Canaries in the Coal Mine: Stanford Economist Bharat Chandar on Young Workers and Surviving the AI Era
EO Korea
Bharat Chandar is an economist at the Stanford Digital Economy Lab who studies how AI affects work. He considers AI's effect on the labor market one of the most important questions in labor economics today. He says it became the core of his research over roughly the past year and a half, once he began using these tools himself and understood what they could do. In this session he covers three things: what his research has found about AI and recent employment, why entry-level workers seem most exposed, and his case for using AI as a learning tool so that careers can look less like a "ladder" and more like a "lattice."
What the data show about AI and recent employment
Chandar describes a study he published with Erik Brynjolfsson and Ruyu Chen, titled "Canaries in the Coal Mine." The researchers compared occupations with high and low exposure to AI. They used data from the payroll services company ADP to track millions of workers in the United States.
Across the workforce as a whole, the study found no large difference in employment trends between more-exposed and less-exposed occupations. The picture changed when the researchers looked only at young workers. In highly exposed occupations such as software development, customer service, and administrative work, employment among young workers declined. In less-exposed occupations, employment among young workers kept growing. Employment among more experienced workers followed its existing trends even in exposed jobs. Chandar summarizes the headline figure this way: young workers in more AI-exposed occupations saw 16% slower employment growth, which he calls a substantial number.
He links this to what many people starting their careers now report, that getting started has become hard. That is why the team chose the title "Canaries in the Coal Mine." They see these results as a possible early signal of a larger transformation and plan to keep tracking them to find out whether that is what they are.
Ruling out alternative explanations, and what remains uncertain
Chandar says it is not certain whether this is a temporary economic fluctuation or a structural shift driven by AI. So the team tested the most plausible alternative explanations.
The first was interest rates. Chandar notes that the occupations most exposed to rate changes tend to be less exposed to AI. Transportation and construction, for example, are very sensitive to interest rates but have almost no AI exposure. He treats this as one of the key points suggesting the cause is not rate movements.
The second was overhiring in the tech sector. The results stayed the same when the researchers excluded tech jobs and computer-related occupations. Across the various alternatives they tested, the findings remained very similar.
His conclusion is conditional. If a structural change in AI capabilities is affecting the labor market, the effect would not be temporary and could be a long-term shift. If the trend persists as they keep tracking it, he would read that as a sign AI is affecting work. He also points out a limit of the research: there is no clean experiment comparing a world with AI to a world without it, and more research is needed to isolate AI's impact properly.
Why young workers have lost their edge
Chandar's explanation for why young workers in particular are affected turns on the type of knowledge they bring to work. When young workers enter the labor market, much of what they do is implementation, meaning carrying out tasks based on what they learned in school. What they are not yet good at, because they lack experience, is work that depends on tacit knowledge.
He defines tacit knowledge as knowledge that depends on very local context, strategic thinking, social interaction, or things that can only be built up through actual work experience. This is the kind of knowledge that is rarely written down in books. The codified, book-learned knowledge that young workers rely on is where AI's capabilities overlap most directly. In that situation, Chandar suggests, experienced workers may hold a relative advantage both over AI and over young workers.
Why firms may underinvest in young talent
Chandar also discusses whether companies will keep training young workers. He says it is natural for firms to want to hire young people if they want to have middle managers and experienced staff in the future. The problem is that this motivation may not be strong enough. Young people don't have to stay with the company that trained them and can leave for another firm at any time. So firms may still hire some young workers, but fewer than would be in society's overall interest. Chandar describes this as a misalignment between the incentives of individual private firms and those of society as a whole.
His more optimistic view is that if AI really helps people learn and works well as an educational tool, the process of building experience could speed up. He adds that this would also require substantial changes to how education systems and universities are organized, so that people can learn better.
What AI won't do much better in the near term: physical work, strategy, and social interaction
Chandar names three areas where he thinks AI will not get much better in the short to medium term. The first is physical tasks, which are hard to perform without major advances in robotics. The second is strategic thinking and deciding what should be done. The third is social interaction.
