A richer economy won’t be enough to win trust in AI
Harvard professor Stefanie Stantcheva examines how people judge technological change: through their own experience, their expectations of fairness and what happens to those who lose out.
In tech circles today, AI is pitched as a technology of abundance. It will improve our knowledge and science, save lives through new treatments, increase productivity, make daily life easier and much more. At the same time, in the media and many policy discussions, the technology sounds like one of scarcity: an imminent job apocalypse, inequality, a loss of learning and education as we outsource our thinking to machines.
The promise of abundance has some clear backing, but so does the possibility of the gains and losses falling on very different groups of people. Aggregate abundance and individual scarcity can coexist.
Many people still do not use AI regularly at work. Yet the citizen who has never prompt-engineered or run agents on their Mac mini will also help shape AI’s political future. They will vote for the governments that regulate it and set policies to either foster or hinder the technology. They will respond to its use in schools and in the workplace. Usage statistics alone cannot tell us what these citizens expect from AI, or which policies they will support.
Currently, a lot of the discussions around AI center on the supply side: What will the technology do? How is adoption progressing? We economists also forecast possible job losses, estimate which occupations will be most exposed and try to predict what will happen to wages and productivity. But that still leaves key questions unanswered: What do people believe AI will do to them, and what do they want policy and society to do with it?
Conversations at the frontier, in your inbox
SubscribeI have spent most of my time over the last few years studying how people understand and reason about economic issues and policies. At the Social Economics Lab at Harvard, we design and run large-scale, in-depth surveys that dive into how people really think about policies and issues that deeply affect their lives. My research has looked at trade, taxation, immigration, climate policy, inflation and now AI. There are lessons across these issues that I believe will be relevant for AI as well.
When people decide which policies to support or oppose, their own material self-interest matters, of course, but it is far from the only factor at play. Their mental models of how these policies work and who gains and loses from them—and, importantly, their judgments about fairness—also play a significant role.
It is possible that AI will deliver aggregate abundance, making the overall economy bigger and wealthier. But public acceptance will depend on whether this abundance trickles down to people’s lives and to the lives of those they care about, and whether it is shared in a way they think is fair. It will also depend on how those who lose out from AI are compensated and whether there are good institutions in place to deal with this transition.
So, what are the factors that will shape whether people welcome AI or resist it, and what can policymakers learn from how people have responded to earlier economic changes?
How people see AI
A new large-scale survey we conducted in France (the research has not been published yet) offers a snapshot of how people think about AI. In it, we were able to match survey responses to payroll data, giving us lots of information about people’s work histories and their firms. (Ironically, AI is becoming an issue for online survey research, which is why having “real” people matched to their actual administrative data is key.)
We found that AI use remains limited at work. Most workers have tried some form of AI, but well under half use it at least once a week at work, and more than a quarter never use it at work. Those who do use AI mainly use chatbots, and more sophisticated uses are quite rare. As with other technologies, adoption follows a familiar divide: College grads and knowledge workers use AI much more than others. A two-speed AI workforce might be starting to form along familiar lines of education and occupation.
There is a clear gap in perceptions and attitudes across this divide. You might think that those who are most exposed to AI and use it most would be the most worried about it. But in our research, the data says they are more optimistic. They see a lot of potential for “augmentation” of their job rather than replacement. For example, they report performing familiar tasks more quickly or better and being able to take on new ones. They do expect more change in their jobs, but they also see more opportunity and report investing more in their skills in order to adapt. By contrast, non-users show greater apprehension about it.
There are at least two interpretations of this pattern. It could certainly be that familiarity breeds acceptance and that exposure to AI makes its uses and limits more concrete, reducing the fear of an unknown and perhaps scary technology. It could also be that the early adopters work in professions that are adaptable and more secure. Separating these channels will require further evidence.
Optimism about the technology can coexist with the demand for guardrails: regulation and workplace protection may make adoption more acceptable.
