01 / 14
// C1 Lesson — Data Journalism & Tech Literacy

The State of AI,
in Charts.

SOURCE: MIT Tech Review STANFORD AI INDEX 2026 EST. TIME: 90 MIN LEVEL: C1

// About this lesson

You'll work through an MIT Technology Review article that translates Stanford's 2026 AI Index into seven headline charts. The lesson develops your ability to read statistical prose, discuss structural trends, and argue with data — core C1 skills for journalism, business, and academic discourse.

// By the end, you will

  • Describe quantitative trends with precision (surge, plateau, close the gap)
  • Use hedging language for uncertainty
  • Interpret charts rather than merely describe them
  • Discuss AI geopolitics and adoption gaps
§ 01 — Warm-up

Before you read

Tap the statements that describe your current relationship with AI tools.
I use AI tools daily for work
I'm paying for a premium AI subscription
I'm broadly optimistic about AI
I'm broadly sceptical about AI
I follow the US–China AI race
I worry about AI's energy footprint
I think AI is in a bubble
I've heard of the Stanford AI Index
I can name three frontier AI labs
I've never used a generative AI model
> Compare selections with your teacher. Which positions feel most contested? Which feel mainstream?
§ 02 — Vocabulary

Match the term to its function

Click a term on the left, then its definition on the right. Terms recur throughout the readings.
// Term
benchmark
frontier model
adoption curve
the jagged frontier
consumer surplus
compute
inference
to close the gap
// Definition
The uneven pattern where models excel at some tasks and fail at trivially simpler ones.
A standardised test used to compare model performance on specific tasks.
To reduce the performance or capability difference between two competitors.
Computational resources (GPUs, data centre capacity) required to train or run a model.
A state-of-the-art AI system operating at the edge of current capability.
The stage when a trained model generates outputs in response to user queries.
The economic value users receive beyond what they pay for a product.
The trajectory showing how quickly a technology spreads through a population.
> Matches complete: 0 / 8
§ 03 — Reading 1

Whiplash

The opening framing — why AI is so hard to read.
[PASSAGE 01 — FRAMING]

If you're following AI news, you're probably getting whiplash.

AI is a gold rush. AI is a bubble. AI is taking your job. AI can't even read a clock. The 2026 AI Index from Stanford University's Institute for Human-Centered Artificial Intelligence — the field's annual report card — cuts through some of that noise. Despite predictions that AI development may hit a wall, the report says the top models just keep getting better.

People are adopting AI faster than they picked up the personal computer or the internet. AI companies are generating revenue faster than companies in any previous technology boom, but they're also spending hundreds of billions of dollars on data centres and chips. The benchmarks designed to measure AI, the policies meant to govern it, and the job market itself are all struggling to keep up.

The result is a landscape full of contradictions. Google DeepMind's top reasoning model scored a gold medal in the International Math Olympiad — yet fails to read an analogue clock roughly half the time. This pattern has a name: the jagged frontier. Models are extraordinary at some things and embarrassingly weak at others, and this unevenness is precisely why opinion on AI is so divided.

§ 04 — Reference catalogue

Eight headline numbers

The metrics that anchor the rest of the lesson. Read once; return as needed.
Metric
Finding
Status
53%
Population adoption of generative AI Reached in three years — faster than the PC or the internet at equivalent points.
surging
2.7%
The US lead on model performance As of March 2026, Anthropic's top model leads the best Chinese model by a razor-thin margin on Arena rankings.
closing
$581.7B
Global corporate AI investment in 2025 More than doubled year-on-year (+129.9%). Private investment alone reached $344.7 billion.
surging
88%
Organisational AI adoption Of surveyed companies globally report using AI in at least one business function.
mainstream
5,427
AI data centres in the US More than ten times more than any other country — but the entire supply chain depends on TSMC in Taiwan.
concentrated
–89%
AI researchers moving to the US Net migration of AI talent into the US has collapsed since 2017, with 80% of that decline in the last year alone.
declining
29.6 GW
AI data centre power capacity Roughly what it takes to power the state of New York at peak demand. Grok 4's training alone emitted ~72,800 tonnes CO₂e.
environmental
+50pp
Expert–public perception gap on jobs 73% of US AI experts view AI's job impact positively, vs just 23% of the general public.
divergence
// Source: Stanford AI Index 2026; Arena Leaderboard; Epoch AI; AIID
§ 05 — Chart 1 of 2

