2026 was the year AI moved from a few new apps to the most important driver of the world economy. The revenues of leading AI labs Anthropic and OpenAI increased by more than 10x year on year. Data centre construction has become a key, if not the, headwind for US GDP. Economists continue to debate back and forth whether AI automation is already having an impact on entry level jobs or not.
Where will it go from here? How do the public feel about it? And just why are we starting to see a backlash against AI among some of those who should be most enthusiastic?
In order to keep an eye on how AI is developing, we:
While AI could be a massive boost to the world economy over the next few years, we found that this is not inevitable. Much of this remains within our control, and dependent on what choices we make to change workflows around it. In our modelling, we found that around half (47%) of the potential economic opportunity in the next five years is dependent on how prepared we are to take difficult choices to integrate AI going forward.
Taking full advantage of AI will require difficult decisions - but in the last year we have seen a distinct fall in global sentiment towards AI. This is particularly concentrated in Western economies like the US and UK, and in some areas, among the young. While sentiment is more positive in Asian and emerging markets, there are some signs that sentiment might be declining too.
As AI becomes more powerful and the general public continue to be ambivalent about the implications of more advanced AGI or superintelligent AI systems, we see increasing readiness to consider more drastic solutions like an AI pause - albeit with doubts over whether it would be effective in practice.
Here are 15 of the most interesting things we found:
AI is the most important technology of our time.
On average, 45% of respondents believed that AI will change the world significantly - and 46% that it already had.
How big an opportunity AI is, is largely under our control.
Even if the technology never changes from here, differences in how fast we adopt and how prepared we are to use AI in more sensitive areas could drive huge differences - up to $14 trillion in economic value by 2035.
Tasks aren’t occupations.
For some jobs AI will act as a force for deskilling; for others, it will be purely complementary.
In the West we are seeing a real AI backlash - but other countries remain highly optimistic.
Over the last two years, net valence for AI in the US and UK has fallen by around 75%
It's still early.
By calibrating from how much workers in our survey were actually using AI to augment their tasks, we can get a better estimate of how far along we are - we estimate in 2026 we have seen only 8% of the potential economic opportunity.
We are beginning to see some evidence of an AI divide in the sophistication of use of tools.
In our survey, we saw that AI usage is increasingly broad, but often shallow. Workers at larger enterprises were around 1.1 times more likely to be a sophisticated user than those at SMBs.
Most people are still uncomfortable sharing sensitive data with AI agents.
Just 29% of people said they would be comfortable giving an AI agent access to their bank account.
Younger workers have the highest levels of worry over AI automation.
By contrast, when we looked at who was most likely to be worried about AI automating their job, we saw little correlation with income, sector or our model's estimate of overall likelihood of automatability.
Respondents increasingly feel that where your AI model come from matters.
And not everybody trusts the US - with 19% of non Americans in our survey saying they didn't trust an American model.
There are very few real luddites, in practice
. While people globally say they want Governments to prioritise jobs over the economic impact of AI, there are no previous labour saving technologies they would have actually slowed down in practice.
Americans are by far the most negative country when it comes to data centres.
Every other advanced economy was clearly net positive on building more data centres, whereas the population in the US was practically evenly split.
For good or for bad, relationships with AI are no longer science fiction.
1 in 5 said they have used an AI for emotional advice, and 1 in 6 that they have a relationship with an AI.
What risk matters most depends how you ask.
We saw concerns shared widely across all the main types of AI model, and statistically it seemed to be a single factor driving AI support or opposition.
Loss of control from an advanced AI continues to be a significant worry.
In our polling we saw that relatively few respondents were aware that lab leaders shared worried about AI risk: just 35% correctly know that AI lab leaders consider superintelligence a serious risk, and under 20% know they've said so publicly.
There are mixed views about the feasibility of an AI pause.
While 56% were in principle in favour of a pause, only around half of those supporters thought one was realistic in practice.
The scale of the AI prize — and how much of it is in our hands.
Across our panel, for good or bad, we saw widespread agreement that AI was an important technology - with few seeing it as a con or a fad.
When asked directly, only 4–13% of respondents in any country think AI is "just a phase that will pass. On average, 48% of respondents believed that AI will change the world significantly - and 43% that it already had. 64% agreed that AI was the most important current technology, and that their country needed to be a leader in it.
For the last four years, we've asked respondents whether they thought AI was developing slower or faster than they expected. Overall, we saw that respondents continued to see AI as developing fast, although relatively speaking this has plateaued off a little over time as respondents get used to AI.
