The 2026 AI Report

The 2026 AI Report · Independent · Updated 10 October 2026

The year AI
became ordinary.

Three young women dance hand in hand in the manner of Botticelli’s Three Graces: one in augmented-reality glasses looking up, one frowning at her phone, one paying with hers.
  1. Brilliantsees more than ever
  2. Unreliablestill gets simple things wrong
  3. Costlyand the bill keeps growing
Scroll
Widely used

53%

of people use generative AI, just three years after launch

Stanford HAI, AI Index 2026 · 13 Apr 2026

Chapter 1 · Widely used vs. worried

Everyday, and uneasy

In three years, half the world started using AI. Half of Americans are more worried than excited about it.

Impressive

35 of 42

points, a gold medal at the world’s top maths competition for students

Google DeepMind · 21 Jul 2025 †

At the International Mathematical Olympiad, under the same time limit and grading as human contestants.

Chapter 2 · Impressive vs. unreliable

Smart, but makes simple mistakes

AI solved five of six problems at the hardest student maths contest. Yet it still misreads a simple clock about half the time.

Helpful up to

83%

less time spent on paperwork

Stanford HAI, 12 Takeaways · 13 Apr 2026

For doctors who use AI to write up patient notes.

Chapter 3 · Helpful vs. unproven

AI in healthcare

AI already saves doctors time on paperwork. But very few studies have tested medical AI on real patient records.

Cheaper about

about 13 times

cheaper every year for the same quality

State of AI Report 2026, citing Epoch AI · 8 Oct 2026 †

Getting the same result from AI costs about half as much every three months.

Chapter 4 · Cheaper vs. costlier

The real cost

Using AI gets cheaper every year. But the costs show up elsewhere: electricity, water, and data centres in people’s neighbourhoods.

Experts

73%

of AI experts think AI will make jobs better

Stanford HAI, AI Index 2026 · 13 Apr 2026

When asked if AI will improve how people do their work.

Chapter 5 · Experts vs. the public

Experts and the public disagree

People who build and study AI are far more optimistic about its effect on jobs than the people who work alongside it. The gap is 50 percentage points.

Summary

Treat AI like any other tool: know what it’s good at, where it fails and what it costs.

AI isn’t finished, it isn’t something to fear, and it isn’t free. This year it won a maths gold medal but still misread clocks. It cut doctors’ paperwork, though the evidence behind medical AI is still thin. It got cheaper to use, while its data centres needed more power and water and faced local opposition. When left to use a computer on its own, it still fails about one task in three (Stanford HAI, AI Index 2026 · 13 Apr 2026). Half of Americans are more worried than excited, and experts are far more hopeful than the public. Both views have a point. Use these numbers, and their sources, the next time AI comes up.

Want to share a number? Press Copy next to it to copy the number with its source.

Sources

Where each number comes from and when it was published. All checked on 10 October 2026. Three marked † were confirmed through trusted news reports because the original page couldn’t be accessed.

  1. 53% of people, three years in. Stanford HAI, AI Index 2026, 13 April 2026.Open the sourceBack to the number
  2. 50% of Americans more concerned than excited. Pew Research Center, How Americans View AI and Its Impact on People and Society, 17 September 2025.Open the sourceBack to the number
  3. 35 of 42 points, Olympiad gold. Google DeepMind, gold-medal standard at the International Mathematical Olympiad, 21 July 2025. †Open the sourceBack to the number
  4. 50.1% of clock faces read right. Stanford HAI, AI Index 2026, 13 April 2026.Open the sourceBack to the number
  5. Up to 83% less time writing notes. Stanford HAI, Inside the AI Index: 12 Takeaways from the 2026 Report, 13 April 2026.Open the sourceBack to the number
  6. 5% of clinical AI studies used real patients’ data. Stanford HAI, 12 Takeaways, 13 April 2026.Open the sourceBack to the number
  7. About 13× cheaper each year (about 47% per quarter) for the same score. Air Street Capital, State of AI Report 2026, citing Epoch AI, 8 October 2026. †Open the sourceBack to the number
  8. 29.6 GW of AI data-centre power capacity. Stanford HAI, 12 Takeaways, 13 April 2026.Open the sourceBack to the number
  9. Yearly inference water may exceed the drinking-water needs of 1.2 million people. Stanford HAI, 12 Takeaways, 13 April 2026 (GPT-4o inference).Open the sourceBack to the number
  10. 71% of Americans oppose a local AI data centre (53% a nuclear plant). Air Street Capital, State of AI Report 2026, citing Gallup (March 2026), 8 October 2026. †Open the sourceBack to the number
  11. 73% of experts vs 23% of the public expect AI to improve how people do their jobs. Stanford HAI, AI Index 2026, 13 April 2026.Open the sourceBack to the number
  12. AI completes about two in three real computer tasks (up from 12%), so it fails about one in three. Stanford HAI, AI Index 2026, 13 April 2026 (OSWorld benchmark).Open the sourceBack to the verdict

† Confirmed through trusted news reports; the original page couldn’t be accessed.

About the artwork

How were all these images made in one style?

Every image on this page was made with AI using one saved style, “Renaissance Mythic Classicism” on StyleRef. Using the same style for each image keeps the colours, texture and drawing consistent, so the pictures read as one series.