Adoption, historically compared

How fast generative AI has spread vs prior general-purpose technologies.
Years to reach 50% of population
// US data; lower bars = faster diffusion
0 5 10 15 20 25 yrs Personal computer 23 yrs Internet 15 yrs Smartphone 9 yrs Generative AI 3 yrs Years from mass-market release to 50% US population adoption
// Source: Stanford AI Index 2026, based on comparable US population adoption data.

How to read it. The bars measure time-to-saturation, not total users. Generative AI didn't just grow — it compressed the adoption curve into a timeframe unprecedented in consumer technology. The caveat: population-level adoption varies sharply by country. The UAE and Singapore top 60%, while the US ranks only 24th globally at 28.3%, despite leading in investment and model development. Building AI and using AI are no longer the same story.

§ 06 — Reading 2

A neck-and-neck race

The geopolitical picture.
[PASSAGE 02 — US vs CHINA]

The gap that effectively closed.

In a long, heated race with immense geopolitical stakes, the US and China are almost neck and neck on AI model performance. In early 2023, OpenAI had a comfortable lead with ChatGPT; by 2024, Google and Anthropic had narrowed it. In February 2025, DeepSeek's R1 briefly matched the top US model. As of March 2026, Anthropic leads, trailed closely by xAI, Google, and OpenAI — while Chinese models like DeepSeek and Alibaba lag only modestly.

With the best models separated by razor-thin margins, competition has moved on. Raw capability matters less than cost, reliability, and real-world usefulness. The index notes that the two countries have different structural advantages: the US still leads in private investment and top-tier model releases, while China leads in publication volume, citation count, patent output, and industrial robot installations.

Meanwhile, a third picture is emerging that complicates the binary framing entirely: the number of AI researchers moving to the US has collapsed by 89% since 2017, with 80% of that decline occurring in just the past year. If talent is destiny, the map is being redrawn in slow motion.

§ 07 — Chart 2 of 2

Model performance, over time

The US–China rankings on the Arena Leaderboard.
Top model score by country — Arena Leaderboard (2023–2026)
// Higher = stronger; lines converge in 2025
1450 1380 1310 1240 1170 Q1'23 Q1'24 Q1'25 Q4'25 Mar'26 Feb '25 — DeepSeek-R1 matches top US model Top US Top CN (−2.7%) Time Arena score (style-controlled) Top US model Top Chinese model
// Source: Arena (formerly LMArena), reproduced via Stanford AI Index 2026. Scores are illustrative of reported trajectory.

Key language. When trend lines get closer, analysts reach for a specific vocabulary: to narrow, to close the gap, to catch up, to converge, to overtake, neck-and-neck, razor-thin margin. Notice how the article uses multiple synonyms in a single paragraph — a hallmark of C1 register. Avoid repetition; vary the metaphor.

§ 08 — Collocations

Language for describing trends

Ten collocations pulled directly from the source, grouped by function.
Function Collocation Example
Sharp growth to skyrocket / to surge Private AI investment surged by 127%.
Sharp growth to more than double Global corporate AI investment more than doubled in 2025.
Convergence to close the gap / to narrow the gap Chinese models have closed the gap with US frontier models.
Convergence to be neck and neck The two economies are neck and neck on AI performance.
Stalling to hit a wall / to plateau Early forecasts that AI would hit a wall did not survive the data.
Stalling to struggle to keep up Regulation is struggling to keep up with capability.
Decline to collapse / to drop sharply Talent migration into the US has collapsed by 89%.
Magnitude by a razor-thin margin The leader is ahead by a razor-thin margin of 2.7%.
Hedging may exceed / is estimated at Inference water use may exceed the needs of 1.2 million people.
Interpretation the data reveals / the data suggests The data suggests a decoupling of AI investment from AI adoption.
§ 09 — Gap-fill