He puts particular weight on strategic thinking and expects it to grow more important. His reasoning is that much future work will likely involve directing AI agents: telling them what to do, steering them, and guiding them as they carry out the work. The ability to state what should be done, or what outcome you want, becomes a key skill. He compares this to the role a manager plays in a company, and suggests that management-style work and giving strategic direction could become quite important capabilities.
His advice for young workers who want to build this capacity is to build and use tools as much as possible and become comfortable with that way of working. The faster people adapt to this change, he argues, the better they will handle labor market disruption and technological change.
Historical comparisons: the Luddites versus the 20th century
Chandar finds it helpful to compare AI with earlier technological shifts, and he points to two contrasting patterns. In the Industrial Revolution, he argues, the most skilled workers faced the greatest risk. His example is the Luddites, who were skilled weavers. Many of them lost their jobs because of the new inventions of the period, even though they were among the more skilled workers of their society. He sees a parallel in that knowledge workers and more highly educated workers may be more exposed to AI.
Electricity and the IT revolution in the 20th century, he says, largely ran the other way. Middle-skill and low-skill work tended to be more exposed to those technologies, and the most skilled and educated people benefited far more from them. Which of these two patterns AI will follow, he says, remains to be seen.
He flags one way AI may differ from past cases: how fast its capabilities are improving. AI today performs a much wider range of tasks than it did three years ago. When new work and new demand are created, a question follows that did not arise in the same way before: will humans do that new work, or will AI capabilities advance quickly enough that AI does it too? Chandar sees this as the point where AI may break from earlier technologies.
Augmentation versus automation
Chandar notes that there is a lot of discussion about using AI to assist workers and strengthen their capabilities, as opposed to replacing them and automating everything they do. In an essay he wrote, he aimed to propose one concrete solution that he believes could substantially assist workers: using AI as a tool to help people learn.
His example of someone already augmenting themselves with AI is the startup founder. The team is very small, but with access to AI it can get much more done. Founders can now handle functions they didn't previously know how to do, because AI tools make those functions accessible. He calls this a good example of augmentation.
He offers a simple test for whether you are being automated or augmented, and says it depends on what kind of work you focus on. Is the range of work you can do expanding, or is this technology shrinking your work? When technology benefits workers, he explains, it usually broadens the range of tasks they can do. Technology that automates or replaces workers removes some of their tasks, so they end up doing less. His goal is to find ways to augment workers so they can do more. He says the best way to do that historically has been education, because education lets workers do far more than they could before.
He then frames the present as an opportunity. He sees a chance for the biggest change in learning capability that we have experienced in 100 years or more, and he expects it to come from personalized learning with AI tools.
How Chandar uses AI in his own work, and where he doesn't
Chandar describes his own use of AI. He uses it most in mathematics. He finds AI very good at writing models and proofs, because reviewing the AI's output is much more efficient than writing everything from scratch. He counts this as an important aid to his work.
He hardly uses AI for writing. His reason is not that he distrusts AI's ability to write. It is that writing helps him think, and doing it himself helps him understand the problem better. If he writes without coming to understand the content, he says, much of the value of writing is lost.
Deciding what to delegate: the human role in "what to build"
From this he moves to a broader point. When we decide what to delegate to AI and what to keep in human hands, he says, the decision depends on what humans want. Part of it is a matter of values about right and wrong, and part is simply expressing our preferences. What do we want to build? What would make us better off and happier? What do we actually want to accomplish with AI? These are things we must communicate to AI.
Chandar says it is unclear to him how such tasks could be automated, at least in the short to medium term, because part of them depends on our own reflection. People have to think about what they truly want, and sometimes they only discover it by reflecting more deeply. For this reason he considers giving direction on what to build and implement to be more distinctively human than AI-like, in the short to medium term. He sees AI as closer to the execution and implementation side.
