Ronnie Chatterji: AI may change work before it changes job counts

Ronnie Chatterji doubts that AI alone will triple unemployment in the next 12 to 18 months. He recalls that U.S. unemployment tripled during the extraordinary shock of the COVID-19 pandemic; an equally rapid increase from AI would require a different economic model from the one he expects.

Reid Hoffman recalls economist Robert Solow’s observation that computers were visible everywhere except productivity statistics. Aria Finger introduces Chatterji, OpenAI’s first chief economist, and the question of whether AI’s economic effects will appear in national data faster than the gains from personal computers did.

Chatterji says OpenAI initially approached him about semiconductors after his government work on the CHIPS and Science Act. He wrote a job description for a chief economist rather than seeking the role, then accepted when the company returned with the same description.

His team combines economists and data scientists with people experienced in managing programs across an organization. He sees a responsibility to provide information about AI’s economic effects and says the job has changed repeatedly since he joined.

Over the next 12 to 18 months, Chatterji expects more companies around the world to connect AI agents to existing applications. He says the technology is still early in much of the enterprise, even if it feels commonplace in San Francisco.

He says job titles have changed less than the work underneath them. In his account, workers increasingly use AI to take on tasks outside their formal roles, such as a finance professional doing marketing work or a communications employee analyzing numbers.

Chatterji expects AI to appear more significantly in productivity data over the next 12 to 18 months. Investment in hardware already contributes to GDP, he says, but productivity is a separate question that economists must measure.

He cautions against casually extrapolating enormous growth rates. An extra one or two percentage points of growth would already be consequential, while even a modest rise in unemployment would prompt serious discussion. He considers a near-term tripling of unemployment unlikely.

Asked where adoption might spread quickly beyond engineering, he names legal and human resources. OpenAI’s lawyers, he says, have used AI to change their workflows rapidly, though he cautions that the company is not a representative sample.

He recalls a lag of roughly four to five months between technical and nontechnical groups at OpenAI adopting agentic tools broadly. He expects structured information and repeatable processes to make legal and people operations areas worth watching elsewhere.

Hoffman asks whether AI will make entrepreneurship cheaper, including making failures arrive faster. Chatterji argues that faster experimentation can help people learn, move on to another company or discover that they would rather join an existing organization.

Finger gives the example of her mother, a piano teacher running a one-person business. Chatterji expects solo operators to use AI for tasks such as marketing or administration, including in places where it is difficult to find a co-founder or hire a specialist.

He cites an OpenAI report estimating that about four million Americans were using ChatGPT somewhere in the entrepreneurial process. Brainstorming can reduce an early barrier he observed in his own research: many would-be founders never share an idea and get feedback on it.

On productivity, Chatterji distinguishes investment spending from evidence that AI itself has raised the efficiency of the economy. Existing data may be consistent with a productivity effect, he says, but determining how much is caused by AI requires more work and may become clear only in retrospect.

He compares that uncertainty with a Federal Reserve debate in the 1990s, when Alan Greenspan and Janet Yellen read emerging productivity evidence differently and the data were later revised. Chatterji expects the payoff to become clearer, though adoption is uneven across firms.

After visiting multiple countries, he sees substantial interest in Japan and South Korea, where firms already using advanced technologies and companies in the semiconductor supply chain may benefit. Elsewhere he sees stronger consumer adoption than enterprise use.

He points to Indonesia, Brazil and India as places where he has seen strong adoption or entrepreneurial activity. He suggests a young population may help explain some use, while describing Europe as engaged in a debate over productive adoption, control of technology and shifting geopolitics.

Asked whether AI investment will raise inflation before productivity brings prices down, Chatterji points to hardware supply bottlenecks such as memory. He expects those pressures to ease as supply grows and thinks lower costs for goods and services could be the larger effect over time, though perhaps not within two years.

In a hypothetical 2031 where AI delivers its promised gains, he expects early adopters to benefit, particularly when firms pair technology with complementary management practices. Even a one- or two-percentage-point increase in growth, he says, would be substantial.