In our research, the workers most exposed to AI also express stronger support for regulation, retraining and protection for those who lose out. Optimism about the technology can coexist with the demand for guardrails: regulation and workplace protection may make adoption more acceptable.
How people assess economic change
Our surveys at the Social Economics Lab are comprehensive and detailed, much closer to structured interviews than to opinion polls. We ask people to explain their reasoning in their own words and think in depth about trade-offs and mechanisms. We also include experiments. Some respondents receive information in the form of text, animations or pedagogical videos that others do not, in a randomized fashion. This allows us to say what actually changes someone’s views, rather than just what is associated with them.
Across policy areas, we consistently find that the following four questions matter for how people assess change, economic phenomena or policies: How effective is it? What will happen to my own household? Who gains and loses? Is it fair? Narrow self-interest is clearly an important factor, but it is not the only consideration. People care about broader effects on the economy and society, and about the distribution and fairness of the process and outcome.
Individual perceptions sometimes differ from what our economic models would predict—and they reflect real differences in what people experience. For instance, a worker who loses their job because of a factory closure learns something very different about trade than a consumer who sees lower goods prices at the store. A low-income household experiences inflation very differently from a household with lots of financial assets and bargaining power in the workplace. The goal of this research is to understand these mental models and how they were built. This is also, I believe, how we should be thinking about AI.
The lessons of trade
When we talk about AI today, I think a lot can be learned from how people have experienced and responded to trade and trade policy. Freer trade was supposed to bring abundance too. Economists have long argued, correctly, that opening up countries to free trade can generate “gains from trade”, i.e., that all countries can be made wealthier by trading based on their comparative advantage.
In the aggregate, trade did deliver these gains. However, economic theory also predicts that there will be losers and winners from trade. Furthermore, in trade, we all have two roles. We are, on the one hand, consumers who can gain from lower prices on the goods we buy, as well as from greater variety. On the other hand, we are also workers who could be exposed to trade’s displacement effects on our job or sector.
The theory predicts that the gains are widespread, but that the losses are acute and concentrated. Indeed, my research on how people understand trade shows that this asymmetry is central to policy views. People recognize that as consumers they reap benefits from trade, but these gains are vague in their minds and hard to picture. It is difficult to imagine what prices would be like in a world without trade. Therefore, consumer gains carry almost no weight in shaping views on trade policy. Meanwhile, the threat to jobs weighs much more heavily on people’s minds, even for those not directly exposed, and moves their views toward protectionism.
Furthermore, two broader beliefs matter for views on trade policy. The first is whether trade makes the economy more productive and competitive as a whole. Those who believe in these efficiency gains are much more likely to support open markets. The second is whether those who (unavoidably) lose from trade can be fairly compensated for that loss; in my U.S. survey on trade, around 62% of all respondents said they thought this could be done with retraining, direct assistance and a general social safety net (such as unemployment insurance or cash transfers). What’s interesting is that even when people thought there would be adverse distributional consequences and more inequality from trade, they still supported it provided they believed that losers could be fairly compensated. But without that belief, even those who thought trade was making the economy more efficient did not support it.
The experimental evidence confirms the importance of redistribution: When we showed respondents evidence about the unequal effects of trade, they most often maintained support for free trade but increased their support for redistribution. The trade example indicates to me that, when people believe there is credible and fair compensation for those who lose out, support for disruptive change can coexist with concern about unequal losses. People seem to believe that change can create aggregate abundance—and is therefore worth having—but also that society may still owe something to those who bear the costs.
Personal exposure matters as well. In the same study, I found that workers in sectors exposed to greater import competition or where the labor market was more strongly affected (e.g., by the “China shock”) were significantly less supportive of free trade. This was partly because of their own direct risk and partly because this experience changed their whole mental model of how trade affects the economy as a whole. The lesson: people’s own experiences can lead them to extrapolate from “This hurts me,” to “This can hurt people like us,” to “This could hurt a lot of people.” Experience shapes our model of the economy, and that can influence views beyond the issue at hand.
