Executive Summary

The Gallup release of September 22, 2026 was headlined "AI Optimism Globally Widespread Despite Uneven Use." It carries the first results of ongoing research with Microsoft, covering the thirty-seven countries that completed fieldwork this year. The project is ultimately heading for 140 countries. What separates this survey from other AI polls is not the size of the sample but the number of axes. Gallup asked five things separately inside one questionnaire: whether people are aware of AI, how often they use it, what impact they expect, how much they trust it, and how it makes them feel.

1.1Three Conditions to State First

This article does not simply copy over the figures Gallup published; it counts them once more. Before that, three conditions of the survey itself. All three change how the numbers read later.

The first is funding. Gallup describes the work as "Gallup's ongoing research with Microsoft," and TechCrunch, writing about the same survey, said Microsoft commissioned it. Whichever wording you take, the same fact is left standing. A survey funded by a company that sells AI products is measuring AI adoption. This article handles the values on the basis of the sample design and the question-design principles Gallup published, and it does not hide that condition.

The second is the population. Roughly 1,000 adults aged 15 and older in each country were asked by telephone, face to face or online, with fieldwork running from April to July 2026. The margin of sampling error runs between ±2.2 and ±4.9 percentage points depending on the country. The Pew survey that comes up later asked adults 18 and older. That difference is the first thing to catch when values from the two surveys are set side by side.

The third matters most. Gallup asked its trust, emotion and expectation questions only of people who are aware of AI. This is not an inference drawn from the report text; it is a design statement Gallup set down three months earlier. In a post on Gallup's methodology blog, two researchers explained why the awareness question comes first.

So every Gallup percentage in this article is a share of the adults in that country who said they were aware of AI. In a country like the United States, where awareness runs above 98%, that is barely different from a share of all adults. Awareness is 12% in Malawi and 22% in Afghanistan. Set those countries' figures next to the American ones and the comparison does not hold unless the difference in denominators travels with them. In the same post Gallup also described the shape of its emotion question: six emotions asked one at a time, yes or no. Why that design matters comes back in Section 6.

1.2A Ladder With Four Rungs

Gallup groups contact with AI under the term adoption ladder. It is a set of steps running from being aware, to having used it at least once, to using it more often, and across all thirty-seven countries the number of people thins out toward the top. The median share aware of AI is 81%, the median share who have ever used it drops to 43%, and the share using it frequently is smaller still. The tables on trust and worry cut that ladder into four rungs: never used, less than weekly (called "occasional" here), weekly, and daily.

Gallup asked AI-aware adults how much they trust AI to provide accurate information. The question is aimed at accuracy alone, not at general warmth toward AI. Across the thirty-seven countries the median share saying they trust it completely or a lot is 36%, so trust itself is not yet abundant. A majority trusts it in only four countries: Israel (67%), Vietnam (59%), China (58%) and Nigeria (52%). Split that value by frequency of use, though, and the picture changes. Below is the table of the fifteen countries for which Gallup published all four rungs, with the higher rungs to the right.

Share (%) saying they trust AI completely or a lot to provide accurate information. The denominator in each country is adults aged 15 and older who said they are aware of AI. Values were extracted from the chart data posted on the Gallup release page; the rightmost column was calculated for this article. Bold orange marks the maximum in that country.

Gallup's own text compresses this table into one line: trust rises sharply along the adoption ladder. The United States shows immediately what that means. Among daily users, 45% say they trust its accuracy; among weekly users, 26%; among occasional users, 15%; among people who have never used AI, 8%. Bottom rung to top rung opens more than fivefold. Across all fifteen countries the distance between daily users and people who have never used AI averages 38.5 points, from 26 points in Russia at the narrowest to 68 points in Bulgaria at the widest. Not one country runs the other way.

1.3The Floor Sits on the Same Rung in All Fifteen Countries

Read the table across and you see the slope. Read it down and you see something firmer. Run an eye down the leftmost column, the rung for people who have never used AI, and that value is the minimum for its country in all fifteen. There is no exception. Thirteen of the fifteen also climb the four rungs without a single step out of order, and even the two that break do not break by much. Vietnam settles from 77% among weekly users to 76% among daily users, a single point, and Ireland has never-users and occasional users tied at 15%.