That said, only around 8% of our survey were completely AI pilled, seeing it as the most important global issue of the day. On average, AI was ranked fourth: coming behind inflation or the cost of living, war and climate change.
2026 was the year AI moved from a few new app to the most important driver of the world economy:
Where will it go from here?
There are many unknowns about AI’s potential impact:
There are many different ways you could combine assumptions on these axes. In order to try and pull them apart, we built a new global model of AI’s impact, building on top of a standard task based model to see what role they might play - and what impact this will have for different countries, sectors and occupations.
Alongside the labour productivity impact of AI, we also incorporated the wider growth impacts of AI: boosting wider innovation, human capital and improving institutions. In our central scenario, these wider impacts are potentially as large again as the direct labour productivity impact.
However, achieving AI’s full economic potential is not inevitable, and does not necessarily happen by default. Much of this remains within our control, and dependent on what choices we make to change workflows around it. In our modelling, we found that around half (47%) of the potential economic opportunity in the next five years is dependent on how prepared we are to take difficult choices to integrate AI going forward.
You can explore different scenarios, or play with some of your own assumptions, in our model here.
One of the key variables in our model is sensitivity: even where AI could technically automate a task, are there cultural, political or ethical reasons we would not want it to? We classified every task on this kind of sensitivity, and applied a discount to the most sensitive ones.
As a sense-check on that classification, we also asked the public how comfortable they were with automating a range of different occupations. For knowledge and clerical work, our scores and the public's comfort lined up closely: people were highly wary of automating sensitive roles like a judge, but far more relaxed about clerical tasks such as translation.1
Sensitivity has a large effect on the headline. The more we try to keep AI out of every occupation where there might be concerns, the smaller the economic prize becomes — we find that a 10-point move in our sensitivity rating swings our estimate of the total global impact by around $1 trillion.
1 The public also did not want physical tasks automated — though here the main reason seems to be that they do not yet trust AI to do those jobs at all, rather than that the work is especially sensitive.
Like most task-based models of AI's impact, our headline model looks at AI's effect on individual tasks, then works out the average impact for each occupation. We were also interested, though, in how the tasks within an occupation interact — and how much it matters which tasks are most affected. That tells us how far AI is likely to be a complementary rather than a substitute force for individual workers.
To explore this we built a new toy model, looking at how much AI affects the tasks within an occupation that create the most economic value, versus the less core ones. This divides jobs into four groups:
That last group is a particularly interesting one. If job roles don't change — if we keep doing occupations the way we always have — it is where the threat of de-skilling lies: even if overall employment holds up, wages could fall. But if we can evolve how occupations are designed, the same group could be where some of the most broadly beneficial outcomes from AI show up.
What over 18,000 people in 15 countries told us about how AI shows up in their lives.
As part of our survey, we ask our respondents what emotions they associate with AI. By looking at the balance between positive and negative emotions - 'net valence' - we can see how people feel about AI overall.
In our data, we see a clear drop for the US and UK, with associated feelings around AI growing clearly more negative. Net valence in the US has fallen from +0.73 to +0.18.
Looking by age, we can see negative feelings are particularly concentrated in two groups: the under 25s and over 65 years olds.
But this backlash isn't true across the world. In emerging economies and Asian countries like Singapore, we see attitudes to AI remain significantly more positive.
This is even more clear when you look across a range of issues. Compared the US to Singapore across six life domains — personal, career, the people you know, the economy, society, the next generation — Singapore sees AI as a positive force, while the US is much more uncertain.
Across advanced economies, we now see around half of adults say that that they are using AI regularly.
Regular AI usage has roughly doubled over the last two waves of our survey in the US and UK - and on current trends looks set to hit full market maturity by around 2030.
Much of this growth is recent. Across advanced economies, 56% of AI users told us they had only started using AI in the last year. Extrapolating from the trend across multiple waves of our US and UK data, we should see the vast majority of the population having adopted AI within the next few years.
Given AI's potential for augmentation however, regular usage by itself is a relatively shallow indicator. When we ask users roughly what proportion of their daily tasks they are using AI for you see a much lower lever, at around 25%.
Integrating this polling data with our economic model gives us a better idea of how much economic growth we should expect to have seen by now. Our model estimates AI could ultimately add around $7.3 trillion in labour-productivity value worldwide; comparing the AI use workers actually reported against the tasks it expects AI to augment, we find we are still early — with roughly 15% of the 2035 labour-productivity opportunity in place so far, or around 8% of the total dividend once wider impacts are included.2 That is set to change quickly, however: if AI follows a normal S-curve diffusion path, over half of the opportunity lands by 2030.
