In this article
Today we're talking about the report the press cites without reading and the figures that circulate with their footnotes cut. Stanford HAI's AI Index 2026 is one of the few serious x-rays of the sector, almost six hundred pages, and almost no one has read it whole. Why does the context lost in trimming matter? Because the 53% figure that circulates in headlines and the 5.5% figure that's in McKinsey describe the same sector and say opposite things. I admit that for months I cited the "53% adoption in three years" myself without knowing it was a population figure, not a company one. A colleague corrected me; it still makes me cringe. Build your informed opinion.
I've spent two summers leafing through the AI Indexes. The 2024 one was already a brick and the 2026 one has beaten it in weight. There are serious figures underneath and there are figures the report itself qualifies with footnotes the press doesn't transcribe. This article deals with six of those figures: the ones most seen in headlines and the ones worst conveyed.
The 53% that circulates in headlines
The 53% figure is the most repeated in the AI Index 2026, and it's almost always cited wrong. The report's figure says generative artificial intelligence has reached 53% global population adoption in three years since its leap into mass consumption, a rate higher than that of the PC and the internet on their respective adoption curves. It's a figure of people, not of companies. And within that figure there are enormous variations: Singapore around 61%, the United Arab Emirates around 54%, the United States around 28.3% (24th in the world ranking), Spain at intermediate figures that depend on the specific survey.
When a headline in the Spanish press says "53% of companies use generative AI," it's confusing the figure. The company figure is different and tends to be higher. The AI Index 2026 documents that 88% of organisations use AI in at least one function, and that the subset that has specifically deployed generative AI in at least one process is around 70%. That means that in three years, between 2023 and 2026, the percentage of organisations with generative AI in production has gone from 33% to roughly 70%.
That last point is news and it's rarely cited that way. The corporate-adoption curve is the fastest the sector has seen. What's worth doing isn't repeating the figure, but understanding what it hides. Because the next figure is the uncomfortable one.
What's missing underneath, the 5.5%
McKinsey published in November 2025 The State of AI in 2025: Agents, innovation, and transformation, a survey with data comparable to the AI Index's but with a specific focus on financial return. The two figures hardest to digest in the report are these. Only 5.5% of organisations report seeing real financial returns attributable to their AI investments. And more than 80% state that their use of generative AI still produces no tangible impact on the company's aggregate operating result.
This doesn't contradict the AI Index's figures. It complements them. Having AI in production and obtaining measurable return aren't the same. There's 70% of companies with generative AI deployed and 5.5% with clear return. The difference between the two figures is the real dimension of the problem and the space the sector will have to cover in the coming years to justify the capital invested.
McKinsey also documents positive indicators at the use-case scale. The software-engineering and IT function reports cost reductions of 10% to 20%, and the marketing and product-development functions report revenue increases above 10%. But that impact still doesn't add up to the corporate EBIT except in 5.5% of cases. The reason is operational: most organisations are still in the pilot phase, not scaling.
The distance between the AI Index headline (88% adoption) and the McKinsey headline (5.5% return) is the one that defines the sector's reality. The Spanish press usually cites the first. It almost never cites the second. And the second is the one that decides whether the current investment was sensible or not.
The market projections, told as certainty
"The global AI market will reach 1.5 trillion dollars in 2030." This phrase, or variants, appears in coverage of the report almost every month. It's worth looking at the source.
The AI Index gathers projections from several independent analysts —Bain & Company, IDC, McKinsey Global Institute, Goldman Sachs, PwC— and presents them as a range, not a single datum. The figures vary between 600 billion and 4.4 trillion dollars for 2030, depending on hypotheses about adoption rate, ROI, substitution of existing markets and demand elasticity. The 1.5-trillion figure is the approximate geometric centre of the range, not the consensus conclusion.