Both ask people to take a leap of faith that a richer economy will eventually create abundance for them or their children.
The parallel with AI is not in the technology itself, but in these policy implications. Both trade and AI might offer diffuse or future aggregate gains alongside acute and concentrated losses that may be much more salient to people. Both ask people to take a leap of faith that a richer economy will eventually create abundance for them or their children. And both will rely on institutions and policy to help those who lose out. The current backlash against free trade and the rise in trade barriers, for example, partly reflect a perception that workers and communities hurt by trade receive too little support. The promise of gains for the economy as a whole is left to make the case for free trade on its own.
The recent lesson of inflation
Inflation provides a more recent example of how aggregate economic conditions can differ from household experience. During the recovery from the pandemic recession, employment strengthened and nominal wages rose in the U.S. But prices were rising rapidly too, with many households reporting lost purchasing power and devoting more of their budget to essentials. To them, it felt like their wages were adjusting much more slowly than prices.
In two U.S. surveys I conducted in late 2023 and early 2024, lower-income households were more likely to feel impacted and to experience costly changes to their consumption and budgets. For these households, the difficulties were real, even though the economy appeared strong overall. Aggregate abundance, with a booming stock market and overall growth, coexisted with a scarcity felt by many households.
These surveys also show a clear asymmetry in how people interpret wage and price increases. Respondents tend to attribute wage increases—some of which are truly inflation adjustments—to their own on-the-job performance or career progression, but attribute price increases to the government and firms. Many believe that employers have substantial discretion over how they set wages, rather than being subject to harsh market forces, and that they elect not to increase workers’ wages to protect their profits.
During the postpandemic rise in inflation, the idea of corporate “greedflation” was very widespread, and consistent with the perception that firms benefit while workers suffer. Furthermore, inflation generated a deep sense of inequity. First, because higher-income households’ wages were perceived to better keep pace with prices. Second, because lower-income households had to make more adjustments to deal with inflation.
In addition, while economists usually think of inflation as something that occurs in two very different scenarios—a booming economy in which demand triggers price increases versus cost shocks triggering both high unemployment and inflation, or so-called stagflation—people generally only think of the second scenario. In their minds, inflation is never the result of a booming economy or the byproduct of other good things, such as lower unemployment. Thus, the trade-offs that policymakers face are not recognized and very few people believe that, to bring inflation down, policymakers need to dampen demand or let unemployment rise. Taken together, these perceptions help explain why inflation triggered a lot of anger and blame.
For AI, this raises the question of how people will interpret changes in their pay and job security as firms adopt the technology. People may recognize that AI has made their firm more productive but still feel that their compensation or work security has declined. As the recent evidence from inflation suggests, if people believe that their employer could have shared the gains more equitably, they may see these losses as the consequence of a deliberate choice.
Understanding these responses requires us to study both what happens and how people explain and experience it. The inflation example shows how important these judgments about distribution and responsibility—even in a booming economy—can be for the politics of AI.
The world AI is entering
These earlier experiences have helped shape the mindsets, the emotions and the erosion of trust with which people now approach AI. To understand how people will respond to AI, we therefore need to examine their views about who gains and who loses from economic change, the emotions they feel and who they trust.
A more zero-sum world
Experiences of scarcity can profoundly affect views about how the economy works. Zero-sum thinking is an important mindset that we have studied. It is the belief that gains for one group or person tend to come at the expense of others. When people view the world in zero-sum terms, they see it as a fixed-sized pie: If someone gets a larger slice of the pie, someone else must be getting a smaller one. This is the backdrop against which AI is transforming the economy. Younger Americans, whose working lives will be shaped by AI, hold more zero-sum views than older Americans. Right now, it is easy to see AI in zero-sum terms. The early gains appear in the valuations and profits of a few firms, while the costs being discussed are widespread changes to career progression and entry-level jobs, pressure on wages and the loss of entire professions. Even if AI increases total output, unless people benefit they may believe that those gains are coming at their expense.