Why this table holds fifteen countries, Gallup does not say. Neither the release text nor the chart notes carry a selection rule. The seven wealthy Western countries Gallup groups together are all in it, but nothing in the table explains on what basis the other eight were added. So every count this article makes from here on is a share of the fifteen countries with published values, not of the thirty-seven. The Palestinian Territories row covers the West Bank and East Jerusalem; Gaza was not surveyed.

To put it plainly: there is little room to doubt that frequency of use and trust travel together. The slope did not come out of one country, and it did not come out faintly in a handful. It comes out of fifteen countries in the same shape and at close to the same strength. What this article does in the next section is not to unsettle that observation. It holds the same ruler against the other axis the same survey asked about one line over.

Gallup asked the same people about emotion as well. It asked separately whether AI makes them feel curious, happy, excited, worried, sad or angry, and worry came out as the most common negative emotion in thirty-five of the thirty-seven countries. The median share saying they feel worried is 32%. Gallup split this question by frequency of use too, and published values for the same fifteen countries. The third headline of the release puts the result this way: "Western countries rank high in AI worry, but daily users worry least."

The headline is not wrong. In twelve of the fifteen countries daily users report the lowest worry, and averaged across the fifteen, daily users are 10.3 points less worried than people who have never used AI. The direction pairs with trust. What does not pair is the shape. Set the table below alongside the trust table from the previous section and you can see two axes moving differently along one ladder.

2.1The Argument Is in the Last Two Columns

The layout repeats the previous section: the lower rungs on the left, the higher rungs to the right. Two columns have been added. One names the rung carrying the most worry in that country, and the other gives the distance between the top rung and the bottom rung. Those two columns carry most of what this article argues.

Share (%) saying AI makes them feel worried. As in the previous table, the denominator is adults aged 15 and older who said they are aware of AI. Values were extracted from the chart data posted on the Gallup release page; the two rightmost columns were calculated for this article. Orange marks the maximum in that country. In Malta the 63% among nonusers and the 62% among weekly users are one point apart.

In the trust table of the previous section, thirteen countries had a higher value on every rung than on the one before. Hold the same ruler against this table and the count is five. Only in the Netherlands, the United Kingdom, Bulgaria, Ukraine and Vietnam does worry come down rung by rung without once reversing. In the other ten it bends backward somewhere. Same survey, same respondents, same ladder, and one axis gives thirteen while the other gives five.

2.2The Distance Travelled Differs by More Than Threefold

The shape is not the only thing that differs; so is the distance. Climbing from the bottom rung to the top, trust gains an average of 38.5 points while worry sheds an average of 10.3. Divide and you get 3.7. Going the whole way up the ladder moves worry only about a quarter as far as it moves trust. The range of the gaps is lopsided as well. On the trust side the narrowest country is +26 points and no country reverses, while on the worry side the largest move is −23 points and the maximum value itself is +5 points. A positive number means there is a country where daily users worry more than people who never use AI.

And reading the table down once more turns up a place that does not pair with what the previous section showed. The minimum of trust sat on the never-used rung in all fifteen countries. The maximum of worry sits on that same rung in eight. The floor of trust gathers in one place, and the ceiling of worry is scattered. Of the seven scattered countries, six have occasional users in that position and one has weekly users.

2.3In One of the Fifteen, the Slope Runs Backward

The country recording +5 points is Singapore. There the least worried group is the people who have never used AI (45%), and the group carrying the most worry is weekly users (59%). Daily users sit in the middle at 50%. In the trust table of the previous section Singapore climbed from 26% to 57% exactly like the others, so the two axes part company further here than anywhere else. Singapore also sits on the top line of Gallup's adoption ladder: 96% of adults are aware of AI, 79% have used it, and 46% use it daily.

