For most of history, expert thinking was scarce. You waited weeks for a specialist, paid a lawyer by the hour and hired a consultant by the day. Now anyone with a phone can ask a model that answers many expert questions well, at any hour, for free or about $20 a month. When something scarce gets cheap, the value moves to what stays scarce: attention, trust, judgment and people.
The change touches everyone who sells what they know: lawyers, doctors, teachers, developers. Who gains, who loses, and what happens to our own thinking? I wanted to know what researchers have actually measured, not what people predict.
This note gathers what I found. The numbers come from Epoch AI, Stanford, MIT, the World Bank, Microsoft and Anthropic, and the links are at the end. When a study is small or not peer reviewed, I say so.
Expertise came by appointment
An expert has only so many hours, and you can't copy them. So we rationed expertise by price and by waiting list.
School shows what that costs. In 1984, the education researcher Benjamin Bloom measured what a private tutor buys. The average student tutored one to one did better than 98% of the students taught in a normal class. He called it the "2 sigma problem": the result was clear, but no school could pay a tutor for every child.
A model changes the arithmetic. Once it's trained, it can run on as many chips as you have, all at once, day and night. In June 2025, Sam Altman, who runs OpenAI, wrote that "the cost of intelligence should eventually converge to near the cost of electricity."
History says to be careful with that kind of promise. In 1954, Lewis Strauss, the head of the US Atomic Energy Commission, said nuclear power would make electricity "too cheap to meter". It never happened: we still pay for every kilowatt-hour. For AI, though, we don't have to take anyone's word for it: the price fall has been measured since 2023.
Before
By appointment
One expert, a few hours a day
Paid by the hour or the day
A private tutor for the families who can pay
now
Now
On demand
One model, on as many chips as you have
Free, or about $20 a month
A tutor in every phone, at any hour
Bloom, 1984One-to-one tutoring beat 98% of a normal class.
Altman, 2025Intelligence will cost about what electricity costs.
1954Nuclear power, "too cheap to meter". It never was.
Ten times cheaper every year
The price of a given level of AI falls fast. In February 2025, Altman put it at about ten times every twelve months. His example: the price per token dropped about 150 times from GPT-4, in early 2023, to GPT-4o, in mid-2024. A token is the unit models are billed in, about three quarters of a word.
Epoch AI, a research group that tracks AI progress, measured it more carefully. For a fixed level of performance, prices fell between 9 and 900 times a year, depending on the task. Getting GPT-4's score on PhD-level science questions became 40 times cheaper each year. Epoch adds that the fastest drops are recent, so they may not last.
Ten times a year is hard to picture, so here is the arithmetic. A task that costs €100 today costs €10 next year and €1 the year after. Altman compares it with Moore's law, the doubling of chip power every 18 months that shaped computing for fifty years. That's about 1.6 times a year.
Cheaper per answer doesn't mean a smaller bill. A team at MIT found that for the best models, the price of a given score falls 5 to 10 times a year. But the price of running the newest models rises 3 to 18 times a year, because they're bigger and think longer before they answer. People move to the newest model, so their bill can grow even as each answer gets cheaper.
That pattern has a name. In 1865, the economist William Stanley Jevons noticed that better steam engines didn't save coal: Britain burned more of it, because more uses paid off. In January 2025, after the Chinese lab DeepSeek released a cheap model, Satya Nadella, Microsoft's CEO, posted: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket."
Today€100
Next year€10
The year after€1
MITRunning the newest models costs 3 to 18 times more each year.
Beginners gain the most
When expertise gets cheap, the people with the least of it gain the most. That's the most hopeful finding I read.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,179 customer support agents who got an AI assistant at work. On average, they solved 14% more issues per hour. New and less skilled agents solved 34% more. The most experienced agents barely changed. The authors think the assistant spread the know-how of the best agents to the new ones.
The same shows up at school. The World Bank ran a trial in Benin City, Nigeria, with Microsoft Copilot, an assistant built on GPT-4. Secondary school students used it for English after school, with their teachers in the room, for six weeks. Their gain was worth 1.5 to 2 years of normal schooling. That's less than Bloom's private tutors, but it took six weeks, with no tutor's salary to pay.
David Autor, an MIT economist who studies jobs and technology, sees a chance for the middle class here. He calls a tool that works with people, instead of replacing them, "a force multiplier for expertise". His example is the nurse practitioner, a nurse with extra training who does work once kept for doctors. AI could let more people without a long degree do expert work like that.
NigeriaSix weeks of AI tutoring after school: 1.5 to 2 years of normal school.
Autor"A force multiplier for expertise."
The first rung is shaking
The same tools that help beginners also do beginners' work. The hiring data now points there.
Brynjolfsson and two Stanford colleagues follow payroll data from ADP, the largest payroll company in the US. In jobs exposed to AI, like software development and customer service, they track people aged 22 to 25. By June 2026, their employment stood 19% below where it would be if it had kept pace with young people in less exposed jobs. Experienced workers in the same jobs show no such gap. Most of it comes from fewer hires, not more layoffs.
The authors first saw the gap in August 2025, and it has grown since. They are careful: other causes, like interest rates, are still being tested. But the drop shows up where AI does the work, and not where it helps people do it.