What the report explains in its methodological sections —and what media coverage practically never conveys— is that five-year projections in a general-purpose technology tend to err by wide margins. The projections made in 2015 about the AI market in 2020 fell short by roughly an order of magnitude. The projections made in 2020 about 2025 erred high in some cases and low in others. There are reasons to think the sector has learned. There are reasons to think the hype curve repeats cyclical patterns. The honest acknowledgement of uncertainty is part of the report, and the coverage erases it systematically.
When a projection is cited, the range should be cited. When the range is cited, what the extremes are based on should be explained. And when neither the range nor the extremes are cited, what's conveyed isn't information about the sector, it's decoration to prop up the headline.
The confidence-vs-use paradox
The report includes figures of citizen confidence in AI by country, cross-referenced with figures of use. The cross-reference produces something many analysts describe as a paradox and that isn't one. In countries with greater AI use in daily life —China, India, the United States, Singapore— citizen confidence in the technology's development is relatively high. In countries with less intensive use —some European ones, Japan— confidence is lower.
The intuitive reading, the one the press repeats, is that "people distrust what they don't know." But the data can be read the other way and the inverse reading is plausible. A legitimate hypothesis is that in countries where use is more intensive, users have internalised the concrete problems —frequent errors, detectable biases, operational dependence— and adjust their expectations downward. The "high" confidence may be confidence in limited functioning, not in the total promise. Whereas in countries of lower use, confidence may be mediated by media coverage, which tends to present AI in aspirational or catastrophist terms, without the correction daily contact gives.
I have no firm evidence to settle which of the two readings is correct. The AI Index doesn't have it either. But media coverage has settled it: it has chosen the "distrust through ignorance" reading and repeats it. That choice isn't neutral. It reinforces the narrative that any resistance is ignorance, which suits the actors who need accelerated adoption.
What the AI Index is good for and what it isn't
I was going to write that the AI Index is the best report in the sector. But I'll correct myself, because "the best" depends on what for. It's worth being precise.
The AI Index is excellent as an aggregator of figures from diverse sources with explicit methodology. When you want to know how many frontier models were published in 2025, where the highest-producing authors are, what proportion of industrial compute is concentrated in the United States and China, how many patents were granted in the sector last year, the AI Index is the most solid and best-cited reference. The figures come with source and methodological note.
The AI Index is mediocre as a unified narrative. The report's structure groups chapters —research, technical, economics, education, policy, public opinion, ethics, science, hardware. That works for a sectoral reading, but produces the optical effect that each figure is comparable with any other. It isn't always. The citizen-adoption figures come from Ipsos surveys. The corporate-adoption figures come from McKinsey and additional consultancies. The investment figures come from PitchBook. The methodologies differ and the nuances are lost in a quick reading.
The AI Index is terrible as an investment guide and as an operational prediction. That isn't its aim. Any use of the report to say "this will happen in 2027" is improper use. The report describes what happened in 2024-2025, with its best estimate, and leaves the projection to others.
To read it honestly, it's worth keeping three operational rules in mind. First rule: any figure in the report carries a footnote. Find it before citing it. Second rule: the adoption figures refer to presence, not impact. The impact figure is in other reports (McKinsey, BCG, Bain) and tends to be an order of magnitude smaller. Third rule: projections beyond three years are scenarios, not predictions. Citing them as predictions is technical dishonesty.
The political question
This is personal opinion but I back it with the report's own trajectory. The AI Index works as an instrument of informal governance. It's read by regulators, legislators, corporate directors, journalists, consultants. Its figures carry over, with or without context, to presentations in Davos, in the Congress of Deputies, in boards of directors. What the report highlights or doesn't, what it qualifies or doesn't, decides in good part what's discussed in public policy and what isn't.