Calling one country an outlier and moving on is easy, but this table is all fifteen countries Gallup published. That the direction reverses in one of fifteen belongs, on its own, to what this section argues. Generalizing from one country is not in order either. The precise sentence is this: worry comes down along the ladder in most countries, and it cannot be said that the rule has no exception.

2.4A Survey of One Country Alone Produced the Same Split

The shape Gallup's fifteen-country table showed turned up again, at small scale, in an entirely different polling house looking at the United States alone. YouGov asked 1,500 American adults online on July 14, 2025, and 82% of avid AI users said they trust AI to some degree or more. Among AI users as a whole the figure is 69%. That trust tracks amount of use matches Gallup. In the same survey, though, the share saying they worry about accuracy came in at 63% among avid users, higher than the 59% among all users. Worry about cost runs the same way, 21% against 17%.

Whether that survey's "avid user" is the same rung as Gallup's "daily user" could not be confirmed, which means the two values cannot be subtracted and compared. What is usable here is direction, and the direction says this. If trust and worry were the two ends of one axis they would have to move in opposite ways, and when two surveys asked with different questions, they did not come out that way.

Back to the "rung with the most worry" column in the previous table. In eight countries that rung is the people who have never used AI. That is the common picture, and it fits the image of someone being most uneasy about a thing they have not gone through. In the other seven countries, though, the rung sits inside the ladder. Six of them are occasional users, and one, Singapore, is weekly users. Collected on their own, the values in those seven look like this.

The seven countries whose most worried rung falls somewhere among people who already use AI. Values are the share (%) saying AI makes them feel worried, with adults aged 15 and older who are aware of AI as the denominator. In the Czech Republic the 58% among occasional users and the 57% among nonusers are one point apart.

The United States is the sharpest case, and Gallup used it as its example too. Daily users report worry at 68%, less frequent users at 80%, and people who have never used AI at 74%. The figure Gallup gives as the American headline value is also 74%, measured across all AI-aware American adults, and it happens to land on the same spot as one of the four rungs. The number that travels into news coverage is usually that 74%. Read either way, that number is not the maximum in this table, and the maximum sits 6 points above it, on the rung for people who use AI occasionally. Canada opens wider still, from 54% among daily users to 72% among occasional users, a gap of 18 points.

That said, the United States sits at the edge of the overall picture this survey draws. Counting the six emotions one at a time, Gallup found positive emotions outweighing negative ones in thirty-four of the thirty-seven countries, and the United States is one of the three where they did not. The other two are Egypt and the Palestinian Territories. Across all thirty-seven countries the most commonly reported emotion is not worry but curiosity, at a median of 64%. This section leads with the United States because its value is the largest and because Gallup used it as the example. Nothing in this table gives grounds for carrying the American shape over to the other thirty-six countries.

3.1Four Explanations for Why the Middle Is the Most Uneasy

Looking for an explanation of this shape meant going through the risk-perception and technology-acceptance literature, and no prior study was found that reproduces an "occasional users peak" result in a general population. So the four below are candidates rather than verified explanations. This article has no grounds for picking one of them.

First, trying a thing a little shows concretely what is at stake. Worry among people who have never used AI comes from stories and news. People who have used it a few times have watched an answer come out wrong, and they have seen the places where it overlaps with their own work. Abstract unease turned concrete registers better in a survey response.

Second, they may be using it without any sense of being able to choose it or stop. What separates daily users from occasional users is not only taste. Someone who meets AI because it arrived inside a work tool lands on a middle rung when the measure is frequency alone, and that contact was not chosen. Contact without control leaves a different mark at the same number of encounters.

Third, the dullest explanation may be the strongest. If a particular age band, occupation or income bracket is clustered in the rung called "occasional users," then what we are looking at is not an effect of frequency but some other property of that group. Ruling that out takes a cross-tabulation carrying frequency and demographics together, and the freely published material does not have one.

Fourth, the size of the rung itself differs from country to country. What Gallup measured is a share within the AI-aware population, so if the "occasional users" rung is large in one country and small in another, the kind of person it holds changes too. Gallup's frequency table puts adults who use AI monthly or less at 23.5% in the United States and 12.7% in Ireland. A rung with the same name is holding close to twice as many people in one country as in another.

