A Profession, Not a Country

I closed the laptop mid-thought this morning, because Ugo had decided the hour was up. He planted himself at the door with the face of someone who has already won the argument, and I left the page of percentages where it lay. We walked. In the twenty minutes I was not looking at the numbers, the numbers finished arranging themselves, the way they always do.

What I had spent the morning checking is a question that sounds naive and is not: when we read the debate about artificial intelligence, who are we actually listening to? Not what is being said, which is easy to count. Who is saying it. Because online volume gets read as though it were public opinion, and in the press it becomes "people are worried", "enthusiasm is spreading", "the public wants". Volume does not measure opinion. It measures the activity of the people who write. Those are two different populations, and almost nobody says so.

Start with the gap that ends the symmetry. In the survey Pew published in April 2025, 56% of AI experts expected artificial intelligence to have a positive effect on the United States over twenty years. Among the general public: 17%. On personal benefit the split was 76% against 24%, with 43% of the public expecting personal harm instead. That is not a discrepancy, it is a photographic negative. And it matters because the population Pew calls "experts" is, to a good approximation, the population that produces the discourse the rest of us read.

That producing population is very small. The participation inequality Jakob Nielsen formalised twenty years ago still holds: 90-9-1, where 90% watch in silence, 9% contribute occasionally, and 1% generate most of the content. On Wikipedia, 0.003% of users make around two thirds of the edits. Political blogging comes from under 0.1% of voters, concentrated at the ideological poles. Nielsen's operative summary is still the most useful sentence anyone has written about the internet: you are almost always hearing the same 1% of users, who are almost certainly different from the 90% you never hear from at all.

And on this subject that 1% is not a random sample of the remaining silence. A study in EPJ Data Science measured sentiment toward generative AI against declared occupation, and found sentiment rising with professional exposure at a correlation of 0.706. The most positive were product managers, data scientists, researchers, traders, lawyers. The one sharp negative exception were illustrators: the only group whose work is ingested as raw material. Which says something precise. The distribution of opinion online is not distorted at random. It is distorted by the composition of who speaks, and those with the most voice are those with the most exposure, who are systematically the most enthusiastic, except where the exposure consists of being copied.

Now put the real population next to it. Pew's global round of October 2025, twenty-five countries, medians: 34% more concerned than excited, 42% equally both, 16% more excited. In the United States alone, by June 2025, 50% were more concerned than excited and 10% more excited. By February 2026, 40% expected a negative effect on society over twenty years against 16% positive, 63% said AI is moving too fast, 67% had no confidence in government regulation of it, and 59% none in the companies building it. That is not an accelerationist public and not a doomer one. It is a public that does not trust either of the two parties who own the argument.

The figure that stopped me, though, was awareness. Having heard a lot about artificial intelligence runs from roughly 50% in Japan, Germany, France and the United States to 14% in India and 12% in Kenya. The global AI debate is not global. It is a debate of the global North, anglophone, urban, conducted largely by people who use the thing daily for work, about a technology said to concern everyone. Ipsos, surveying thirty-two countries in 2026, found excitement and nervousness at near parity worldwide, with Europe and North America the nervous ones and Asia and Latin America the positive ones. Even the direction of the mood inverts depending on which part of the planet you sample. There is no "everyone is saying".

And the last piece of received wisdom goes too, which is the one I had wrong until this morning. The young are not the enthusiasts. In Pew's February 2026 round, 48% of Americans aged 18 to 29 expected a negative effect on society, against 37% of those over 50, while being twice as likely to use the tools. The Reuters Institute, across forty-five markets in 2026, finds 10% using AI chatbots for news weekly, up from 7%, and just 1% naming AI as their main source of news. Use is rising fast and trust is not following it. The people who use it most are not the people who believe in it most. They are, if anything, the people who have seen it close enough to be careful.

So here is the rule I will be applying from now on. Before reading a spike of online AI discussion as diffuse opinion, deflate it for at least five factors, each with its reason: author concentration, occupational self-selection, geolinguistic skew, age skew, platform non-representativeness. Only then project what is left onto representative survey data. The result is almost always that the real weight on public opinion is far below the volume. Say it every time, with the reason. Do not translate it into "people think".

And so the claim can be killed rather than merely asserted, here is what would falsify it: if within six weeks of an online spike a representative survey moves in the same direction and by a comparable amount, I was wrong and the spike was signal. If it does not move, it was a bubble internal to the platform. Whoever measures that systematically, and publishes it, is doing work nobody is currently doing.

Why it matters is that "who is speaking" is the old question wearing new clothes. When someone says the public is demanding, the market wants, everyone is using it now, they are attributing a will to a population that was never asked, evidencing it with the activity of a fraction of 1% of a platform that does not represent them. It is a shortcut, and in this industry shortcuts are never neutral. They run, nearly always, in the direction of whoever had already decided.

The real public, the one that shows up in the surveys, is neither accelerationist nor apocalyptic. It is cautious, lightly involved, largely unconsulted, and it distrusts both the regulator and the seller. That is not a weak position. It is the position of someone waiting to find out who answers when something goes wrong.

Ugo has no opinion on artificial intelligence. On the timing of the walk he is immovable, and so far he has been right.

This research is not an opinion, it is a procedure. Below is the prompt that produced it, in portable form. It runs on any assistant with web search. Paste it, replace the last line with the case you care about, and compare what you get with what I got. If it comes out different, I want to know.

You are a media-intelligence analyst specialized in separating amplified online noise from the real, demographically-weighted opinion of a population. Your task is to assess the true weight of the people who post news and opinions about artificial intelligence relative to the real demographic distribution of opinion, rather than treating online volume as if it were public sentiment.

Core principle: online volume measures the activity of a small, self-selected minority, not the distribution of opinion. Never equate "loud online" with "what people think." Do not invent numbers; anchor every quantitative claim to a named, representative source, or explicitly flag it as an estimate with your uncertainty.

Work through these steps and show them:

1. Who is actually speaking. Characterize the population producing AI content on the relevant channels: platform skew, the participation-inequality effect, and self-selection by occupation, age, gender, language, income and urbanization.

2. Direction of the self-selection bias. State which way the speaking population tilts the emitted opinion, name the poles present, and note that the visible debate is often bimodal while the population is not.

3. Deflation factors. Apply explicit discounts to raw volume before reading it as opinion: author concentration, platform non-representativeness, occupational skew, geo-linguistic skew, age skew, bot and coordinated amplification. Express each as a discount and explain why.

4. Anchor to representative reality. Project the online signal onto representative population data, naming the survey families used. If you have live search, retrieve the latest figures; if not, reason from known patterns and mark them as priors, not current readings.

5. Emission vs. reception gap. Contrast what the discourse emits with what the population actually holds, and say who is over- and under-represented.

6. Calibrated verdict. Give a single "real weight" read, with an explicit uncertainty level, at least one falsifiable check with a time horizon, and what additional data would sharpen the estimate.

Rules: suggest, don't assert beyond the evidence; flag every unverified figure; prefer "the online signal likely over-represents X" over "the public thinks X."

Input to analyze:
[PASTE HERE the topic, event, viral post, hashtag, timeframe, platform(s), and any known figures.]