That's why it's worth the non-specialist reader knowing two things. The first is that Stanford HAI —the institute that publishes the report— is a serious academic institution, but funded in part by the same companies whose sector is measured in the report. There's no conflict of interest in the legal sense, but there is a risk of framing. The report presents accelerated adoption as mostly neutral good news, when there are sectors where accelerated adoption without measurable return is a serious problem. The second is that Spanish-language media coverage of the report is reduced, in most general-interest outlets, to a translation of Stanford's press release. That release highlights the showiest figures. The sections on the confidence-use paradox, limited ROI, and projections with uncertainty don't enter the release and don't enter the coverage.
I'd ask, with no realistic hope of getting it, two things of the Spanish press. One: that when the AI Index is cited, the page and chapter be cited, not the executive summary. Any figure in the report can be checked; it's editorial laziness that produces the trimmings. Two: that when an AI Index adoption figure is cited, it be complemented with a return figure from McKinsey or BCG. The double citation doesn't triple the work and gives the reader a complete picture back.
The latest concrete datum from the AI Index 2026 itself, published by Stanford HAI on 8 April 2025, sums up the problem with two figures together: 88% of organisations with AI in some function, against the 5.5% that reports clear financial return according to McKinsey State of AI in 2025 (November 2025). That difference is the whole sector. Whoever understands it understands the coverage.
Definitions
Adoption: the documented presence of the use of a technology in a population or in a set of organisations. It doesn't measure intensity, profitability or continuity. A company that tests a pilot counts as adoption; a company that has closed the pilot without scaling can also count depending on the survey.
EBIT: Earnings Before Interest and Taxes. A standard financial indicator to measure a company's operating result before financial and tax charges. When McKinsey says more than 80% of companies see no EBIT impact from AI, it's saying the effect, if it exists, isn't felt in the consolidated income statement.
Market projection: a quantitative estimate of the future size of a market under explicit assumptions. It isn't a prediction. Projections beyond three years in general-purpose technologies have typical errors on the order of 50% or more.
Citizen confidence: the percentage of respondents who declare a positive view of a technology or of the institutions that develop it. It's an indicator of perception, not of behaviour. It's measured with methodologies that vary between studies.
References
Stanford HAI, Artificial Intelligence Index Report 2026 (8 April 2025, hai.stanford.edu/ai-index/2026-ai-index-report). The full report and all chapters.
McKinsey & Company, The State of AI in 2025: Agents, innovation, and transformation (November 2025, mckinsey.com/quantumblack). Source of the 5.5% figure of organisations with clear financial return.
Bain & Company, Technology Report 2025: AI Investment and Returns (2025). Independent projections and qualification of the range.
International Data Corporation (IDC), Worldwide AI Spending Guide (semi-annual update). AI market projections to 2030 from an IT-investment perspective.
Goldman Sachs, The Economic Impacts of Generative AI (sectoral report 2024-2025). An alternative source for macro projections.
Ipsos, AI Monitor Global Survey (annual, 2023-2025). The survey used by the AI Index for citizen-confidence figures.
PitchBook, AI investment data cited by the AI Index. The original source of the deployed venture-capital figures.
To go deeper
Erik Brynjolfsson & Andrew McAfee, The Second Machine Age (W. W. Norton, 2014). A historical framework for understanding why productivity takes time to aggregate even when the technology is in use.
Daron Acemoglu & Simon Johnson, Power and Progress (PublicAffairs, 2023). An analysis of why the accelerated adoption of a technology doesn't guarantee equitable distribution of its return.
Robert J. Gordon, The Rise and Fall of American Growth (Princeton University Press, 2016). An antecedent on how the impact of major technologies on macro statistics is measured and unmeasured.
Tim Harford, The Truth-Detective's Handbook (Bridge Street Press, 2020, revised edition 2024). A manual for the critical reading of public figures, useful for facing any sectoral report.
You might also like
- What the AI Index Report 2026 says that the Spanish press doesn't tell you
- Why Stanford HAI rules and Spanish universities count for nothing
- How the public conversation about AI broke
- The real cost of training GPT-5
Comments0
No comments yet.
Leave a comment