Complete the analyst's summary

Use each word once. Type exactly as shown in the bank.
// WORD BANK
skyrocketed
plateau
closed
outpacing
razor-thin
collapsed
exceed
surplus
whiplash
jagged
01. Following AI news is a study in contradiction — the 2026 Index helpfully cuts through the .
02. Capability keeps rising: early fears that AI would have been contradicted by the benchmark data.
03. The US–China performance gap has effectively , with top models separated by a margin.
04. Global corporate investment in AI past $580 billion in 2025.
05. Generative AI is spreading through populations faster than any prior technology, both the PC and the internet.
06. US consumer from generative AI is estimated to $172 billion annually.
07. Meanwhile, inbound AI talent migration to the US has by 89% since 2017.
08. Capability remains uneven — the so-called frontier, where models win Math Olympiads yet misread clocks.
§ 10 — Comprehension

Multiple choice

One answer per question. Explanations appear on selection.
Q1What is the main argument the MIT Technology Review article makes about the state of AI?
AAI development has hit a wall and progress is slowing.
BAI is advancing so fast that benchmarks, policy, and labour markets cannot keep up.
CChina has clearly overtaken the United States in AI.
DPublic adoption of AI remains limited to a small technical elite.
B is correct. The article repeatedly foregrounds the mismatch between capability speed and institutional response. A is contradicted by the data; C overstates the gap (the US lead has narrowed to 2.7%, not disappeared); D is false — 53% population adoption in three years.
Q2Which statement best captures the "jagged frontier"?
AAI models perform worse over time as data quality degrades.
BAI benchmarks are deliberately designed to be unfair.
CModels excel at some tasks and fail at surprisingly simple ones.
DAI systems have reached rough parity with humans on all tasks.
C is correct. The canonical example: gold-medal performance on the International Math Olympiad, yet failure to read analogue clocks half the time. The unevenness is the defining feature.
Q3What is the most defensible reading of the US–China adoption data?
AThe US dominates AI adoption because it dominates investment.
BBuilding AI and using AI have become separable questions; the US leads the former, trails in the latter.
CChina leads global AI adoption at the population level.
DAdoption rates correlate only with population size.
B is correct. The US hosts 5,427 data centres and leads in investment, yet ranks 24th at 28.3% population adoption. The UAE and Singapore lead. A is false; C is unsupported by the data shown; D is wrong — adoption correlates with GDP per capita, not population size.
Q4Which piece of evidence best supports the claim that AI infrastructure is a strategic vulnerability?
ANearly every leading AI chip is fabricated by a single foundry, TSMC, in Taiwan.
BAI investment reached $581 billion in 2025.
CAnthropic's top model leads by 2.7%.
DThe US hosts more than 5,000 data centres.
A is correct. Concentration risk — the entire global supply chain depends on one company on one island — is the vulnerability. B, C, and D describe scale or leadership, not exposure.
Q5What accounts for the 50-point gap between experts (73%) and the public (23%) on AI's job impact?
AThe general public has not yet encountered AI tools.
BExperts are paid to be optimistic by AI companies.
CSurveys systematically miss non-users.
DThe two groups have divergent lived experiences of the technology — power users on frontier models see a different product than occasional users.
D is correct. The article explicitly points to the gap between $200/month power users on coding and research tasks versus casual users encountering the "jagged frontier". The experience is different, so the opinions diverge.
§ 11 — True / False

Fact-check the claims

Based on the 2026 AI Index data presented.
1. Generative AI reached 53% US population adoption faster than either the PC or the internet.
2. As of March 2026, the best US model leads the best Chinese model by more than 15 Arena points.
3. The US ranks only 24th globally in population-level AI adoption, despite leading in investment.
4. The volume of AI researchers migrating to the US has dropped by roughly 89% since 2017.
5. Experts and the general public broadly agree on AI's likely impact on the job market.
6. AI data centre power capacity has reached roughly 29.6 GW — comparable to New York State at peak demand.
7. The Stanford AI Index concludes that AI development has clearly plateaued.
8. Almost every leading AI chip is fabricated by a single company (TSMC) in Taiwan.
§ 12 — Discussion