The mechanism is easy to see in software. Juniors used to learn on small tasks: a config change, a first test, a fix in a README. Coding agents can now take those tasks. You learn judgment by doing small work and getting some of it wrong. So who becomes the next senior? An earlier note on this site covers PRAISE, a Microsoft program built on that question. It pairs each junior with a senior whose job is to "ship seniors".
- SeniorExperienced workers in the same jobs: no comparable gap
- Mid-careerExperienced workers in the same jobs: no comparable gap
- Entry level, 22 to 2519% below where they would be, mostly fewer hires
The paradoxAI helps beginners the most, and it does beginner work.
Cheap to buy, unequal to use
Access is cheap, but using it well isn't spread evenly. That's true between countries, and between people.
Anthropic, the company behind Claude, publishes how people use its model. In its September 2025 report, use follows income. A country with 1% more income per person uses Claude about 0.7% more, counted per working-age person. Inside the US, the link is even stronger: 1.8%.
Then there's skill. In late 2023, a team led by Stanford researchers gave 50 US doctors a set of clinical cases. Half could use GPT-4 on top of their usual resources, half had only their usual resources. The doctors with GPT-4 scored 76%, the doctors without it 74%: no real difference. GPT-4 on its own scored 16 points higher than the doctors without it.
The model was there, and it was good. The doctors didn't get much out of it. So the gap that matters now is in how people work with the model, more than in who can buy it.
Anthropic1% more income per person: about 0.7% more use of Claude.
Thinking you can skip
The risk for each of us is quieter: we can skip the thinking and still get the answer.
In 2025, a team at the MIT Media Lab asked 54 people to write essays over several sessions. Each person wrote with their brain only, with a search engine, or with ChatGPT. They recorded brain activity with EEG, sensors on the scalp. The brain-only group showed the strongest networks, the search group came next, and the ChatGPT group showed the weakest. The ChatGPT users also had trouble quoting their own essays. The study is small and came out as a preprint, so I read it as a warning sign, not as proof.
A survey by Microsoft Research and Carnegie Mellon University points the same way. They asked 319 knowledge workers about 936 real tasks they did with AI. The more people trusted the AI, the less critical thinking they reported. The more they trusted their own skills, the more they checked.
The same risk applies to teams that let AI agents write code: they can ship more changes than anyone can read. Two habits from earlier notes help: review the plan before an agent starts, and ask what each green check proved. Both keep the decisions with people.
1Brain only
The strongest, most spread-out networks
2Search engine
Moderate
3ChatGPT
The weakest, and trouble quoting their own essay
Microsoft and CMUMore trust in the AI, less critical thinking. More trust in yourself, more checking.
What stays scarce
When one thing gets cheap, the value moves to what it needs and can't make itself. In 1971, the economist Herbert Simon put it in one line: "a wealth of information creates a poverty of attention." Here's what I see staying scarce when intelligence is cheap.
- Attention. Answers are free, but reading them takes your time, and nobody gets more than 24 hours.
- Trust. Graphite, an SEO company, counts about half of new web articles as mostly written by AI. Yet in a 2025 sample, 86% of the articles on Google's first two pages were written by people. Knowing who wrote something, and who stands behind it, is worth more than before.
- Someone who signs. A model can suggest a diagnosis or an architecture. A doctor signs the prescription. An engineer signs off on the design and answers for it when it breaks.
- People. In a four-week study by the MIT Media Lab and OpenAI, 981 people used a chatbot. Those who chose to use it more ended up lonelier and saw other people less. It's a link, not proof: lonely people may simply use it more.
- Electricity and chips. Cheap thinking runs on power plants and chip factories. The International Energy Agency expects the US to use more electricity in 2030 for processing data than for making aluminium, steel, cement and chemicals combined.
For anyone who sells what they know, that list is the part of the job that keeps its value. It's attention to the real problem, the trust of the people you work for, and a name that answers for the work. Knowledge on its own gets cheaper every year.
Cheap nowStays scarceThe sign
AnswersAttention: your 24 hoursNobody gets more time
TextTrust in who wrote itAbout half of new web articles are AI-written
AdviceSomeone who signsThe doctor, the architect
ConversationTime with peopleHeavier chatbot use goes with more loneliness
ThinkingElectricity and chipsUS data centres pass heavy industry by 2030
Simon, 1971"A wealth of information creates a poverty of attention."
Try it
Every source is below. If you only open one, take the Stanford paper on young workers: it's updated as new payroll data comes in.
- Sam Altman, Three Observations and The Gentle Singularity, 2025.
- Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, and the MIT team's The Price of Progress.
- Brynjolfsson, Li and Raymond, Generative AI at Work, and Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine?
- The World Bank, From Chalkboards to Chatbots, and Goh and colleagues, Large Language Model Influence on Diagnostic Reasoning.
- Anthropic, the geography of AI use.
- Kosmyna and colleagues, Your Brain on ChatGPT, and Lee and colleagues, The Impact of Generative AI on Critical Thinking.
- My note on the price side, intelligence on tap, from Garry Tan and Mark Pincus.
Then pick one thing you paid an expert for this year: a letter from a lawyer, a tax question, a code review. Ask a model first. Write down what you still needed a person for. That's your own list of what stays scarce. If you try it, tell me what was on yours.
Plain text: abundance-of-intelligence.md · Atom feed · All notes