2 This is one reason it can be hard to distinguish the effect of AI in the economic statistics. National accounts are imperfect, frequently revised, and noisy with the business cycle and other macro trends. As a rough rule of thumb, until AI is adding a full percentage point to labour-productivity growth for a few years running, it may be hard to unambiguously pick it out of the official numbers.
In our data, we find that how you use AI tools matters as much as how often you use them - with the more sophisticated users reporting significantly higher benefits. Techniques like GPTs or skills, agents, or different prompting strategy can all make a real difference.
Across our sample, we saw some signs of a divide in AI sophistication:
The divide isn't just about who uses AI — it's about how they feel about it. Pooling across all 15 countries, super-users are strongly positive on AI, while non-users are net-negative. Sophistication and optimism move together.
One driver of these differences was how supported users felt at work, or their workforce enablement: how much did they have access to the right tools, and how clear was it how they were or weren't allowed to use them?
In our data, we saw a significant difference across both countries and size of business in the level of workforce enablement.
If two terms dominated the AI discourse in 2025, it was vibe coding and agents.
While vibe coding tools have seen rapid take up in development communities, in our survey data we saw that it was still early days for the general population: across advanced economies less than 10% of the general population said that they were using them regularly. Out of those who were using them, the most common use cases were for building personal web apps or analysing data.
Beyond coding, agent based work flows can help people across their daily tasks: but are likely to be most powerful when they have access to your personal context. In practical terms, that could mean everything from your bank account to your health data. How comfortable are people with sharing that kind of sensitive data?
In our data, we actually saw on average respondents were comfortable with sharing their health data and their calendar - but were much less confident with anything with potential financial implications.
Some of this however may be just less experience with the tools. In our data, we saw that the more frequent or sophisticated an AI user was, the more prepared they were to share their data.
Jobs, sovereignty, data centres, intimacy.
In our data, we saw widespread agreement that AIs would be able to helpful with their jobs in the next few decades, with 80% believing that this was likely.
By contrast, we saw much less conviction that AIs would be able to take over the job entirely: on a scale from 0 (definitely can't do my job) to 10 (definitely can do my job), the average response was 5.7.
Across all the countries we asked in, we saw the same hierarchy of occupations which people thought acceptable to automate: translators, IT support, software developers, customer service agents are the most-acceptable-to-automate. Nurses, childcare workers, hairdressers, and surgeons are the most protected.
Interestingly, when we looked at who was most likely to be worried about AI automating their job, we saw little correlation with income, sector or our model's estimate of overall likelihood of automatability. The pattern of worry by age also looks very different in advanced vs emerging economies. In advanced economies, worry falls monotonically with age — the young worry most, the old worry least. In emerging economies the curve is flatter and worry stays high right through mid-career, only dropping past 65.
When asked a forced choice question over whether governments should prioritise protecting jobs or maximising the economic benefits of AI, we saw strong support for the former across countries. On average, in advanced economies 68% chose protecting jobs compared to 32% picking wider economic benefits.
How strong this belief is in reality, is less clear. When we asked about previous labour saving technologies - from the self-service checkout to the original luddite cause of mechanised textiles - only a minority believed that the Government should have slowed any of them.
With the exception of Japan and the US itself, across the rest of our countries we saw a clear belief that China is now ahead of the US on AI capability and innovaton.
By contrast, we saw a much more mixed picture when it came to which countries people trusted models from. Across the board, people prefer models from their own home country most - and then after Japan was overall most trusted, with the US slightly behind.
This support however masked significant disagreement with significant minorities saying they both trusted and distrusted models from the US and China.
One practical implication of this is that we saw overwhelming supporting for some version of data sovereignty, with 77% saying it was important to keep data within the country.
That said, most countries also recognised they couldn't build the whole AI tech stack by themselves - just 57% thought they should build their own AI chips - and even among the 19% who said they distrusted models from the US, in practice 56% of them were still using one of ChatGPT, Claude or Gemini.
Every other advanced economy was clearly net positive on building more data centres, whereas the population in the US was practically evenly split. Across the board, we saw that Americans were less likely to think data centres were beneficial
Looking across the full five-statement battery — local jobs, growth, business competitiveness, personal data protection, and tech leadership — the US dot consistently sits to the left of every advanced-economy comparator.