3.2A Similar Shape Has Been Reported in a Professional Group

Nothing turned up in a general population, but a narrow sample has produced a result running the same way. A national survey of dermatologists in India reported that doctors who use AI flagged concerns about misuse and overuse more often than doctors who do not, 72.9% against 54.2%, with the direction holding after adjustment for other variables. The authors named the result the adopter's paradox. General concern about algorithmic accuracy, on the other hand, was similar in the two groups.

That study cannot be read across to a general population. The sample is professional, and what they worry about is one particular job, clinical practice. One thing can be carried over. Using a thing and worrying about it are recorded as not pushing each other out in another field as well. What this section found in Gallup's table is an observation of the same kind.

What the previous two sections showed is a state of affairs, a map of who trusts how much and who worries how much right now. Look at that map and the next question follows on its own. Did using it often build the trust, or did the people who trusted it already end up using it often? On that question Gallup, instead of producing an answer, wrote down that it cannot answer. It is the last sentence of the paragraph on trust.

The same reservation repeats in the report's bottom line. The very sentence noting that frequent users are generally more trusting and less worried carries, at its end, the qualifier that the survey cannot determine which comes first. A survey that measures many people at one moment is structurally unable to separate the two. Either the same people have to be followed over time, or the researcher has to decide who gets to use the thing. The Pebblous blog has worked through this point once before in a piece on estimating causal effects from virtual interventions, so what follows stays with this survey and this table.

4.1No Study With a Design That Answers It Could Be Found

If the survey cannot separate them, the next thing is to look for a study that has. The design required is clear. Randomize who gets to use AI, put attitude toward the technology and unease about it on the outcome side, and follow them over time. No such study was found in the published literature. That does not mean none exists; it means none came up in the range this article searched. The three that came closest each fall short by one piece.

Answering "does using it reduce worry" requires separating self-selection from treatment effect. Each of the three designs above fails to meet one of those conditions.

So the reservation Gallup wrote down is not a circumstance of this one survey. It is closer to a circumstance of the whole question. Observations of use and attitude moving together come out of many places, and the ground for rewriting those observations as "trying it changed people" is still thin.

4.2If Attitude Were One Axis, the Two Tables Could Not Differ

Setting aside the direction of causation, the table in Section 2 is already saying one thing. If trust and worry were the two ends of one ruler, the two tables would have to come out as mirror images, and they did not. A remark Gallup's senior scientist Pablo Diego-Rosell made to TechCrunch lands on exactly that spot.

Measurement instruments in the research literature assume the same structure. One AI attitude scale validated in Germany, China and the United Kingdom treats attitude toward AI as two factors, acceptance and fear, and that structure came out similarly in all three countries. The two factors correlate negatively but remain separate axes, and trust is one item on the acceptance side. How independent the two axes actually are, and what the correlation coefficient is, this article could not confirm, so no figure for it is given here.

Another organization's survey put nearly the same thing in nearly the same sentence. The 2026 edition of the AI poll Ipsos has run across thirty-two countries for five years asks the same person about both anticipation and nervousness. The report's line reads: "In many cases, it's the same people feeling both emotions." Among people under 35, nervousness came in at 52% and anticipation at 56%, the highest on both sides. If two feelings grow together inside one person, a metric with one axis copies that person down only halfway.

There is only one way to separate what a single-moment survey cannot. Ask again, in the same words, over several years. The United States has material of that kind. Pew Research Center has been asking, with the same question wording since 2021, how much people have heard about AI and how they feel about AI spreading through daily life. All six readings appear in the topline of the thirty-seven-country report published on September 17, 2026, the appendix that carries the question wording and the response distributions as they stand.

The phrase "the same question" carries only as much weight as was checked. Pew attaches a footnote and breaks the trend line whenever question wording or answer options change, and the two questions in this topline carry no such footnote. The same six values appear again on a separate trend chart page Pew published. What could not be done was opening each topline from 2021 through 2025 and comparing the wording letter by letter.