The lesson from studying zero-sum thinking is that mindsets follow lived realities.
If the gains remain concentrated and the feeling of insecurity spreads, more people will view AI as zero-sum and support policies of restriction and protection. The lesson from studying zero-sum thinking is that mindsets follow lived realities. Avoiding this thinking requires changing how the gains from AI are distributed, alongside an emphasis on its aggregate gains.
An angrier world
AI is also spreading at a time of rising anger in political discourse. In fact, even from a long-run historical perspective, this seems to be an exceptionally angry time. With my coauthors, I have studied congressional debates dating back to the 1870s, campaign speeches and millions of social media posts by politicians and citizens. We have documented a sharp rise in anger in American political communication over the past decade.
Even over the full period we can measure, starting right after the Civil War, anger in congressional speeches is the highest it has ever been. This matters because emotions shape policy views. Anger is an emotion that makes us seek out who is to blame. Our experiments show that negative emotions increase support for protectionism, anti-immigration policies and redistribution. Positive emotions are much harder to trigger and don’t have equally large effects in the opposite direction. Anger stands out specifically: It seems to trigger the desire for policy action, while fear has no detectable effect. The idea that “anger mobilizes, fear paralyzes” appears true when it comes to policy views.
AI thus arrives at an unprecedented time in terms of anger levels. Our research shows this can create a feedback loop: The more people are exposed to anger, the more likely they are to react with anger to subsequent events. And there are many ways in which AI might trigger anger in future, from layoffs that are attributed to it, to more intrusive monitoring at work, to rising inequality, especially if people see these outcomes as choices that firms have made, allowed by governments.
A less trusting world
The pandemic put trust in scientists and government under strain. In some countries, including France, trust in scientists fell substantially in 2020. However, there are important differences in who people trust across countries, groups and institutions and those differences impact how they assess AI. People may, for instance, trust scientists while doubting firms or government. In our research across 12 countries, we found that people who trusted scientists were more supportive of public health measures, more likely to comply with them and more favorable toward vaccination. Trust in government depended in part on whether government recommendations agreed with those of scientists.
We see a related pattern in our French worker data. Workers who have greater confidence in management are more likely to view AI as something assisting and transforming their work. Beyond these survey findings, the way employers use AI also matters. Deploying it to monitor employees can become a form of surveillance and reduce their control over their work. Whether workers have confidence in how their employer uses AI will also depend on decisions about who is consulted, what data is collected and whether employers keep to their commitments on how AI will be used and how workers will be supported.
What can be done
While the evolution of both technology and the economy is still highly uncertain, the lessons learned from past experiences and the backdrop against which AI is spreading suggest some avenues for policy- and decision-makers. The following are thoughts rather than concrete proposals, for which much more work is needed.
First, fairness is a deep concern for people and should be taken seriously. People care about how gains are distributed and who gets to make those decisions.
There needs to be credible support for those workers who lose their jobs or income before displacement starts happening. The example of trade shows how long these losses can last. It is possible to act much earlier on AI by putting support systems in place now—such as wage insurance, transition support, portable benefits, direct assistance and a revamped social safety net. People need to know that funding is and will be available.
Second, it is not sufficient to highlight the aggregate abundance that AI could create without linking that abundance to people’s lives. People care about how their jobs will improve, which services might become more affordable and, importantly, how they will be helped during this transition.
Third, perceptions need to be measured. We need recurrent data across countries to understand not just how people use AI, but what their broader views and concerns are. It is important to include people who do not currently use AI. Their views matter politically and are formed through news and political discourse rather than experience.
AI may create abundance on a scale that is difficult to predict, but the social and political response will in large part depend on how people perceive it. Recent history shows that aggregate gains alone are not sufficient. The rise in zero-sum thinking and anger and the erosion of trust, together with these key economic experiences, will shape how people respond to AI.
The technology will determine what is possible, but social and political choices will determine what is actually done.