Argue with the data

Pick two prompts. Speak for 2–3 minutes each. Use at least three collocations from § 08.
[ PROMPT 01 ]
The US leads in AI investment and model releases, yet ranks 24th in population adoption. Is this a policy failure, a cultural difference, or a statistical artefact? Defend your position.
hint — compare to the UAE (64%), Singapore (60.9%), France (44%); consider GDP per capita, digital infrastructure, trust, and language.
[ PROMPT 02 ]
73% of US experts see AI's job impact positively; only 23% of the public agrees. Whose view do you trust more, and why?
hint — think about epistemic incentives, lived experience, the jagged frontier, and who bears the cost of being wrong.
[ PROMPT 03 ]
Given that the US–China performance gap has effectively closed, what should the EU's strategy be — sovereignty, specialisation, or abstention?
hint — consider the Mistral / Stargate / DeepSeek landscape, regulation (AI Act), and talent retention.
[ PROMPT 04 ]
If a single Taiwanese foundry (TSMC) fabricates almost every leading AI chip, how much of the "AI race" is actually a geopolitics-of-manufacturing story?
hint — decouple the software narrative from the hardware dependency; consider what "winning AI" would even mean in this frame.
[ PROMPT 05 ]
As an English teacher or edtech builder, which data point in this lesson changes your plans for the next 12 months — and how?
hint — adoption speed, expert–public gap, jagged frontier, or the shift from capability to cost-and-reliability.
§ 13 — Grammar focus

All / whole · each / every · some / any

Preston, Unit 29 — the quantifier distinctions that separate B2 from C1.

// Why these quantifiers matter here

The AI Index text is saturated with quantifier choices. The whole supply chain runs through one foundry. Every frontier lab now trains on comparable compute. All of the top four models are separated by razor-thin margins. Some countries lead in adoption; any of them could overtake the US. Each choice carries a slightly different meaning — and C1 writing depends on getting the nuance right.

// Core distinctions

all + plural / uncountable noun for the general set: All data centres need cooling. Before a specific noun, add the or of the: all (of) the data centres in Texas. All also means "the only thing": All we know is the top-line figure.

whole / entire emphasise completeness with a singular noun. Note the article position: all (of) the report → the whole report / the entire report. Both are C1-natural; entire is slightly more formal.

EVERY

The whole group collectively. Often paired with adverbials of frequency.

Every lab publishes a safety card.
The index is released every year.
every five minutes · every other day

Combines with nearly, almost, and abstract nouns like reason, chance, hope.

EACH

Individual items in a set, one by one. Used with smaller, definable sets.

Each model was evaluated separately.
The labs each reported their results.

Can sit mid-sentence: The models were each retested. — note the plural verb.

Does not combine with a numeral: ✗ each five minutes.

any in affirmative clauses means whichever / it doesn't matter which: Any of these benchmarks would do. In negative clauses it means none / not a single: I didn't see any inconsistency.

some appears in questions when you expect a positive answer or are offering something: Would you like some context? With a numeral, some means approximately: some twenty labs participated. With a singular noun it can mean a certain, unspecified: some researcher once argued…

Pronoun reference. The indefinites anyone / someone / everyone / no one are singular in form but take them / their as a neutral pronoun: If anyone claims AGI, ask them to define it.