Lay the values of the two questions on one chart and the title of this section becomes visible at a glance. The orange line above is "heard a lot about AI," the dark line in the middle is "more concerned than excited," and the pale line below is "more excited than concerned."

5.1The Value That Stood Still While the Other Doubled

The orange line climbs with barely a pause. Americans who have heard a lot about AI went from 26% in December 2022 to 48% in February 2026. Over the same stretch, those saying they have heard nothing at all fell from 15% to 3%. Measure exposure as awareness and the American figure roughly doubled in a little over three years.

The dark line moves differently. Those saying worry outweighs anticipation stood at 37% in November 2021 and 38% in December 2022, nearly the same, then jumped once to 52% in July 2023. After that it reads 51, 50, 52. Four years now in the same band. Pew, writing about the past year, said views in the U.S. have been relatively stable, and stretching the window out in both directions does not change that sentence.

The pale line bent at the same moment. Those saying anticipation outweighs worry fell from 18% in 2021 to 10% in July 2023, then stuck there at 11, 10, 9. So the stretch in which American attitudes actually moved runs from late 2022 to the summer of 2023, about half a year, and all three rungs have been still since.

The part of this picture that matters most is what happened after attitudes stopped. Awareness climbed another 15 points, from 33% in July 2023 to 48% in February 2026. The stretch in which exposure kept growing is exactly the stretch in which attitude was frozen. Put another way, across the past three years, as more Americans came into contact with AI, there is no trace of worry going down.

A qualifier Pew attached to its own data belongs right next to this. Pew observed in several countries that people with high awareness answer the emotion question more optimistically, and it used Germany as the example. Among Germans who have heard a lot about AI, 26% say worry outweighs anticipation; among Germans who have heard a little or nothing at all, 41% do. Inside that same paragraph, in parentheses, it says this: "In many countries surveyed, people who are less aware of AI are also less likely to provide a response." The job-loss question carries the same qualifier. Less awareness may mean not less worry but no answer, and the accompanying explanation is that "not sure" clusters among respondents with lower income and less education.

5.2Three Other Surveys Did Not Go Down Either

Building this section on one organization's material would tie the argument to that organization's questions. Three more surveys of the United States and the world over the same span go alongside it. The values in the table below must not be added to or subtracted from one another. The questions differ, the populations differ, the dates differ. Direction is the only thing that can be set side by side.

The question wording and the population differ across all four surveys, so the values cannot be bound into a single number. The two moments in the KPMG and University of Melbourne study cover different countries: the 2022 survey asked seventeen countries and the 2025 survey asked forty-seven, so it is not a time series measuring the same countries repeatedly.

The sentence in the report KPMG and the University of Melbourne produced from more than 48,000 people across forty-seven countries comes close to rewriting this article's argument in another hand. Compared with the earlier survey of seventeen countries run in 2022, before ChatGPT was released, it says people have become less trusting and more worried about AI as adoption has increased. That comparison runs between two moments covering different countries, so the reading is not a strict time series, and dropping the condition would make this article break in Section 7 the very rule it is about to state.

Nor do the four surveys all point one way. Pew's United States is flat, while Quinnipiac and Bentley–Gallup got worse. So the sentence "AI anxiety is spreading" is not something this material holds up. One line is held up: no survey yet shows worry going down over a stretch in which use went up.

This blog has written the sentence "worry and use grow at the same time" before, when it looked at the gap between AI exposure and bargaining power. The 52% that piece cited is what American workers answered in an October 2024 survey about AI use in the workplace, and the 52% in this section is what American adults answered in a June 2026 survey about AI spreading through daily life. Same number, different thing measured. What that piece put in one sentence is what this section has laid out along a time axis.

The two reports came out in the same season, both asked thirty-seven countries, and both deal with public opinion on AI. That makes them easy to quote side by side inside one story. The moment they are set side by side, though, four places start to blur. In all four the numbers are exact, and what differs is what each number counted.