Two scenarios for inequality
Chandar then describes two scenarios with different implications for inequality. In the first, AI sharply reduces the advantage of having learned something new. If the barriers to reaching the top of a field or performing at the highest level in an occupation fall a great deal, because AI can handle many of the hardest tasks, labor market inequality could be much lower. The gap between people who learned a lot in school and are highly capable, and people who didn't try very hard in school, might not be very large. He finds it interesting that this implies a possible tradeoff between inequality and investment in learning.
In the second scenario, AI increases the advantage conferred by strategic thinking or social skills. Then working hard in school could become even more valuable, because anyone who can develop strategic thinking could be highly valuable in the labor market. Chandar does not say which scenario will prevail.
His advice to students is to use AI tools as much as possible, build things themselves, and focus on developing that strategic thinking. The questions he wants them to ask are how to use these tools well, where the tools still fall short, and where they as humans can add value.
Keeping critical thinking in the loop
Chandar acknowledges that AI's effects on how people think are very complex. He mentions a range of new educational interventions involving AI that are designed to keep people focused on critical thinking rather than simply handing off tasks. His example is Khan Academy, which has a feature that lets students use AI but does not give them the answer directly. Instead, the AI helps them work their way to the answer.
From career ladder to career lattice
Chandar ends with his vision. If we really harness AI's ability to support learning, he imagines a future where people can move between different occupations far more easily, as demand for those occupations shifts over time. If some occupation becomes much more economically important and we can find ways to help people move into it faster, he believes a great deal of potential could be unlocked.
He does not claim this outcome is assured; he presents it as a hope. In his words, he sincerely hopes we move away from a career ladder, which carries more risk from technological change, and toward something closer to a career lattice that better suits workers.
Young workers in occupations more exposed to AI show 16% slower employment growth. That's a significant number. The structural shift in AI capabilities that is affecting the labor market is not a temporary change. If we properly harness AI's capacity to support learning, transitioning between different occupations will become much easier. Rather than a 'career ladder,' which carries high risk from technological change, I hope we move toward a 'career lattice' structure that fits workers better.
I'm Bharat Chandar. I'm an economist at the Stanford Digital Economy Lab, and I study the impact of AI on labor. Looking back over the past year and a half or so, I think one of the most important questions in labor economics today is how AI affects the labor market. That started after I began using these tools and came to understand them. This became the core of my research, and I thought it was an important question that would have a big impact on society.
Canaries in the Coal Mine? - Facts about the recent effects of AI on employment
I published a study with Erik Brynjolfsson and Ruyu Chen. We studied how occupations are changing, comparing occupations with high and low AI exposure. Using data from the payroll services company ADP, we tracked millions of workers in the United States.
The main finding of the study is that, overall, there was no big difference in employment changes based on AI exposure. But when we focus on young workers, the difference is clear. In occupations more exposed to AI, such as software development, customer service, and administrative jobs, we saw employment declining. Meanwhile, in occupations less exposed to AI, employment continued to grow. Employment growth for more experienced workers also kept to its existing trend.
Young workers in occupations more exposed to AI showed 16% slower employment growth. Many people are just starting their careers and finding that the process isn't easy. That's also why we chose the title 'Canaries in the Coal Mine.' We're tracking these results to see whether they are a signal of the potential transformation AI will bring.
How much is AI driving this phenomenon, and how will it change going forward? We can't know for sure whether this is a temporary change in the economy or an AI-driven structural change. So we actually tested the most plausible alternatives, including interest rate changes. Occupations more exposed to interest rate changes are actually less exposed to AI. One of the important points is that fields like transportation and construction are very sensitive to interest rates but are actually barely exposed to AI. The cause doesn't appear to be interest rate changes.
Looking at overhiring in the tech sector, even if we exclude tech or computer-related jobs, the results are the same. We tested these various alternatives, but still got very similar results. As you said, if a structural shift in AI capabilities is affecting the labor market, it won't be temporary. It could be a long-term change. If we track this phenomenon and the trend holds, that's a signal that AI is affecting work.
Of course, we're not in a situation where we can run a clean experiment comparing a world with AI and a world without it. More research is definitely needed to properly isolate the effect of AI here.