// Exercise 1 — select the correct option
> click one option per slot. In a few items both may be valid — pick the stronger fit.
01. You need to read the wholeall the chapters of the AI Index before drawing conclusions.
02. She sat looking at entireall the analysts and thinking that the wholeentire conference was just theatre.
03. He had eachevery reason to distrust the benchmark results.
04. Is anybodysomebody still defending the "AI wall" hypothesis?
05. We have fifty frontier models, and everyeach one was evaluated separately, yet eachevery six months the methodology changes.
06. The mayor's reaction wasn't anythingsomething special — anybodysomebody with a policy brief could have written it.
07. He played safe, predictable models his entireall his career, so branching into agentic systems was definitely somethinganything extraordinary.
08. She cited the whole of theall the Arena leaderboard in a single footnote.
09. AllEverything we saw was a rather mediocre panel with speakers who were past their prime.
10. The wholeAll the situation felt absurd because everyeach ten minutes a new press release arrived.
// Exercise 2 — complete with the correct quantifier
> one word per gap. choose from: all · whole · entire · each · every · some · any · anybody · anything
01. The index is released year in April.
02. The global supply chain depends on a single foundry in Taiwan.
03. frontier lab now trains on comparable compute, and three months the rankings shift again.
04. If tells you the race is over, ask them to cite Arena.
05. The judges scored model independently, after which the scores were aggregated.
06. She reviewed the findings for chapter, but skipped the appendix.
07. We didn't find conclusive, and neither did the peer reviewers — the experiment has to be redone.
08. the top four models cost under ten cents per million tokens — lab competes on price now, not capability.
09. Is there at your organisation who still hasn't used a generative model?
10. you really need is a clear evaluation on your own workload.
// Exercise 3 — find and correct the error
> each sentence contains one quantifier error. rewrite the corrected version. tap "reveal" to check.
01.The every model we tested failed on the clock-reading task.
// model answer
Every model we tested failed on the clock-reading task. — we do not say "the every"; every already scopes the whole set.
02.I've read the most of the 2026 AI Index report.
// model answer
I've read most of the 2026 AI Index report. — no article before most of.
03.All data centre requires cooling infrastructure.
// model answer
Every data centre requires cooling infrastructure. — all + singular countable is ungrammatical; use every (or all data centres).
04.Anybody who reads this report can see the expert–public gap.
// model answer
Anybody who reads this report can see the expert–public gap. — no error. sentence is already correct; anybody in an affirmative clause means "whoever, no matter who".
05.They interviewed each five experts from major labs.
// model answer
They interviewed five experts from each major lab. — each does not combine with a numeral; use every, or rephrase as each + singular noun.
06.The whole of analysts agreed that capability was accelerating.
// model answer
All of the analysts agreed that capability was accelerating. — whole is for singular nouns; all of works with the plural.
07.Is there some questions you'd like to ask the panel?
// model answer
Are there any questions you'd like to ask the panel? — neutral question with plural subject — any and plural are. Some would imply you expect a yes.
08.He wrote the entire report in all weekend.
// model answer
He wrote the entire report in one weekend. — or: …in a single weekend. All weekend is adverbial ("throughout the weekend"), incompatible with in.
§ 14 — Takeaways

Lesson summary

What the data tells us; what it doesn't.
  1. Capability is still accelerating. Forecasts of a plateau have been contradicted; benchmarks like SWE-bench Verified rose from 60% to near-100% of human baseline in a single year.
  2. The US–China performance gap has effectively closed. The top four frontier models are separated by razor-thin margins. Competition has moved from raw capability to cost, reliability, and real-world usefulness.
  3. Adoption is historically unprecedented. Generative AI reached 53% US population adoption in three years — faster than the PC or the internet. But adoption and investment have decoupled: the US leads the latter and trails the former.
  4. Infrastructure is structurally fragile. One foundry on one island makes almost every leading chip. The hardware dependency is a shared systemic risk, not a US-versus-China one.
  5. The public–expert gap is the real story. Divergent lived experience of the technology — "jagged frontier" vs power-user frontier — produces divergent beliefs about AI's consequences.
  6. Benchmarks are necessary but insufficient. A model can win mathematical olympiads and still misread a clock. Evaluate on the work that matters to you.
The headline finding of 2026 is not that AI won. It is that AI is running faster than the institutions built to measure, govern, and absorb it.
// Post-lesson checklist — tap to mark done
I can describe three quantitative trends using varied collocations.
I can explain "the jagged frontier" in my own words.
I can argue a position on AI adoption using at least one number from § 04.
I have identified one data point relevant to my own professional context.