All four are values each organization recorded accurately in its own report. What fails to line up is not the values but what each of them counted.

6.1Measure the Same Country Twice and the Answers Differ

The third row opens widest. American "AI worry" is 74% on one side and 52% on the other. A 22-point difference is not a question of which organization is more accurate. Gallup asked AI-aware adults whether AI makes them feel worried, yes or no. Pew asked all adults to pick which is larger, anticipation or worry, and gave them a middle box for equally concerned and excited. In the United States, 37% chose that middle box. One asked whether worry is present, and the other asked whether worry is larger than anticipation.

Nor is it a difference produced by different polling houses. One of the organizations that directed Pew's international fieldwork is Gallup. The American data, though, was collected separately on Pew's own panel, so that logic cannot be extended to the American comparison. Which countries Gallup handled could not be confirmed country by country from the published methodology documents.

Worry is not the only place where names overlap. So is "daily user," a term used here since Section 1. In Gallup's frequency table, 22.8% of American adults said they use AI daily. The figure Pew put to American adults in February 2026 is 24% if narrowed to chatbots, and roughly 38% if widened to using AI in any form several times a day. Those three values cannot be subtracted from or weighed against one another. Gallup asked people 15 and older while Pew asked people 18 and older, and one asked about AI as a single block while the other split chatbots from other AI. "Daily" and "several times a day" are not the same threshold either.

So "daily" in this article's tables refers to the rung Gallup drew. That is not a number to set beside a daily-use rate seen in some other story, and the first thing to check is whether that story's daily and this table's daily hold the same people.

What happens when one name points at several things is something this blog touched on in a piece about evaluation metrics that share a name while measuring different things. What this section is saying is not that. Neither of the two surveys attached a wrong name to anything. They were built to measure different things to begin with, and they only look like the same name at the moment they are placed side by side.

6.2The Mismatch Showed Because Several Axes Were Kept

Retrace Sections 1 through 3 and there was a condition that made it possible. Gallup measured awareness, frequency of use, expectation, trust and emotion separately inside one survey. Had those five been combined into a single "AI acceptance" score, the table in Section 2 could not exist. The observation that worry does not come down as far as trust goes up holds only while the two axes are each still there. In a summed score that mismatch cancels itself out and disappears.

Keeping several axes is not itself rare. The KPMG and University of Melbourne survey measures frequency of use, willingness to trust, AI training received and worry at once across forty-seven countries. Three in five respondents in emerging economies trust AI systems while two in five do in advanced economies, self-reported AI knowledge runs 64% against 46%, and training received runs 50% against 32%. Ipsos has spent five years asking one respondent to rate both anticipation and nervousness. Surveys that keep several axes do exist.

The rare thing is elsewhere. Even those surveys do not hold the same ruler to the same subject repeatedly. KPMG's study is in its fourth round, and the countries covered changed with every round, from five to seventeen to forty-seven. Gallup's survey here is the first wave of a project heading toward 140 countries, so there is no previous round to compare against. What allowed Pew to say "four years in the same place" in Section 5 was not a large sample but six askings in the same words since 2021. Standing several axes up and holding those axes unchanged for years are two different jobs.

6.3Three Things to Take From These Tables

What fails to line up when a number from one survey is set beside another is something this blog went through once, back when it took up opposition to data centers. What that piece looked at was eight surveys asking the same thing in different ways, and what this one looked at is two surveys that measured different things to begin with. A piece on AI that employees brought in themselves, which put the same question inside the workplace, belongs alongside them. Following the route this article took, three things come first when two surveys give different pictures of the same subject. One, who was taken as the denominator. Whether it is people aware of AI or all adults makes the same question produce entirely different values in countries with low awareness. Two, were the emotions asked one at a time, or traded off against each other. A question asking about six emotions separately lets one person say yes to both curiosity and worry, and a question making people pick between anticipation and worry does not. Three, was it measured once, or asked again in the same words. The first two are ways of reading today's value, and only the third answers whether anything changed.

