20260226-pro-worker-ai-hamilton-project

AI + work: Building pro-worker AI

The Hamilton Project at Brookings Institution virtual event, 2026-02-25.

Panelists: Daron Acemoglu (MIT), David Autor (MIT), Simon Johnson (MIT) Moderator: Natasha Sarin (Yale Law School)

Accompanying essay: "Building Pro-Worker AI" by Acemoglu, Autor, and Johnson, released by The Hamilton Project.

Table of Contents

• Introduction and Opening Remarks
• Conceptual Framework: Defining Pro-Worker AI
• Concrete Examples of Pro-Worker AI
• Policy Proposals
• Q&A Session

---

Introduction and Opening Remarks [00:01:04]

Aviva Arendine: Good afternoon everyone and thank you so much for taking the time to join us today. I'm Aviva Arendine, director of the Hamilton Project at the Brookings Institution. The Hamilton Project is a nonpartisan organization that puts forward analysis and policy proposals to support broad-based economic growth. As we all know, we're already experiencing rapid growth of artificial intelligence in and out of the workplace. Earlier this week, the Hamilton Project released an essay from Daron Acemoglu, David Autor, and Simon Johnson, all of MIT, entitled Building Pro-Worker AI. The essay defines pro-worker AI as technology that makes human skills and expertise more rather than less valuable. It also argues that the market is underinvesting in the development of pro-worker AI and that policy interventions can help.

Natasha Sarin: Thanks so much, Aviva. I have a million questions and we shared so many thoughtful questions from the audience today. So, what I agreed with our authors to do is to spend about 10 or 15 minutes up top laying out some of the core arguments in the paper and then using that as a launchpad for our conversation.

---

Conceptual Framework: Defining Pro-Worker AI [00:03:42]

Daron Acemoglu: 52% of American workers are worried about AI but this isn't a fear of the unknown. 42% of those currently using AI think that it will lead to job loss compared to only 30% among those who are not using AI. Americans are right to be worried because AI is presented as the greatest automation technology of all time. AGI — artificial general intelligence — being the apogee of this vision. Automation can increase productivity, but it has a number of adverse effects on workers who may suffer earnings losses, displacement, and the inequality between capital and labor can worsen.

This paper is arguing that there is a different pro-worker direction of AI. We define pro-worker technologies as those that expand human capabilities and make worker skills and expertise more valuable. Pro-Worker AI is AI that serves this purpose.

Tasks as the Unit of Analysis

Tasks are the constituent units of production — what you need to do in order to deliver a good or a service. For example, for textile production you need spinning, weaving, finishing, dying, and a range of marketing and white-collar tasks. A key question for firms is how to perform different tasks — which factors to allocate them to.

Labor often performs tasks using broad skills (analytical knowledge, physical strength) but also expertise — acquired knowledge for performing those tasks. For example, a welder requires hand-eye coordination and physical strength, but also expertise about machinery and production processes. It is part of the reason why welders in factories used to be paid the equivalent of more than $40 an hour. Expertise is often quite critical for workers to have well-paying jobs.

Five Types of Technological Change

We need to understand technology within the context of how it reallocates tasks and how it makes task production more productive. It is useful to distinguish five different types of technological changes:

1. Labor Augmenting Technologies

What economists call labor augmenting technologies make labor more productive in the tasks already assigned to it. For example, better hammers and better tools would make workers in construction more productive. That increases productivity, no doubt — the same worker with better tools can perform the same tasks as two workers were able to do before.

But it has ambiguous effects on things workers care about:

• Labor share may not increase and may in fact decrease
• As fewer workers are necessary to perform the same amount of tasks, the price of those tasks can go down
• The expertise that workers had (e.g., for advanced building trades) gets devalued

There isn't a natural win-win situation for labor from labor augmenting technologies.

2. Capital Augmenting / Automation Technologies

The industrial revolution was about automation in the textile industry. Robots for welding rationalized manufacturing, but they were decidedly mixed news for workers. Many welders lost their jobs and the premiums they enjoyed.

In general, automation technologies:

• Displace workers
• Always reduce the labor share of production (capital substitutes for tasks previously performed by labor)
• May increase or reduce wages — there is no law of economics that says automation should be bad for workers in the absolute, but there is a distinct possibility that so many are displaced that wages can even go down
• Commodify expertise — making that expertise unnecessary. For welders, all that expertise gained in specific industries became worthless once robots started performing those tasks

Automation can take the form of small task productivity improvement but a lot of task displacement. It's decidedly a mixed bag for workers — definitely not our definition of pro-worker technology.

3. New Task Creating Technologies (Pro-Worker)

Pro-worker technologies create new tasks — new opportunities for workers to work alongside tools, gain new expertise, and make their existing expertise more valuable. This reinstates labor rather than displacing it.