Young workers have lost their competitive edge. What is the new competitive edge?
When young workers enter the labor market, a lot of what they do is about actual implementation — carrying out tasks based on the knowledge they learned in school. On the other hand, what they aren't good at due to lack of experience is work that relies on tacit knowledge, or work that requires the kind of experience you can only get by actually doing the job. It also requires more social interaction and more strategic thinking.
I think of tacit knowledge as relying on very local context, strategic thinking, social interaction, or things you can only build up through actual work experience. This is the type of knowledge that isn't written down much in books. For young workers, this is where their work overlaps more directly with AI's capabilities. And in this situation, more experienced workers can have a relative advantage not only over AI but also over young workers.
When it comes to training young workers, it's natural that companies want to hire young talent if they want to have middle managers or more experienced employees in the future. The problem here is this: even if there's an incentive because future workers are needed, in practice that incentive may not be strong enough. So from a societal perspective, companies may not hire or train enough young talent, because those people don't have to stay at the company forever. They can leave for another company at any time. So companies will try to hire some young talent, but they may not hire as many as would be in society's overall interest. This is ultimately a mismatch between the incentives of individual private companies and those of society as a whole.
The more optimistic perspective I can offer is this: if AI really helps people learn and is capable as an educational tool, that process could become faster. This will also require many changes to education systems and how things are organized at the university level so that people can learn better.
There are three things AI won't get much better at in the short to medium term. The first is physical tasks that are hard to perform without major advances in robotics. The second is strategic thinking and leading what should be done. And the third is social interaction.
I think strategic thinking is becoming increasingly important and will become even more important going forward. That's because in the future, a lot of work will likely take the form of directing and leading AI agents on what to do and guiding them to actually execute it. So I think that kind of strategic thinking — the ability to express what should be done, or what output I want — will be a very important skill going forward. And this is similar to the role a manager plays within a company. This kind of management work and the ability to provide strategic guidance could be quite an important skill going forward.
When I think about how young workers can develop these skills, it's important to build and use tools as much as possible and get used to working that way. The faster you adapt to these changes, the better you'll be able to cope with labor market disruption or technological change.
From career ladder to career lattice - Young talent who use AI will get ahead
I think it's very helpful to compare AI to these historical shifts. For example, the Industrial Revolution. Comparing AI with that period, I think it was actually a case where the most skilled workers faced greater risk — risk stemming from the Industrial Revolution. One example that comes to mind is the Luddites, who were skilled weavers. Many of them lost their jobs because of new inventions that appeared during the Industrial Revolution. They were the more skilled workers in society. The similarity here is that more knowledge workers or highly educated workers may be more exposed to AI. So I think that's an interesting comparison.
Compared to the electricity or IT revolutions, it was basically the opposite during the 20th century: middle-skill or low-skill labor tended to be more exposed to those technologies. Meanwhile, the most skilled and highly educated people benefited much more from the development of these new technologies. So it remains to be seen whether AI will be like the first case or the second.
I think there's something worth keeping in mind here. One way AI may differ from past historical cases is the speed of capability improvement. In fact, AI today performs a much wider range of tasks than it did three years ago. And as new work is created and demand arises, a question follows: will that work be done by humans, or will AI capabilities advance quickly enough that AI does that work too? I think this is where AI may differ from previous technologies.
There is a lot of discussion underway about using AI to assist workers and augment their capabilities, as opposed to replacing workers and automating everything they do. What I wanted to say in this essay was to propose one specific solution that I think could significantly assist workers: using AI as a tool to help people learn.
I think one example of someone augmenting themselves with AI right now is a startup founder. It's a very small team, but because they have access to AI, they can do much more work. With access to AI tools, they can now handle on their own many different functions they didn't know before. This is a good example of augmentation.
