All Else Unequal

The average user doesn't exist.

What a $124 average hides about who actually gets the value of AI.

Plate 1The trade

In March 2026, Stanford economists offered 1,908 American AI users one random sum to give up every AI tool (ChatGPT, Gemini, Claude, Copilot) for a month, starting tomorrow. Before you see how they answered: what is the least you would take?

Your price for a month$22

Source: Brynjolfsson, Collis et al. (2026), figure 2 Survey: March 2026 CSV CC BY 4.0

The price of a month

Opinion

If someone offered to pay me to stop using AI for a month, starting tomorrow, I would ask for somewhere between $100 and $500.

That number isn't about convenience. It's about what the month would cost me. I'm a product owner for internal tools and reports at a large company, and the analyses I deliver, and most of the code behind them, now go through AI. The few times I've had to program without it, I did the work by hand. It got done, but my usual work would take me ten times longer. AI is no longer a tool I open; it is built into how the tasks get done. I don't work the same way without it.

A year ago I would have asked for much less. Hold on to that sentence. The rest of this essay is about it.

The question

Data

The question you just answered is the one economists at Stanford's Digital Economy Lab put to 1,908 American AI users in March 2026. Not how much they liked AI, but a trade. Each person was offered one random amount ($1, $10, $20, $50, $100, $200 or $500) and could only say yes or no.

Here is how they answered. Offered a single dollar, three in ten took it. At $10, half did. Past that, more money buys fewer people: doubling the offer from $100 to $200 moved only three more in a hundred. And at $500, the highest offer, 12% still refused.

Mean ≠ median

Model

The paper's headline is the average: the average user would need $124.50 to give AI up for a month. Multiply by 115 million American users and twelve months and you get $172 billion of value a year, more than twelve times the $14.2 billion that OpenAI, Anthropic, Google and Microsoft earned from consumers worldwide in 2025. By this measure, users keep most of the value AI creates. That is the finding most headlines took away.

But the median is $11.48. Half of American users would sell a month of AI for less than the price of a lunch.

When an average is ten times its median, it describes almost nobody. Line a hundred users up and raise the offer to the “average” price: only 28 still refuse the money. Seventy-two in a hundred value AI less than the average user.

Plate 2A hundred users

The same answers as a crowd of 100. Raise the offer and whoever it reaches lies down. Point at anyone to see their price.

$11.48

Point at someone to see their price.

Source: Brynjolfsson, Collis et al. (2026), figure 2. Each person's price is read between the seven offers. Survey: March 2026 CSV CC BY 4.0

Model

There is no typical user who values AI at $124. There is a large majority for whom it is handy and replaceable, and a minority for whom it has become part of how they work and think: twelve in a hundred would not give it up even for $500. The $124 is what you get when you average the two. And because the survey stops at $500, the real average is higher still. The tail pulls harder than the number shows.

Not money. Practice.

Data

So what separates the $11 user from the $500 one? The paper looks at demographics and usage together, and the answer is not what you would guess.

Paying for a subscription adds $2.90 a month to the median valuation. A college degree adds nothing measurable. Income barely matters, and neither does which tool you use. What matters, by a wide margin, is how often you use it: frequent users value AI $36.90 a month more than occasional ones. Using it at work adds another $9.10.

Plate 3What moves the value

Extra median value of AI a month that goes with each trait, holding the others constant. How often people use it dwarfs everything else.

No significant effect: college degree, living in a rural area, using several AI products, which tool. Income matters only weakly.

Source: Brynjolfsson, Collis et al. (2026), figure 3 Survey: March 2026 CSV CC BY 4.0

Model

AI is an unusual good in this respect. For most goods, what a thing is worth to people rises with what they can afford. Here it rises with how much they have practised.

The loop

Opinion

The survey cannot tell which way that runs. Maybe people who value AI use it more. Maybe using it more is what makes it valuable. My bet is both, feeding each other: you use it, you learn what it is good at, it becomes more useful, you use it more.

Data

The paper holds a trace of that loop over time. In late 2024, the median user would have given AI up for $0.52. By July 2025 it was $3.39. By March 2026, $11.48: twenty-two times higher in fifteen months.

Plate 4The median climbs

The median value American users put on a month of AI, in three surveys: twenty-two times higher in fifteen months.

Source: Brynjolfsson, Collis et al. (2026), table 2 and note 4 Surveys: Dec 2024 to Mar 2026 CSV CC BY 4.0

Opinion

Better models explain part of that. Practice, I think, explains the rest.

I am one of those points. A year ago I asked AI questions. Now I build with it. What changed was less the tool than me: I learned to ask for the right thing, to check what comes back, to chain one task into the next. Every one of those lessons made the next month of use worth more.

Europe: two gaps that add up

Data

No one has run this survey in Europe. What exists is Eurostat's first count of who uses generative AI at all: in 2025, 32.7% of Europeans aged 16 to 74.

That average hides a spread of its own. In Denmark, 48% of adults use it; in Romania, 18%. And the gap doesn't stop at access. In the countries where more people use AI, a larger share of those users also use it for work, the kind of use the American data links to higher value. Across the 27 member states the two move together (r = 0.75). Combined, 27% of Danish adults use AI at work, against 5% of Romanians.

Plate 5Two gaps

Each dot is a member state. Where more adults use generative AI, more of those users also use it for work (r = 0.75).

