It can be surprisingly difficult to think about AI’s labor impacts. Motivated by having made our own fair share of mistakes, we hope to address five that are especially alluring:
Presuming employment falls roughly linearly with automation.
Conflating capability and cost-effectiveness.
Conflating employment and wage effects.
Conflating short-term and long-term employment effects.
Ignoring comparative advantage.
As suggested throughout, each tends to underestimate the prospects for long-term human employment. They also tend to overstate its importance.
1. The relationship between employment and automation could be highly non-linear.
New technologies have generally maintained or increased the level of human employment. Today, the U.S. unemployment rate is less than 5% despite wave after wave of new general purpose technologies over the past 200 years. That’s because previous technologies have merely automated a portion of the broad band of human skills, allowing humans to take up occupations that utilize other portions.[1]
However, many reasonably predict AI’s labor impacts will, in the end, be quite different. As the argument goes, AI (along with robots) will automate every human skill, performing every task cheaper and better, and thus leaving no economic refuge for humans. As we note in the fifth section of this post, this argument is a little quick. But even if it holds, it tells you very little about the level of human employment in the years leading up to this end state (or about end states just shy of it). For, if AI automates half of all tasks or even almost all tasks (e.g., 99%), this does not imply that human employment will be roughly 50% or 1%, respectively, of current levels.
Indeed, it does not imply that human employment will decrease at all. Instead, at such points, AI will resemble technologies of the past. It will have automated only a portion of the broad band of human skills, allowing humans to specialize in their remaining skills. If these skills are complementary to those automated or AI grows the scale of the economy sufficiently, even 1% of all current jobs (say, just the job of therapist or craftsman) could be sufficient to employ an arbitrarily large portion of humans.[2]
Hence, (i) the final level of human employment tells you little about its trajectory to that point, and vice versa, and (ii) worlds with nearly complete automation of tasks could have much greater human employment than worlds with complete automation.
2. AI may be more capable than humans yet more expensive.
In recent years, AI capabilities have greatly improved due to heavy compute spending by frontier labs. Much of that compute is used for large training runs that create new models—a fixed cost amortizable by widespread use of the new model. However, a significant and growing portion of such compute is used for inference, such as the “thinking” models perform before they answer your queries, which is a variable cost.[3] That’s why it’s more expensive (i.e., costs more “usage”) to use a model that thinks for longer. If this trend continues, frontier AI models could become quite expensive to use.[4] Hence, even if those models are more capable than humans at a given task, it may still be economical to employ humans for that task.[5]
Assuming training compute produces model instincts akin to humans’ intuitive system 1 thinking while inference compute is required for model reasoning akin to humans’ deliberate system 2 thinking,[6] the latter—all else equal—would be especially costly to automate. If so, humans may be in especially high demand for the slow, analytical thinking for which we are known, even in worlds where AI is generally more capable.
3. Automation might lower wages rather than employment.
Even if AI lowers the demand for human workers, it does not follow that the human employment rate will decline. In an efficient labor market, laborers will still be employed, even as they become less viable, just at a lower wage. In such markets, the wage offered simply drops until it is mutually beneficial for the employer and laborer.
Long-term unemployment or exit from the labor force can occur when the potential laborers’ outside options—e.g., unemployment benefits and increased leisure time—are more attractive than working. In a world with a high-paying UBI, declining wages could more easily push humans out of the workforce; but in worlds without UBI and modest or no unemployment benefits, humans might continue working even for very low wages.
4. Spikes in unemployment could be due to short-term rather than long-term effects.
Some labor impacts from AI are unlikely to persist in the long-run. Unemployment can result from labor shocks that cause people to shift from one sector to another. While those people are in transit, they are often considered unemployed. Much of the current unemployment caused by AI is plausibly due to such shocks.[7] Certain domains, such as law or software engineering (and perhaps merely sub-domains thereof), are being automated, leading people to transition in pursuit of higher wages elsewhere.
However, just because people are transitioning away from their old jobs does not imply that they won’t find new ones. Indeed, such transitions are the typical result of new technologies. Just as the relocation of buggy drivers to other parts of the economy didn’t imply that the car would lead to widespread unemployment, the relocation of lawyers and software engineers does not imply, by itself, that AI will lead to widespread unemployment.[8]
That said, such transitions can take a long time. If AI automates a sufficiently wide swath of jobs, it may be difficult for a lawyer or software engineer to find an unautomated job for which their talents are easily transferable, leading them to accept a job with a lower wage, as aforementioned, or spend a longer time searching. Still, one should expect that coming generations will be trained not to be lawyers or software engineers, but in sectors where their skills are least likely to be automated. Insofar as these new sectors are indeed refuges, widespread short-term unemployment we may see in the coming years or decades need not portend any long-term decline.
5. Comparative advantage could support full employment even if AI is absolutely better at everything.
As noted, if AI is cheaper and better at accomplishing any individual task than every human, it’s natural to think that human employment should go to zero. However, even in such worlds, Ricardian trade could maintain full employment for humans. Much as poor countries are still economically valuable even if rich countries can accomplish any economic task more efficiently, humans could remain economically valuable due to such comparative advantage. For example, even if an AI robot could perform your job more efficiently than you, so long as the compute required for that performance could even more profitably be used elsewhere, an employer may be willing to continue to employ you at a wage below the value of that compute.[9]
Of course, while Ricardo’s difficult idea is surprisingly powerful, it is no magic bullet. For, comparative advantage would only support human employment if the economic benefits of trading with or hiring humans outweigh the integration costs. For example, we do not employ chimpanzees because the cost of communicating with and training them to produce economically valuable work is greater than the value of the work they’d produce.[10] Still, producing enough economic value to exceed integration costs is a much lower bar than maintaining any absolute advantage on AI. Thus, again, the result that human employment could be much more robust to automation than typically imagined.
Conclusion
We do not mean to suggest the labor impacts of AI are not worrying. Full human employment at extremely low wages is obviously not an outcome most would find desirable. The point is that long-term, permanent unemployment is receiving far too much attention. Not only because it is less likely than many presume but also because it’s a misleading indicator of human welfare. Those who have suffered from globalization and computerization have suffered in large part due to reduced real wages rather than mass unemployment. If history is any guide, it would be yet another mistake to narrowly focus on employment levels in the coming era.
David H. Autor, “Why Are There Still So Many Jobs? The History and Future of Workplace Automation,” Journal of Economic Perspectives (2015), https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3.
Of course, inference costs of once-frontier capabilities generally fall over time, so such model capabilities are unlikely to remain expensive forever. Ben Cottier, Ben Snodin, David Owen & Tom Adamczewski, “LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks,” Epoch AI (2025), https://epoch.ai/data-insights/llm-inference-price-trends.
Seyed M. Hosseini & Guy Lichtinger, “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data” (2025), https://ssrn.com/abstract=5425555.
Here, unlike in this post’s second section, human employment persists due to the opportunity cost of AI use rather than its direct cost. For comparison: section two is about worlds where humans, e.g., cook meals because robot performance of that task would cost $200 (ignoring opportunity costs) while this section is about worlds where a robot could cook meals for $1 (ignoring opportunity costs) but because the required compute has even more profitable uses in other economic sectors, the task is still performed by humans. See Pascual Restrepo, “We Won’t Be Missed: Work and Growth in the AGI World” (2025), https://bpb-us-w2.wpmucdn.com/campuspress.yale.edu/dist/c/4765/files/2025/10/AGI_v4.pdf.
[Co-authored with Seyed M. Hosseini; cross-posted from my Substack. We'd be grateful for feedback.]
It can be surprisingly difficult to think about AI’s labor impacts. Motivated by having made our own fair share of mistakes, we hope to address five that are especially alluring:
As suggested throughout, each tends to underestimate the prospects for long-term human employment. They also tend to overstate its importance.