Whether you are more automated or more augmented depends on which tasks you focus on. Is the range of work you can do expanding? Or is your work shrinking because of the introduction of this technology? The reason this can be seen as great in terms of augmentation is that, when technology benefits workers, it often broadens the range of work they can do. On the other hand, technology that automates tasks or replaces workers eliminates some of the tasks workers have to do, so they now do less work. My goal is to find ways to augment workers so they can do more. And historically, the best way to do that is through education. With education, workers can do much more than before.
I think there's something worth keeping in mind here. One way AI may differ from past historical cases is the speed of capability improvement. And I think now is an opportunity — an opportunity for the biggest change in learning capacity we've experienced in 100 years or even longer. And that comes from personalized learning using AI tools.
For me, there are a few strands to augmentation with AI. The area where I use AI the most is math. AI is very good at writing models or proofs, because reviewing AI's output is much more efficient than writing it all yourself from scratch. So I think this is also an important aid to my work.
On the other hand, there are areas where I don't use AI — I rarely use it for writing. That's because the process of writing helps me think, and I understand the problem better when I do it myself. It's not because I don't trust that AI tools can write. If I don't understand the content by writing it myself, the value I get from writing drops significantly.
When we decide what to delegate and what to leave as the human domain, that decision depends on what humans want. Part of that is a matter of values about what is right and wrong. Part of it is simply expressing our preferences. What do we want to build? What will make us better and happier? What do we actually want to realize with AI? This is something we must express to AI.
I think it's unclear how those kinds of tasks could be automated, at least in the short to medium term, because some of them depend on our reflection. We have to think about what we truly want. Sometimes the more deeply we reflect on it, the more we discover what we truly want. So I think guidance on what to build and realize is a more uniquely human trait than an AI one in the short to medium term. I see AI as being closer to the execution and implementation side.
Imagine that AI really reduces the advantage of learning something new. What's interesting about this is that it's potentially a world with much lower inequality in the labor market — if the barriers to reaching the top of a field or achieving the best performance in a particular occupation become much lower. Because AI can do many of the hardest tasks, inequality could actually be lower in such a world. That's because the gap between people who learned a lot in school and are very capable, and people who didn't try that hard in school, may not be that large. This is somewhat interesting — there could be a potential trade-off between inequality and investment in learning.
On the other hand, if AI really increases the advantage in strategic thinking or social skills, the value of working hard in school could actually become greater. Because if you can develop strategic thinking skills, you can become truly valuable in the labor market.
I'd encourage students to use AI tools as much as possible, build things themselves, and really focus on developing that kind of strategic thinking. What's the best way to use these tools? Where are they still lacking, and where can you, as a human, add value?
When you think about the impact AI will have on how we think, there are very complex aspects. This includes a variety of new educational interventions involving AI use, designed so that people focus on critical thinking skills rather than simply handing off tasks. So there are various platforms. For example, Khan Academy has a feature where you can use AI tools, but it doesn't give you the answer right away. Instead of AI giving you the answer directly, it helps you get to the answer.
If we truly harness AI's ability to support learning, I imagine a future world where people can transition between different occupations much more easily, based on how demand for those occupations changes over time. And if an occupation becomes much more economically important, finding ways to help people transition to it faster could unlock a lot of potential. So I sincerely hope we move toward something closer to a career lattice that fits workers better, rather than a career ladder that carries more risk from technological change.
First of all, I had quite a hard time coming up with questions that ChatGPT would get wrong. It was a shock. I wondered how to stay ahead of AI, but that's actually the wrong question.
My name is Ken Ono. I'm a mathematician, and I also work in AI for mathematics. I'm a professor at the University of Virginia, currently on leave, and a founding mathematician at Axiom.
My view of intelligence has changed a lot. I believe we are not doing our best in educating our children. I'm not saying this to criticize educators. I'm an educator myself, and visiting kindergarten or first-grade classrooms is always a joy. On parent visiting days, when parents come in and talk about the wonders of math, I wish I could bottle up all that overflowing energy and keep it.
If only we could keep that sense of wonder about the world and the energy children have when everything is new. Think about where we would be today. The potential and productive creativity of someone like Ramanujan is inside all of us. Whose identity is it? It's yours.
Article published