The gap is the story

Two feelings, one population

Across 32 countries in spring 2026, 51% of adults said products and services using AI make them excited. 50% said the same products make them nervous. Those are not two camps of matching size standing on opposite sides of a room. In most markets they are substantially the same people, holding both feelings at once.

That near-parity is the most important fact in the survey record and the first casualty of any headline. A summary that leads with rising optimism is telling the truth. A summary that leads with rising anxiety is telling the truth too. Neither is telling you the thing that matters, which is that the two moved together.

The Anglosphere is the tightest bloc in the data, and the gloomiest: Australia and Canada at 67% nervous, New Zealand 65%, Ireland 63%, the United States 64%, Great Britain 62%. All six sit above the global average on nervousness and below it on excitement. No other way of grouping these countries — by income, by region, by regulatory regime — sorts them half as cleanly.

Not opposite ends of one scale

If excitement and nervousness were one attitude measured twice, countries would fall along a downward diagonal: enthusiasm here, dread there. They do not. Knowing where a country sits on one axis tells you surprisingly little about where it sits on the other — and the countries furthest from the line are the ones worth naming.

Ipsos reads the split as being about the status quo rather than about development. The argument: where people want their existing arrangements changed, AI looks like a shortcut; where they want them protected, it looks like a threat. That is a story told about a correlation, not a mechanism anyone has tested, and it deserves that label. It does survive one obvious objection better than the usual rich-versus-poor framing — Singapore, South Korea and China are not poor countries, and all three sit firmly on the enthusiastic side.

Where the split becomes concrete

Force the choice — AI will create new jobs and new ways of working, or it will eliminate jobs and disrupt industries — and the geography of Figure 1 reappears almost intact, with the edges sharpened. Canada and the United States are the two most pessimistic markets surveyed. Nigeria, Japan and Mexico are the three most optimistic. There is no middle to speak of.

Note what this measure is not. It asks about the job market in the abstract, not about the respondent's own job — and the two answers diverge everywhere. In the United States, 30% expect AI to improve their own job against 18% who expect it to improve the job market. That asymmetry is routine in labour-market polling, and it is not evidence of muddle. People are distinguishing their own bargaining position from everyone else's, which is a perfectly coherent thing to do.

The expert gap, and the part of it that inverts

The most-cited comparison in this literature sets what AI experts expect over the next twenty years against what the American public expects. The gaps are enormous, and outside a handful of domains they all run the same way: experts are more optimistic, by as much as 50 points.

Where the two groups agree, they agree gloomily. Elections, the news, personal relationships: roughly one in ten of each expects AI to improve any of them. Consensus in this dataset is a warning sign, not a reassurance.

The more revealing chart is the occupational one, because there the gap changes sign. Ask which jobs AI will reduce, and experts and the public converge on cashiers, factory workers and journalists. Then they cross. Experts think truck drivers are at nearly twice the risk the public does. The public thinks teachers and doctors are at substantially more risk than experts do.

Read the crossings as a disagreement about task structure rather than about capability. Experts price in physical autonomy they expect to arrive, and discount the credentialed, liability-bearing, in-person residue of teaching and medicine. The public prices it the other way round: it has seen a machine write a lesson plan and read a scan, and it has not seen a truck drive itself down its own street. Each group is extrapolating from what it can observe. Neither has been proved wrong yet, and the occupations where they cross are the ones to watch.

Does using it make you like it?

The industry's working assumption has been that familiarity produces comfort. The most repeated counter-claim is that the heaviest users are now the angriest. Both are drawn from the same Gallup survey of 14- to 29-year-olds. Only one of them survives contact with the crosstabs, and it is not the one you would guess from the headlines.

Within 2026, daily users are dramatically more positive than non-users: 44% excited against 4%, 18% angry against 59%. Across the two years, every group fell — and daily users fell fastest, down 18 points on excitement.

So "exposure breeds distrust" is not what the data says. What it says is narrower and more uncomfortable: using the tools still predicts liking them, and the return on using them is falling every year. A gradient across people and a slide across time are different objects. Quoting the second while asserting the first runs the arithmetic backwards, and it is the single easiest error to make with this dataset.

The say-do gap, measured directly

Ipsos puts the question in the bluntest available form: I don't always trust AI tools, but I use them anyway. Globally, 63% agree. In the United States, 49% do — the lowest figure of any country surveyed. Read carefully, that is not a sign of American confidence. It is a sign of lower American use: you cannot report using something warily if you are not using it.

Everyone wants a referee. Nobody trusts the referee.

The United States records the lowest trust in its own government to regulate AI responsibly of any country surveyed — 31%, against a global average of 54% and a Singaporean high of 81%. The same country leads the world in private AI investment by a factor of twenty-three, hosts more than ten times as many data centres as anyone else, and produces most of the frontier models. Capability and confidence have come apart completely, and nothing in the data suggests they are on their way back together.

The American crossover

Until recently, distrust of government AI regulation was a Republican position. Between 2024 and 2026 the parties changed places: Democratic distrust rose 20 points, Republican distrust fell 9, across a change of administration. The same crossover, smaller, shows up in what each side thinks of the companies. Whatever this measures, it is not a settled view about artificial intelligence.

Underneath the partisanship, the direction of demand is stable and one-sided. In all fifty states, more people say federal regulation will not go far enough than say it will go too far. The gap widens sharply with age and with education. Party barely registers: 5 points between Democrats and Republicans, against 13 points between the youngest and oldest respondents and 12 between the least and most educated.

The stable finding, then, is not that people want regulation. It is that people want someone to be accountable and cannot identify who that would be. Distrust of government and distrust of industry move independently, they have now swapped partisan owners once inside two years, and the one thing that has not moved is the share of people who answer "not sure."

Where the distance goes missing

Every figure in this note was checked against the instrument that produced it. Fifteen did not survive the check in the form they circulate in.

None of them is a fabrication. A column gets transposed. Two surveys fielded two years apart get merged into one row. A release date stands in for a field date. A correlation acquires a mechanism on the way to the summary. These are the ordinary failure modes of secondary reporting, and their defining property is that they read perfectly well — nothing looks wrong until someone opens the source.

Seven of the fifteen change a conclusion rather than a decimal.

Checked and confirmed

The following were traced to primary sources and stand exactly as commonly stated. The check is not a verdict on the literature — most of it holds.

Open, not checked

How to read any of this

Three cautions carry across every chart above.

Online panels are not populations. Ipsos states plainly that its samples in fifteen of the thirty-two markets — including India, China, Indonesia, Brazil and South Africa — are more urban, educated and affluent than the countries they represent. Those are precisely the markets driving the "Global South is enthusiastic" story. The finding may still be real. It is measured on the connected minority.

Question wording moves these numbers more than reality does. Global workplace AI use is 58% in one instrument and 21% in another for the United States alone. Neither is wrong; they ask different questions of different populations with different reference periods. Any comparison across surveys needs the wording in front of it.

Sponsorship is not disqualifying, but it is disclosable. Several figures in the circulating literature come from studies commissioned by parties with a stake — a bank on affluent investors, a progressive think tank on worker attitudes, a search company on AI priorities. The fieldwork is generally sound. The question selection is never neutral.