The article that seeded this one closed on the line that greater exposure alone will not resolve the unease around the technology. This article does not dispute it. What it did was count how far that holds. Trust climbs on every rung in thirteen of the fifteen countries, and worry comes down that way in five. In seven countries the heaviest worry falls on a rung inside the user population. On the time axis, worry stayed between 50 and 52 while awareness climbed another 15 points. Those four lines are everything this article added.

Footnote. The summary text of the Pew report says Israel is the only nation where more people are primarily excited than primarily concerned, while the topline table in the same report puts several countries, South Korea among them, on that side, and the report text itself notes South Korea's 21% five lines later. Which one to take is left to the reader. South Korea's position is handled as one item in the questions below.

The problem this article has followed does not live only in opinion polling. It has the same shape as a problem Pebblous meets again and again in data quality work. The three passages below are not about a product. They are about how the structure the earlier sections showed turns up again in practice.

7.1Reading a Cross-Sectional Slope as Evidence of Improvement

What Section 4 showed is the place where Gallup stopped itself. The step from "daily users worry less" to "using it reduces worry" is one the survey itself said it cannot take. In front of a dashboard, that step gets taken almost automatically. A usage graph going up reads as settling in.

The same step comes up in data quality work. Going from the observation that a dataset run through a cleaning pipeline has a high quality score to the conclusion that the pipeline raised that score takes one more thing. It takes ruling out the possibility that data already in good shape was what got routed through that pipeline. No amount of single-moment material separates those two. Measuring the same objects before and after processing, or taking the choice of what gets processed out of human hands, separates them. What random assignment does in a survey, those two devices do here.

7.2What Answers "Did It Change" Is a Measurement Program, Not a Dataset

What allowed Pew to say in Section 5 that American worry has been in the same place for four years was not the size of the sample. It was having asked in the same words six times since 2021. Gallup's survey carries thirty-seven countries and a small margin of sampling error, but as a first wave it has no time axis yet. So it answers "how much is it now" and cannot answer "did it change." The difference between the two surveys lies not in the size of the material but in whether there is a record of measuring the same subject the same way again.

Data quality metrics work the same way. Measure once and a grade comes out, and with that grade the present state can be described. Whether it improved can only be said once the same schema and the same checks have been laid against the same objects again and that history has been kept. If lineage records where a value came from, repeated measurement records what that value was and when. They are different layers, and the second one usually goes missing first. Putting a first score on something gets commissioned as a project, and holding the same ruler for several years rarely does.

7.3Axes Can Be Added, but a Record Does Not Appear in a Year

To an organization that has to report on AI adoption, this article hands back two questions. How many items are we measuring, and how many of those did we measure last year in the same words? What let Gallup see the mismatch between trust and worry was measuring five axes separately in one survey. Had the axes been collapsed into one, that mismatch would not have existed in the material at all.

And as Section 6 showed, keeping several axes is less rare than it sounds. What is rare is not the number of axes but the record. The forty-seven-country survey named in that section has covered a different set of countries in each round, which limits comparison across years, and Gallup's survey here has no earlier round to be set against at all.

What Pebblous can put forward here is not much. It is having worked on measurement designs that apply the same standard repeatedly rather than on scores taken once and finished. As a small example, articles on this blog are published with a record of the generation pipeline attached. Which stage was handled by which model, and which checks it passed, go into the page configuration. The reason is that a reader should be able to go back over the figures written here.

The figures in this article come from three strands. Items 1 through 7 are the primary material this article is built on, read directly: the Gallup release and the chart data posted on its page, Gallup's methodology post, and the Pew report with its full topline. Items 8 through 12 are material from other polling houses used in the time-axis comparison in Section 5, and items 13 onward are academic literature on measuring attitudes toward AI. Where an item was available only at summary fidelity, the article cited direction alone.

Primary survey material

• 1.Gallup (2026-09-22). "AI Optimism Globally Widespread Despite Uneven Use." 37 countries, about 1,000 respondents each, aged 15 and older, fieldwork 2026-04 to 2026-07. news.gallup.com
• 2.Gallup. Embedded chart data posted on the release page above — frequency of use across 37 countries, the 15-country trust cross-tabulation, the 15-country worry cross-tabulation, and the emotion question. The tables in Sections 1, 2 and 3 were extracted from this data.
• 3.Lau, C., & Nugroho, A. (2026-06-24). "How Gallup Developed Global Survey Questions About AI." Gallup Methodology Blog. Gallup's own account of the awareness screener and the parallel emotion questions. news.gallup.com
• 4.Pew Research Center (2026-09-17). "Globally, More People Expect AI to Cause Job Loss Than Growth." 42,151 people across 36 countries, fieldwork 2026-02-08 to 2026-05-13, with the United States surveyed twice separately. pewresearch.org
• 5.Pew Research Center. Topline PDF for the report above — question wording for Q54 to Q57, country tables, and the six-year American time series. The values in Section 5 were carried over from this document.
• 6.Pew Research Center. International survey methodology. International fieldwork was directed by Gallup, Langer Research Associates and the Social Research Centre, and the American data was collected on the American Trends Panel.
• 7.TechCrunch (2026-09-23). "Even Americans who use AI every day are worried about it." The Diego-Rosell quotation and the description of the survey design were carried over from here. techcrunch.com

Other surveys used for cross-checking

• 8.Gillespie, N., Lockey, S., Ward, T., Macdade, A., & Hassed, G. (2025). Trust, Attitudes and Use of Artificial Intelligence: A Global Study 2025. University of Melbourne and KPMG. More than 48,000 people across 47 countries, fieldwork 2024-11 to 2025-01. DOI 10.26188/28822919
• 9.Ipsos (2026). AI Monitor 2026. 23,532 people across 32 countries, fieldwork 2026-03-20 to 2026-04-03, Global Advisor online panel.
• 10.Quinnipiac University Poll (2026-03-30). 1,397 American adults, fieldwork 2026-03-19 to 2026-03-23, live-caller random digit dialing, ±3.3 percentage points.
• 11.Ray, J. (2026-07-28). "Americans Cool Toward AI." Bentley University–Gallup. 3,270 American adults, fieldwork 2026-05-04 to 2026-05-11, Gallup Panel web, ±2.4 percentage points.
• 12.YouGov (2025-07-14). "Most Americans use AI for quick answers, but avid users trust it more." 1,500 American adults, online. This article cited direction only.

Academic literature

• 13.Sindermann, C., Sha, P., et al. (2021). "Assessing the Attitude Towards Artificial Intelligence: Introduction of a Short Measure in German, Chinese, and English Language." KI — Künstliche Intelligenz. The two-factor structure of acceptance and fear.
• 14.Schepman, A., & Rodway, P. General Attitudes towards Artificial Intelligence Scale (GAAIS). Another scale reporting a two-factor structure.
• 15."The Adopter's Paradox" — national survey of dermatologists in India. 72.9% of doctors using AI flagged misuse and overuse, against 54.2% of doctors who do not. arXiv:2607.01252. A professional sample.
• 16.Fang, C. M., et al. "How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal Randomized Controlled Study." arXiv:2503.17473. This is the first row of the table in Section 4.

Related Pebblous articles

• 17.Pebblous, "The Most AI-Exposed Workers Have No Seat at the Table." report/ai-exposure-union-gap-2026
• 18.Pebblous, "IC² estimates intervention effects from observation alone." report/ic2-virtual-intervention-causality
• 19.Pebblous, "Employees Bring Their Own AI, and Employers Still Have No Rules." blog/shadow-ai-konstanz-workplace-survey
• 20.Pebblous, "Did Data Center Opposition Grow, or Did the Question Change?" report/datacenter-opposition-reasons-pa-study
• 21.Pebblous, "The bias check AI labs report ranks models like a reasoning test." report/benchmark-concept-validity-2026-09
• 22.Pebblous, "Can You Trust a Company's Own Answer That Its Data Is Ready?" blog/data-readiness-self-report