This reinstatement is critical. If welders lose jobs to automation, they may find other jobs, but it wouldn't be new work — it would be jobs that lower-paid workers used to do. But with new task-creating technologies, you have a host of new possibilities. That's why advanced machinery starting in the early 20th century led to many new high-paying jobs: new machines required operators, technicians, supervisors, and so on.

4. Expertise Leveling Technologies

GPS-based navigation tools enable drivers who previously couldn't deal with complex roads (such as those in London) to do a better job. But this is not new task creation — workers with greater expertise (black cab drivers) could already do that. This enables less experienced workers to take over some tasks, so the value of expertise is mixed — an ambiguous kind of technology when it comes to pro-worker direction.

5. (Additional types exist, but these four capture the essential distinctions.)

The Core Argument

AI has tremendous potential to create new tasks and new ways of making expertise and workers collaborate with machinery. We are not the first to make this point — many pioneers of computer science foresaw this possibility 60–70 years ago. But importantly, this is not the direction that AI is going. And that's why building pro-worker AI — in fact, building pro-manual-worker AI — is a possibility, is socially beneficial, and is technically quite feasible.

---

Concrete Examples of Pro-Worker AI [00:11:45]

David Autor: Let me set the scene with a hypothetical example, then some real examples.

Hypothetical: Aviation Maintenance Technicians

There are 161,000 aircraft maintenance technicians in the US, median pay about $80,000/year, federally certified skilled vocational work. Imagine three hypothetical tools:

1. The AMT Automator (NOT pro-worker) Makes the work so simple that any high school kid with a screwdriver could do the job — the worker is just the eyes and hands of the machine. Even if this hired lots of people, there would be no expertise involved. They'd never make high pay because it would be a commodity.

2. The AMT Assistant (Pro-worker, with caveats) The technician remains the expert decision-maker, but AI extends their diagnostic and repair capabilities, handles paperwork, supports training via flight-simulator-like features. This takes the expertise of that technician and leverages it — allows them to solve harder problems, do more reliable work, use their judgment but go further. A mixed bag: creates new competition between junior and senior technicians.

3. The AMT Liftoff (Pro-worker, new task creation) Helps aircraft maintenance technicians transition their skills into the spaceflight sector, which is rapidly growing and requires a whole new set of skills.

The key point: These are the same occupation, the same underlying goal, but these tools have radically different implications for workers. The design choices matter. A pro-worker technology is emphatically NOT a technology that enables anyone to do anything without any training. The practical test: does the technology make human skills and expertise more useful rather than less necessary?

Real Example 1: Schneider Electric (Skilled Trades)

A global electrical services firm with field technicians who do on-site diagnosis and repair. These people have tons of training and skills — some have PhDs. But what they don't have is real-time field support of the depth of Schneider's whole knowledge enterprise.

They built a tool that allows field technicians to upload photos, diagnostic data, and field notes. The AI matches that with databases and provides:

• Real-time guidance
• Access to plans and repair manuals
• Recommended steps
• Regulatory compliance advice
• Translated reports

This is an enabling technology — it allows technicians to leverage their expertise to go further once they get to a job site.

Real Example 2: US Patent Office (White-Collar Government Work)

Patent examiners used to spend half a day or a full day doing old-style boolean searches (author last name, date of release) just to figure out what prior art was relevant. Now the Patent Office uses an AI tool called "More Like This" that allows conceptual matching of patents according to substance and category. This allows the patent examiner to focus on the judgmental task of evaluating what's novel rather than the paperwork task of looking for what else has been written.

Real Example 3: Hearing-Impaired Delivery Workers in China (Low-Wage Gig Work)

China has approximately 200 million gig workers. A non-trivial number have hearing impairments — which is a big problem for delivery workers because they're expected to call customers to get gate codes and directions.

A developer built a speech-to-text-to-speech tool: the worker types what they want to say, it speaks to the customer; when the customer speaks, it translates back to text in real time. The technology is prosaic (Siri has had this for 15 years) but transformative:

• Closed the customer review gap between hearing-impaired and non-hearing-impaired workers
• Opened a performance gap where hearing-impaired workers actually performed better (more experienced, more dedicated, fewer outside options)

The Takeaway

These examples span skilled trades, white-collar government work, and low-wage gig work. Pro-worker AI is not a niche concept just for high-tech jobs. In fact, it has even more potential in non-degree work that has been historically underserved by technology. The question is not whether this is technically feasible — these tools exist today. The question is: why isn't the market producing more of them?

---

Policy Proposals [00:18:54]

Simon Johnson: The key elements of pro-worker AI are twofold: the creation of new tasks, and those new tasks must involve and value human expertise. Three categories of policy ideas:

Category 1: Government as Leader and Experimenter

Healthcare and education are sectors where the federal government and state/local governments are really important players — they spend a lot of money, they procure things. These are places where government can say: we want pro-worker AI, we want to enhance human capabilities, we want to provide better services. These are places to experiment, lead, and have big demonstration effects.

Government building AI expertise at federal, state, and local levels — not just "let's automate, let's replace workers, let's reduce costs" but "let's be creative, let's come up with new things we can do."

Grant making through NSF, NIH, Department of Defense, Energy, Agriculture — these grant processes could explicitly and directly ask for efforts to develop pro-worker AI. It has to have new tasks, and new tasks have to value human expertise in order to count.

Competition is needed — both in what government does and within the private sector. We're not saying anyone can pick a winner.

Category 2: Tax Code Reform

The tax code has come to favor excessively the adoption of algorithms over the promotion and development of human expertise and human tasks. This asymmetry could be redressed. We understand there are revenue considerations, but changing that asymmetry is really important.

Category 3: New Markets and Worker Voice

Creating new markets and activities around:

• Worker voice
• Worker ownership of expertise
• Ownership of data

Category 4: Breaking Down Barriers

Addressing barriers caused by licenses in various professions — if we're going to expand human capabilities (e.g., for nurse practitioners or technicians), they have to be allowed to do more than they've been allowed to do before.

The Framing

We're not anti-technology. We're not trying to prevent AI from being adopted. We know it's coming. The direction of technology can be changed. Choices can be made. The goal is opportunity for all humans at all skill levels, irrespective of education or background. The path we're currently on is not the pro-worker AI path — that's unfortunate, regrettable, avoidable, and we're working to change it.

---

Q&A Session [00:24:20]

Natasha Sarin (moderator): One thing that seems hard about pro-worker AI is that it's somewhat amorphous and hard for a policymaker to evaluate in real time. When has government successfully shaped the contours of technological progress?

Historical Precedents and the Direction of Technology

David Autor: Pro-worker AI needs to be more clearly measured and delineated. But many things we aspire to do are also amorphous — the whole Silicon Valley ecosystem is burning with the desire to "beat China," which is also amorphous. The question is changing the priorities of researchers. If tomorrow 40–50% of AI talent decided to use AI to create better tools and information for workers rather than automating everything and chasing AGI, we would soon get pro-worker AI.

Going back to the industrial revolution: the big disruptions came from rapid mechanization of weaving, which led to a more than 2/3 drop in weavers' wages and horrible factory conditions. Alternatives existed that could have been done more slowly, with more effort on improving other technologies. The Chartist movement, the union movement, new industries — all played a role in changing the direction. We don't need to repeat those mistakes. But if you put all the onus on government alone to steer us, I don't think that's going to work. It needs to be a collective effort.

Simon Johnson: Three things:

1. DOD Self-Driving Vehicle Grand Challenge with $1 million prize money got an entire industry jump-started. Let's do a Grand Challenge for Pro-Worker AI.
2. Digital advertising and social media — the creation of the attention economy was a deliberate government policy that brought us all kinds of trouble. "Fix it later" doesn't work.
3. Section 230 of the Communications Decency Act (1996) — that was a government policy, a decision.

On Mixed-Bag Technologies: Uber and Waymo

Natasha Sarin: How should we evaluate technologies that are great for some workers but bad for others? Uber opened driving to millions but commodified incumbent expertise. Waymo could mean the end for ride-hailing drivers but is great for consumers who get safer commutes.

David Autor: Automation has many benefits, but the speed of change matters enormously. We learned during the China trade shock and the industrial revolution that the rate of labor market adjustment isn't nearly as fast as the rate at which technologies can change.

We are not advocating for slowing things down — it's a question of which direction we're pushing in. Self-driving cars took two decades of multi-trillion-dollar investment. That was a choice of what to prioritize. If you want to run a surveillance state with real-time content moderation, that's a huge investment too — some countries have done that. What we use AI for is not just what Claude says at the prompt — it's what capacities we choose to build out.

A metric of pro-worker AI: does it enable people without elite education to do high-value services — healthcare delivery, paralegal services, software, kitchen design, contracting, skilled repair? Does it enable people to take expertise they've mastered and apply it to a more valuable set of problems?

Simon Johnson on Waymo vs. Uber:

• Uber replaced taxis: bad news for taxi medallion owners, but a lot of new people entered the urban chauffeur business. Relative wages in that sector didn't change much. Maybe 1–1.5 million people drive for Uber today.
• Waymo is completely different — it's an automation technology. It replaces human drivers. Some tech support jobs exist, but most are not in the United States. "People who close Waymo doors when left open by passengers — I don't think those jobs come with health insurance or 401ks."
• Waymo has 20% market share in San Francisco and is expanding. Consumers like it — it's not a scam.
• We never said "stop Waymo, slow it down." We're saying: create new tasks. You have to run faster if you want to create new jobs and drive demand for labor.

On the "Automation = Progress" Assumption

Natasha Sarin: Over longer time spans, automation has been a significant engine of human prosperity. The aggregate effect has been to make societies richer. What time horizon should we evaluate pro-worker policy against?

Daron Acemoglu: You've assumed that automation in the past automatically created new jobs and prosperity. The US had 50% of its labor force in agriculture, now 2–3%. A lot of new work was created — but that was not automatic. David and co-authors show that two-thirds of jobs performed today didn't exist in their current form 60 years ago.

If agricultural workers had simply been thrown into construction, that would have pressed down construction wages, capital share would go through the roof, and wages would probably fall. What was critical was that new technologies and new organizations created new work in manufacturing, new industries, new clerical work. That required both entrepreneurship and new technologies to be developed.

David Autor: Two historical data points:

1. Mechanization of textiles (1770s–1830s UK): Real wages for textile workers in Manchester and Lancashire did not really move between the 1780s and the 1830s. Huge fortunes were made — entrepreneurs became some of the richest people in the world — but no gain in real wages for 50 years until railways arrived and created new tasks.

2. Digital transformation (1980s–present): While headline growth looks respectable, we've had massive bifurcation of the labor market — a crushing of middle skills, middle education, middle America pushed down to the lower end while high-skilled people did extremely well. This is 40–50 years. If AI reinforces that established pattern, we're not talking about a pause but a continuation and intensification of labor market polarization. The social and political consequences are pretty obvious to everybody.

Simon Johnson: The China trade shock of 20 years ago was like technological progress — it lowered prices, was pro-consumer. It was also crushing for manufacturing-intensive communities and scarring for the workers. Those communities have started to recover 20 years later, but not the people who were doing those jobs — completely different set of people. Nobody would now argue that was the best possible way we could have handled it.

On Market Failures and Why Firms Don't Build Pro-Worker AI

Natasha Sarin: If pro-worker AI could be profit-maximizing, why aren't firms pursuing it themselves? What market failure are we correcting?

Daron Acemoglu: The market is great at allocating apples and oranges but doesn't always work well when there is monopoly and external effects. Standard Oil was very profit-maximizing but not great for American society.

When it comes to innovation direction, profit-maximizing is not always good for society. Reasons:

• Business model lock-in: AI is led by companies with a business model around selling automation tools and digital advertising. None have specialized in pro-worker technologies, which are difficult to monetize (the workers who gain new expertise capture some of the benefits).
• Concentrated industry: AI requires enormous resources. Promising startups get bought by established players — a very restricted market.
• Automation is actually hard: Even with impressive robots and software, productivity gains from automation are quite small (work with Pascal Restrepo). Wholesale automation of tasks requires difficult organizational changes.
• Pro-worker AI could jumpstart more powerful productivity growth by making human resources — at the center of many businesses — more productive.

The trade-off is not between progress-with-adverse-distributional-effects versus pro-worker. It's about a different direction that could be much better for both social outcomes and productivity.

Simon Johnson: The uniformity of vision and power of fashion around automation in the corporate sector is "a little oppressive." Computer science graduates who were told since age four to learn to code can't get jobs now — the fashion has shifted toward "Claude this week, something else next week."

Daron Acemoglu: The entire AI and computer science field was shaped from early on by a vision of machines replicating the human mind — the Dartmouth conference participants thought that was within a few years' reach. Science fiction plus that background has conditioned the industry to think AGI is the most natural path. But it's a very difficult path and there are many other things we could do.

On China and the AI Arms Race

Daron Acemoglu: If China did not exist, Silicon Valley would have had to reinvent it — it's such a boogeyman that helps the AGI agenda. The AGI craze and zero-sum framing is actually hurting American businesses. In many technology areas China is more advanced — Chinese companies are now better at implementing robots and approaching AI integration into manufacturing. We have a lot to learn from them. If we move from AGI to pro-worker AI, it will help relations with China and kickstart a better agenda for improving manufacturing.

On All Policy Levers

Simon Johnson: Full disclosure — the T in MIT stands for Technology. I'm all about science and technology — more of it, but for what purpose? Creating more good jobs. That's the bumper sticker: More Good Jobs.

Accelerate what matters — what drives demand for human expertise. On policy: why can't we have it all? More redistribution, more worker training, more worker voice — AND change the direction of technology. Tax code reform is one element, but also encouraging innovation to be less algorithm-intensive and more human-intensive.

David Autor: We should be using all levers available:

• Better social safety net
• Wage insurance
• Universal basic capital giving people an ownership stake not just vested in their labor
• We can't slow AI down — we're definitely in a prisoner's dilemma race
• But we're missing opportunities — not saying automation is evil, but there are opportunities to make human expertise more effective

There's huge upside, huge risk. We should look for upsides, mitigate risk, and modestly steer in a socially beneficial direction.