The strongest stated arguments against data centres were their perceived impact on electricity bills, taxpayer costs and environmental damage.
To what extent these environmental concerns are the true cause of concern in the US are less clear: those who worried about environmental impact were no more likely to rate environmental concerns seriously in other contexts we saw a significant minority (35%) who incorrectly thought data centres were likely to put water pressure on their local environment
AIs might be tools, but almost nobody treats them as a dishwasher. Across countries, we found that just under half (43%) think its important to be polite to an AI.
Across the board, we saw a significant minority were turning to AIs for something closer to the kind of relationship you might have with a human:
For the majority of people, these kind of relationships can be harmless. For a vulnerable minority however, we already seeing real world examples of concrete harms created by unhealthy paraosocial relationships with AI, including delusion, self harm and paranoia.
In our polling, we saw these worries were particularly charged when it came to children talking to AIs unsupervised. In our survey, 77% agree unhealthy AI–child relationships could repeat the harms of social media, and 62% say the risks of children talking to AI agents outweigh the benefits even with safeguards.
AGI, loss of control, and the case for slowing down — what to do with what we know.
What risks from AI do people worry about most? Is it short term more concrete risks like deepfakes or misinformation, or broader trends like unemployment or existential risk? Does it differ by demographic or is it fairly aligned?
In our previous research, we've found that which risks get the most prominence depend a lot on how questions are framed and the format they come in. For this wave of research then, we asked about risks and worries in multiple different ways to let us triangulate what we could find.
Overall, we found that the majority of AI concerns comes from a single pro or anti AI sentiment. Using a statistical segmentation, we found that there was largely only a single factor driving attitudes - we didn't detect one group that was more worried about the environment and then another jobs, for example.
When asked to rank their level of worry across a long list of risks, it is the more prosaic risks that come highest: deepfakes, misinformation and cybersecurity, with employment a little behind. That order was relatively consistent across countries.
By contrast, when we let people answer in their own words — naming the single biggest danger from AI — existential and long-term risks came much more to the fore. Among advanced economies, "an AI taking control" was the most common unprompted answer of all, even though the same risk sits well down the list when people are handed a fixed menu. This skew toward existential risk was a little stronger in advanced than in emerging economies, but held across the board.
Finally, when asked to pick the single risk they thought governments should prioritise, they chose An advanced AI being misused by criminals, terrorists or hostile governments (26%), AI enabling misinformation, deepfakes or badly affecting vulnerable users (22%), and AI causing unemployment or the loss of jobs (21%).
While the framing changes which risks rise to the top, what made far less difference was the respondent's country: asked on a like-for-like basis, the ranking of risks looked broadly similar everywhere.
Over the last year, talking about AGI, superintelligence and even the singularity has become steadily less taboo among the tech industry.
In our survey, 61% agreed that AGI - an AI at least as capable as a human - was physically possible, and 32% expected it to arrive before 2030. On average, the median expected year for AGI to arrive was around 2032.
On average, most people in our survey were wary about the implications of AGI, with a net score of +6% on good vs bad for the world, and -15% on safe vs dangerous. A superintelligent AI was seen as even more dangerous, with 50% on average believing it would be dangerous.
One key reason for worry about advanced AI was that only 29% think a superintelligent AI would be "aligned with human interests by default", while 52% think it would seek to take control away from humanity
The general public are not alone in worrying about the risks from advanced AI. In previous years, we've seen AI lab leaders including Demis Hassabis, Dario Amodei, Sam Altman and Elon Musk all warn publicly that AGI could be dangerous if not managed correctly.
In our polling we saw that relatively few respondents were aware that lab leaders shared these concerns: just 35% correctly know that AI lab leaders consider superintelligence a serious risk, and under 20% know they've said so publicly.
Given the potential risks - and benefits - from continuing to develop ever more advanced AI, how should the world respond?
On the face of it, in our data, we saw clear support from a majority in nearly every country for a pause on AI development, with 56% agreeing.
In practice, however, just 56% of those thought a slowdown was realistic - while enough significant proportion dropped support when framed against the possibility of falling behind China in technological supremacy.
This position was surprisingly non partisan. In the US and UK, [we saw relatively little divide in support or opposition to a pause based on politics.]
A less drastic intervention than an outright pause would just be implementing co-ordination. Again, we saw initially strong support for this across the board - but this dropped away by a significant proportion if it was seen as unhelpful to their own country, with a majority in the US preferring to maintain domestic control.