their own products — there are three rules for reading these numbers. First, look at change within a single wording: under the new wording, all businesses rose from 17.3% to 24.6%, and large businesses from 30% to 40%. That climb is real, and it says AI is moving from "someone is trying it" to "most large businesses use it." Second, do not compute multiples across wordings: 12% to 40% is not "tripling in a year," because a broadened question sits in between; using it to argue that "AI is already everywhere" rests on evidence that does not stand up. Third, see the gap between large and small businesses for what it is: 40.7% for firms with 250+ employees, only 21.1% for firms with 5–9. Small-business adoption is about half the large-business rate. For a small team, that is both a gap and an opening — the places where large businesses are already using AI at scale are exactly where a small team can use AI tools to close the gap, and that is what the vibe-coded projects on VibeFix are doing.

There is also a reminder here for anyone who cites data: the next time you see a headline like "X% of American businesses already use AI," ask three questions first — which survey, what exactly was asked, and weighted by what? If you cannot answer them, the number is not worth citing. Willison's approach is worth copying: he published the raw data file, a snapshot of the questionnaire, and his own method, and the chart's footnotes state "arithmetic mean of six successive survey estimates, calculated separately for each question wording; no bridging of the break" — even flagging the gap left by the 2025 government shutdown. Honesty in data journalism is not about how good the numbers look, but about showing how the numbers were made.

Finally, back to the name "Super Intelligence" itself. It is absurd, but it is also accidentally honest: when official documents start calling everyday data analysis, marketing copy, and customer-service replies "Super Intelligence," the term has been diluted until it means no more than "AI." The question genuinely worth asking is what the data after November 5 will show: once the question is called "Super Intelligence," will the adoption reading jump again? And if it does, is that because businesses used more of it, or because respondents were startled by the new word? The next release of this dataset is the real result of the renaming experiment. Until then, every grand narrative about the "AI adoption rate" should first look down at what the questionnaire actually asks.

One sentence to sum it up: the most-cited trend line for American business AI adoption had its question changed twice and its name changed once within a year — the 40.7% figure for large businesses is real, but it was not measured with the same ruler as the 12% of a year ago. The numbers did not lie; the people who joined two segments into one line did.

Views 5Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

An analyst seen from behind in a dark security operations center, facing large screens showing world maps and network topology
News
The Patch Just Shipped and the Exploit Is Already There: Google Finds the First Fully Autonomous AI Exploitation Framework

On 16 November 2026, Google's Threat Intelligence Group reported the first fully automated vulnerability discovery and exploitation chain: no human in the loop. The evidence is an exposed server running exploit_agent, a framework that diffs code around patch announcements, writes exploits and fires them. The same report shows a North Korean cluster turning Claude Code's settings file into a backdoor trigger — and stolen cloud compute running up $100,000 bills in a day.

Industry TrendsAI CodingSecurity & Privacy
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates