Facts About AI in 2026: The Numbers, and What They Leave Out

Modern AI is dominated by systems that learn patterns from data rather than following only hand-written rules — though older approaches like rule-based expert systems and search algorithms still count too. Facts about it go stale fast: on SWE-bench Verified, a coding benchmark built around real GitHub issues, scores climbed from 60% to nearly 100% of the human baseline in a single year, per Stanford HAI’s 2026 AI Index.

How Fast Is AI Adoption Actually Growing?

Widespread, but not the same thing as scaled. Those are two different numbers, and conflating them is where most coverage goes wrong.

McKinsey’s 2025 global survey found 88% of respondents’ organizations used AI in at least one business function, up from 78% the year before — a figure also reported in Stanford’s 2026 Index.

Scaling tells a different story. Nearly two-thirds of organizations haven’t begun scaling AI across the enterprise, with roughly one-third reporting they’ve started.

Consumer adoption moved fast too. Generative AI reached 53% of the population faster than the PC or the internet, per Stanford’s Index — though pace varies by country and correlates strongly with GDP per capita.

Are AI Agents Actually Being Used, or Just Talked About?

Mostly still early. 62% of organizations report at least experimenting with AI agents, per McKinsey’s survey, but only 23% say they’re scaling an agentic system somewhere in the enterprise.

Even “scaling” undersells how contained this still is. Most companies scaling agents are doing so in just one or two business functions, and in any single function, no more than 10% of organizations report scaling agent use there.

How Much Has AI Inference Actually Cost Fallen?

Sharply, on the specific metric Stanford tracks, over a specific window. The cost of matching GPT-3.5-level performance fell from $20 per million tokens in November 2022 to $0.07 per million tokens in October 2024 — a 280-fold drop, per Stanford’s AI Index. That’s a 2022–2024 figure, not a 2026 price point.

Hardware and efficiency moved in the same direction over that period, as separate metrics: AI hardware costs dropped roughly 30% a year, and energy efficiency improved about 40% a year.

Investment climbed at the same time cost fell, without one figure explaining the other. Corporate AI investment reached $252.3 billion in 2024, with private investment climbing 44.5% — cheaper inference and heavier funding ran in parallel, not in a straight cause-and-effect line.

What Does One AI Query Actually Consume?

Depends on which layer gets measured, and most viral claims only measure one.

On-site cooling water is the narrowest number, and the smallest. Gemini’s text query uses about 0.26 milliliters on-site; OpenAI has cited a similar figure near 0.3 milliliters.

Grid electricity adds a second, larger layer most headlines skip. Generating that electricity often consumes water too, so the full footprint runs well above the on-site figure alone.

Task type changes the energy math substantially. A typical chat query uses roughly 200 times the energy of a basic classification task, and generating an image runs roughly 1,450 times more energy-intensive than a basic text query — with video generation higher still.

Inference, not training, drives most of that demand. Day-to-day use accounts for an estimated 80–90% of total AI energy consumption, not the initial training run — a pattern the UN University’s 2026 report attributes to sheer volume: ChatGPT alone processes an estimated 2.5 billion prompts a day.

Infrastructure totals are the widest lens, and the one usually quoted out of context. Global data centers used an estimated 448 terawatt-hours of electricity in 2025, projected to reach 945 TWh by 2030, with a water footprint projected at 9.3 trillion liters — a figure covering data centers broadly, not AI workloads alone.

Disclosure hasn’t kept pace with any of this. A 2021 study in npj Clean Water found fewer than a third of data center operators measured their water consumption at all — an old figure, but no more current industry-wide number has replaced it since.

Are People Becoming More Worried About AI?

Yes, and the trend has accelerated. 52% of Americans say they’re more concerned than excited about AI’s growing role in daily life, up from 37% in 2021, and 9% are more excited than concerned, per Pew’s survey fielded in late June 2026.

Young adults moved the most. A majority of adults under 30 — 55% — now say they’re more concerned than excited, the first time that’s happened since Pew began asking in 2021.

Experts see something else. Pew’s parallel survey of AI experts found 47% more excited than concerned versus 11% of the general public, and Stanford’s 2026 Index puts the expert-public gap on whether AI will help people do their jobs at 50 percentage points.

What Is AI Doing to Jobs So Far?

Nothing broad. Something specific and widening at the entry level.

Stanford’s Digital Economy Lab, using payroll data through mid-2026, finds no widespread, economy-wide job displacement associated with AI. Experienced workers show no comparable employment gap.

Young workers are a different story. Employment among workers ages 22–25 in highly AI-exposed occupations now stands about 19% below where it would be if it had kept pace with less-exposed peers — up from a 15% gap measured in July 2025.

The mechanism matters as much as the number. The adjustment operates mainly through reduced hiring of young workers, not increased separations — companies aren’t cutting junior staff loose so much as not posting junior roles in the first place. Stanford’s own team calls this a descriptive pattern, not proof of causation, and the entry-level hiring paradox is already showing up in wider labor-market data independent of this one study.