1. The relationship between employment and automation could be highly non-linear.
New technologies have generally maintained or increased the level of human employment. Today, the U.S. unemployment rate is less than 5% despite wave after wave of new general purpose technologies over the past 200 years. That’s because previous technologies have merely automated a portion of the broad band of human skills, allowing humans to take up occupations that utilize other portions.[1]
However, many reasonably predict AI’s labor impacts will, in the end, be quite different. As the argument goes, AI (along with robots) will automate every human skill, performing every task cheaper and better, and thus leaving no economic refuge for humans. As we note in the fifth section of this post, this argument is a little quick. But even if it holds, it tells you very little about the level of human employment in the years leading up to this end state (or about end states just shy of it). For, if AI automates half of all tasks or even almost all tasks (e.g., 99%), this does not imply that human employment will be roughly 50% or 1%, respectively, of current levels.
Indeed, it does not imply that human employment will decrease at all. Instead, at such points, AI will resemble technologies of the past. It will have automated only a portion of the broad band of human skills, allowing humans to specialize in their remaining skills. If these skills are complementary to those automated or AI grows the scale of the economy sufficiently, even 1% of all current jobs (say, just the job of therapist or craftsman) could be sufficient to employ an arbitrarily large portion of humans.[2]
Hence, (i) the final level of human employment tells you little about its trajectory to that point, and vice versa, and (ii) worlds with nearly complete automation of tasks could have much greater human employment than worlds with complete automation.
2. AI may be more capable than humans yet more expensive.
In recent years, AI capabilities have greatly improved due to heavy compute spending by frontier labs. Much of that compute is used for large training runs that create new models—a fixed cost amortizable by widespread use of the new model. However, a significant and growing portion of such compute is used for inference, such as the “thinking” models perform before they answer your queries, which is a variable cost.[3] That’s why it’s more expensive (i.e., costs more “usage”) to use a model that thinks for longer. If this trend continues, frontier AI models could become quite expensive to use.[4] Hence, even if those models are more capable than humans at a given task, it may still be economical to employ humans for that task.[5]
Assuming training compute produces model instincts akin to humans’ intuitive system 1 thinking while inference compute is required for model reasoning akin to humans’ deliberate system 2 thinking,[6] the latter—all else equal—would be especially costly to automate. If so, humans may be in especially high demand for the slow, analytical thinking for which we are known, even in worlds where AI is generally more capable.
3. Automation might lower wages rather than employment.
Even if AI lowers the demand for human workers, it does not follow that the human employment rate will decline. In an efficient labor market, laborers will still be employed, even as they become less viable, just at a lower wage. In such markets, the wage offered simply drops until it is mutually beneficial for the employer and laborer.
Long-term unemployment or exit from the labor force can occur when the potential laborers’ outside options—e.g., unemployment benefits and increased leisure time—are more attractive than working. In a world with a high-paying UBI, declining wages could more easily push humans out of the workforce; but in worlds without UBI and modest or no unemployment benefits, humans might continue working even for very low wages.
4. Spikes in unemployment could be due to short-term rather than long-term effects.
Some labor impacts from AI are unlikely to persist in the long-run. Unemployment can result from labor shocks that cause people to shift from one sector to another. While those people are in transit, they are often considered unemployed. Much of the current unemployment caused by AI is plausibly due to such shocks.[7] Certain domains, such as law or software engineering (and perhaps merely sub-domains thereof), are being automated, leading people to transition in pursuit of higher wages elsewhere.
However, just because people are transitioning away from their old jobs does not imply that they won’t find new ones. Indeed, such transitions are the typical result of new technologies. Just as the relocation of buggy drivers to other parts of the economy didn’t imply that the car would lead to widespread unemployment, the relocation of lawyers and software engineers does not imply, by itself, that AI will lead to widespread unemployment.[8]
That said, such transitions can take a long time. If AI automates a sufficiently wide swath of jobs, it may be difficult for a lawyer or software engineer to find an unautomated job for which their talents are easily transferable, leading them to accept a job with a lower wage, as aforementioned, or spend a longer time searching. Still, one should expect that coming generations will be trained not to be lawyers or software engineers, but in sectors where their skills are least likely to be automated. Insofar as these new sectors are indeed refuges, widespread short-term unemployment we may see in the coming years or decades need not portend any long-term decline.
5. Comparative advantage could support full employment even if AI is absolutely better at everything.
As noted, if AI is cheaper and better at accomplishing any individual task than every human, it’s natural to think that human employment should go to zero. However, even in such worlds, Ricardian trade could maintain full employment for humans. Much as poor countries are still economically valuable even if rich countries can accomplish any economic task more efficiently, humans could remain economically valuable due to such comparative advantage. For example, even if an AI robot could perform your job more efficiently than you, so long as the compute required for that performance could even more profitably be used elsewhere, an employer may be willing to continue to employ you at a wage below the value of that compute.[9]
Of course, while Ricardo’s difficult idea is surprisingly powerful, it is no magic bullet. For, comparative advantage would only support human employment if the economic benefits of trading with or hiring humans outweigh the integration costs. For example, we do not employ chimpanzees because the cost of communicating with and training them to produce economically valuable work is greater than the value of the work they’d produce.[10] Still, producing enough economic value to exceed integration costs is a much lower bar than maintaining any absolute advantage on AI. Thus, again, the result that human employment could be much more robust to automation than typically imagined.
Conclusion
We do not mean to suggest the labor impacts of AI are not worrying. Full human employment at extremely low wages is obviously not an outcome most would find desirable. The point is that long-term, permanent unemployment is receiving far too much attention. Not only because it is less likely than many presume but also because it’s a misleading indicator of human welfare. Those who have suffered from globalization and computerization have suffered in large part due to reduced real wages rather than mass unemployment. If history is any guide, it would be yet another mistake to narrowly focus on employment levels in the coming era.
David H. Autor, “Why Are There Still So Many Jobs? The History and Future of Workplace Automation,” Journal of Economic Perspectives (2015), https://www.aeaweb.org/articles?id=10.1257/jep.29.3.3.
That said, the extent of such employment will depend on the percentage of the human population that can adequately develop these non-automated skills.
Toby Ord, “Evidence that Recent AI Gains are Mostly from Inference-Scaling” (2025), https://www.tobyord.com/writing/mostly-inference-scaling; Deloitte, “Why AI’s Next Phase Will Likely Demand More Computational Power, Not Less” (2025), https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.html
Of course, inference costs of once-frontier capabilities generally fall over time, so such model capabilities are unlikely to remain expensive forever. Ben Cottier, Ben Snodin, David Owen & Tom Adamczewski, “LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks,” Epoch AI (2025), https://epoch.ai/data-insights/llm-inference-price-trends.
Fleming, Svanberg, Li, Goehring & Thompson, “Beyond AI Exposure: Which Tasks Are Cost-Effective to Automate?” (2026), ssrn.com/abstract=6691902.
Daniel Kahneman, Thinking, Fast and Slow (2011).
Seyed M. Hosseini & Guy Lichtinger, “Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data” (2025), https://ssrn.com/abstract=5425555.
Hyman, Lahey, Ni & Pilossoph, “How Retrainable are AI-Exposed Workers?” (2025), nber.org/papers/w34174.
Here, unlike in this post’s second section, human employment persists due to the opportunity cost of AI use rather than its direct cost. For comparison: section two is about worlds where humans, e.g., cook meals because robot performance of that task would cost $200 (ignoring opportunity costs) while this section is about worlds where a robot could cook meals for $1 (ignoring opportunity costs) but because the required compute has even more profitable uses in other economic sectors, the task is still performed by humans. See Pascual Restrepo, “We Won’t Be Missed: Work and Growth in the AGI World” (2025), https://bpb-us-w2.wpmucdn.com/campuspress.yale.edu/dist/c/4765/files/2025/10/AGI_v4.pdf.
H/t Carl Shulman, Dwarkesh Podcast (2023), https://dwarkesh.com/p/carl-shulman.