Keyboard: focus the chart, then use the arrow keys to step through the countries.

Source: Eurostat, isoc_ai_iaiu, people aged 16 to 74 Updated 5 Jun 2026 CSV CC BY 4.0

Opinion

So Europe faces two gaps at once: between countries, in who has started, and between users, in how deep they have gone. Official statistics count the first. Nobody in Europe measures the second: Eurostat counts who uses AI, not what it is worth to them. If the American data is any guide, the second gap is the one that decides who gets the value, and it is the one we cannot see.

Data

The European and American figures are not directly comparable. Eurostat counts use in the last three months among 16-to-74-year-olds; the American survey uses a different question and age range.

A practice gap, not an access gap

Opinion

If the value of AI depended on money or credentials, we would know what to do: subsidise access, widen education. The data points somewhere less comfortable. Access is already nearly free. Most users pay nothing, and paying barely changes how much they value it. What separates users is practice: daily use, at work, on real problems.

That has a consequence for schools and training that goes beyond “AI literacy”. Knowing what a model is doesn't move anyone along the curve. Two things do: using it every day on real tasks, and the judgement to know when to trust what it gives back. The second only grows out of the first. And a gap built on practice compounds, because every month of use makes the next one more valuable. Education that treats AI as a topic will widen the gap. Education that makes it a daily tool, and teaches people to check it, might close it.

The gap also shows where pricing is heading. Today the price of AI barely tracks its value: a heavy user and an occasional one pay about the same, often nothing. I expect the two ends to move apart. Competition and open models will keep the basic tier cheap for most people. The heavy users, the ones who would not give it up for $500, will pay more, through plans priced by use. Much of the surplus that flows to the tail today is what the next pricing model will come for.

The honest doubt

Opinion

There is a question this data can't answer, and it is about me.

When I say I'd ask for a few hundred dollars to give up AI for a month, is that because it gives me that much value? Or because I no longer work the same way without it? Economists have long known that people demand more to give something up than they would pay to get it in the first place. The survey measures exactly that: the price of losing something you already have.

I don't think the two can be separated. If my usual work would take ten times longer without AI, that is value and dependence in the same sentence. The most valuable tools tend to become the ones we can't work without. The people at the top of the curve, myself included, are also the people for whom a month without it would hurt most. Value and dependence may be the same number, read from two sides.

How this was made

The data

  • United States. Brynjolfsson, Collis, Eggers, Kazinnik and Nguyen, What is Generative AI Worth?, Stanford Digital Economy Lab working paper, 13 April 2026. Not peer reviewed. An online survey on Prolific with US quotas by age, gender and ethnicity, of generative AI users only: 1,908 people in March 2026, 1,491 in July 2025 and 502 in December 2024.
  • Europe. Eurostat, Individuals - use of generative AI tools (isoc_ai_iaiu), 2025, updated 5 June 2026. It is the first year of the series, so there is no trend yet. No cell used here carries a low-reliability flag.
  • Consumer revenue. The $14.2 billion that OpenAI, Anthropic, Google and Microsoft earned from consumers worldwide in 2025 is Gartner's figure, as cited in the paper (note 5).

How the plates were built

  • The seven observed shares are the centres of the markers in the paper's figure 2, read from the PDF's vector geometry. With about 270 people per offer, each carries a sampling error of about ±6 points, which is why $10 and $20 come out equal.
  • The fitted curve is the paper's logit on the log of the offer, with its median of $11.48 and a slope of 0.414 measured on figure 2's fitted line. Capped at $500, it implies a mean of $123.70, against the paper's $124.50.
  • Each of the hundred users in plate 2 has a price read between the seven observed shares, joined in log price. Who lies down first is a fixed shuffle: the plate shows shares, not real people.
  • The European correlation (r = 0.75, n = 27) is our own calculation from Eurostat's table.
  • The paper's text gives $2.27 for the July 2025 median in two places; its table 2 and abstract give $3.39. We use table 2.

Limits

  • The survey asks what people would accept to give AI up, not what they pay for it. People tend to ask more to give something up than they would pay to get it, so values stated this way can overstate it. The order of magnitude of the surplus is sturdier than its exact figure.
  • The authors list hypothetical bias, an online sample, a one-month horizon, the $500 cap and AI built into other products left uncounted, and say this is not a full welfare analysis.
  • The traits in plate 3 go with higher value; the survey cannot say they cause it.

Cite this essay

Víctor Simó Baeza (2026). The average user doesn't exist (≠01). All Else Unequal. 
@misc{aeu01,
  author       = {Víctor Simó Baeza},
  title        = {The average user doesn't exist},
  year         = {2026},
  howpublished = {All Else Unequal, ≠01},
  url          = {}
}

Data and licence

Every plate and every data file on this page is published under CC BY 4.0: use them in a class, a slide or an episode, adapted or not, with credit to All Else Unequal.

  • demand-curve-us-2026.csvThe share of American AI users who keep AI at each of the seven offers, observed and fitted. Plates 1 and 2.
  • value-drivers-us-2026.csvExtra median value a month by trait, and the traits with no significant effect. Plate 3.
  • median-value-us.csvThe median value of a month of AI in each survey, with its number of respondents. Plate 4.
  • genai-use-eu-2025.csvUse of generative AI in the EU-27 and its members, 2025: overall, for work, by age and by sex. Plate 5.