Why Are AI Facts So Hard to Compare?

Definitions and denominators shift under every headline number, which is the real reason “facts about AI” pages disagree with each other.

Benchmarks keep getting saturated faster than researchers can redesign them — Stanford’s own framing. Documented AI incidents rose from 233 to 362 year over year, per the 2026 Index’s Responsible AI chapter, a signal that safety and policy tracking is struggling to keep pace with deployment.

Even AI labs have started grading themselves, with uneven transparency. Anthropic recently published its own R&D automation metrics, scored by its own models against its own rubric — a real data point, but also a case study in why self-reported numbers need a second source before they become “facts.”

Expert disagreement compounds the confusion at the top of the industry too. Anthropic executives, independent researchers, and rival labs openly disagree over how dangerous frontier AI actually is, leaving even risk estimates contested rather than settled.

Is the Industry Getting More Transparent — or Less?

Less. The average transparency score across major AI companies fell from 58 to 40 out of 100 between 2024 and 2025, after several years of gradual improvement.

The drop wasn’t uniform. Meta’s score dropped from 60 to 31 and Mistral’s from 55 to 18, while IBM stood out at 95, against the lowest scorers, xAI and Midjourney, at 14. Companies are most opaque about training data and training compute, and about what happens to a model after deployment.

Capability convergence is moving in the opposite direction. U.S. and Chinese models have traded the performance lead multiple times since early 2025, closing a gap that looked much wider just two years ago.

The companies with the most competitive frontier models are disclosing less about how they’re built, not more.

AI Facts That Look Contradictory — But Aren’t

88% adoption, yet only a third are scaling. Different stages of the same rollout — trying a tool once counts as adoption; running it enterprise-wide doesn’t happen at the same rate.

AI use is rising while public concern is rising, too. Behavior and sentiment aren’t the same measure — more people using something doesn’t mean more people trust it.

Inference is getting cheaper while AI spending explodes. Unit cost and total demand move independently; cheaper per-query pricing can coexist with far more total spending if usage grows faster than price falls.

No economy-wide job collapse, yet a 19% employment gap for young workers. Aggregate labor data and one narrow, heavily-exposed age group inside it are different denominators entirely.

0.26 milliliters per query, yet trillions of liters industry-wide. One is a single on-site measurement; the other is the sum of global infrastructure over a year. Neither number is wrong — they’re not measuring the same thing.

What the Data Says Is Changing Fastest

Four things move on a documented curve rather than a guess: adoption climbed from 78% to 88% of organizations in one year; inference cost for a fixed capability level fell over 280-fold between 2022 and 2024; global data-center electricity use reached an estimated 448 TWh in 2025; and the entry-level employment gap for AI-exposed 22- to 25-year-olds widened from 15% to 19% over the same stretch.

Public sentiment moved in the opposite direction from adoption. Concern has climbed every year since 2021 and shows no sign of reversing on its own.

FAQs

Q. What percentage of businesses use AI in 2026?

About 88% of organizations report using AI in at least one business function, according to McKinsey’s 2025 global survey cited in Stanford’s 2026 AI Index. However, only about one-third have begun scaling AI across the enterprise, so adoption is much higher than organization-wide deployment.

Q. Is AI taking jobs in 2026?

AI has not caused broad economy-wide job displacement, according to Stanford labor research through mid-2026. However, employment among workers ages 22–25 in highly AI-exposed occupations is about 19% below the level expected if it had kept pace with less-exposed occupations. The research indicates that reduced hiring, rather than increased layoffs, accounts for much of this gap.

Q. How much water does one AI query use?

A short AI text query can use less than half a milliliter of on-site cooling water, based on figures reported for major AI systems. This does not include water consumed indirectly to generate the electricity powering data centers, so an AI query’s total water footprint can be higher.

Q. Are people becoming more concerned about AI?

Yes. 52% of U.S. adults said they were more concerned than excited about AI in Pew Research Center polling conducted in June 2026, up from 37% in 2021. Only 9% said they were more excited than concerned.

Q. How much has the cost of AI fallen?

The cost of achieving roughly GPT-3.5-level AI performance fell about 280-fold, from $20 per million tokens in November 2022 to $0.07 in October 2024, according to Stanford’s AI Index. This measures the historical decline in inference cost for a fixed capability level; it is not a current 2026 API price.

Q. Is AI adoption growing faster than enterprise deployment?

Yes. 88% of organizations report using AI in at least one business function, while only about one-third have begun scaling it across the enterprise. The gap shows that experimenting with or adopting AI is far more common than deploying it broadly across an organization.

Q. Is the AI industry becoming more transparent?

No. The average transparency score for major AI companies fell from 58 out of 100 in 2024 to 40 in 2025, according to Stanford’s Foundation Model Transparency Index. Major gaps remain around training data, computing resources, and what happens to AI models after deployment.

Related: AI Transformation Is a Governance Problem (Not Tech) — 2026 Truth