Chatterji expects education and healthcare to employ more people in a richer economy: societies tend to spend more on these services as incomes rise. He says he would prefer AI to help reduce class sizes and improve care rather than simply replace teachers or clinicians.

The quality of those jobs matters. If people move into education and care roles that pay less than the jobs they leave, the transition may be less beneficial. Hoffman also raises entertainment and asks how much demand there will be for the much larger supply of AI-assisted content.

Finger asks who would pay for expanded care and education. Chatterji says the answer depends on AI actually producing a productivity dividend; if it does, governments could use the resulting revenue, potentially with changes to taxation of labor and capital, to fund services people value.

He sees another opportunity in navigation: an agent might help someone work through a financial-aid application, school access or a complex healthcare process, rather than merely supply information. He imagines public institutions using tools of their own to manage scarce resources under democratically chosen rules.

Finger asks whether removing friction could overwhelm systems that already lack enough resources. Chatterji says tools would be needed on both sides of an application or allocation process. He also sees potential value in providing an initial source of assistance where professional expertise is otherwise hard to access.

Hoffman asks what changes when services can run around the clock. Chatterji says persistent agents can work without sleep, but some actions still need a person’s clarification or permission. Systems receiving their output can also be overwhelmed.

He describes setting boundaries inside his own team around how much AI-generated material enters a shared channel and who reviews it. In his view, humans need rules for handling persistent and parallel agents, not just more generated work.

Asked whether labs can choose to augment workers rather than automate tasks, Chatterji agrees with economist Daron Acemoglu that the direction of technology involves choices. He says innovation can create new products and industries as well as reduce costs, but doubts one lab can settle the direction alone.

He argues that labs, governments and other institutions must work together on the distribution of gains and the availability of productive jobs. Sustaining that cooperation will be difficult in a competitive global economy, he says.

Finger points to the China trade shock as an example of concentrated losses and widely dispersed benefits. Chatterji agrees that harm affecting a relatively small share of the population can still have deep effects on particular communities and on politics.

He uses a free AI-assisted vacation plan to illustrate consumer surplus: a benefit to the user that is not necessarily captured in GDP. He expects some benefits to spread widely, while losses may be concentrated by industry or region. His view is that policy must respond to those uneven effects.

Asked how to build trust with people worried about AI, Chatterji starts with empathy and evidence about both gains and harms. He says a person can have a real reason for anxiety even if an aggregate unemployment figure looks reassuring.

He calls for strengthening existing automatic stabilizers, including state unemployment systems and benefits that expand when unemployment rises. He describes that as preparation that would be useful whether AI causes major job disruption or another shock arrives.

People also need to feel they retain control, he says. He names concerns from professionals about their expertise, young people about entry-level opportunities, and communities about data-center effects. Government, in his view, has to show how it will address those specific harms.

Drawing on work with former Commerce Secretary Gina Raimondo, Chatterji points to the Commerce Department as a place to influence AI adoption. He cites standards work at NIST, regional development and the department’s role in semiconductors as examples of its economic-policy reach.

On exporting models and chips, he says efficiency and scale are not the only considerations. Concentrated semiconductor manufacturing can create security and geopolitical risks. He argues that access to advanced capabilities and national security cannot be cleanly separated from economic policy.

Chatterji expects some deep technologies to face a different balance between an open global market and national security than consumer internet products did. He anticipates continuing disagreements between governments and companies over where to draw that line.

Inside a frontier lab, he sees two reasons for a chief economist: interpreting data about AI’s use for people making decisions outside the company, and applying economic reasoning to the behavior and organization of agents. He says OpenAI has an obligation to present the information without a slant.

Finger asks about credibility when research comes from OpenAI. Chatterji describes working with outside academics, sharing as much data as possible, seeking peer review and funding independent teams. He says collaborators can see things his own team may miss.

His team prioritizes evidence about how people and firms use AI now rather than long-range predictions. He thinks current data can help policymakers and business leaders make decisions, although it can seem less surprising than a forecast.

He says OpenAI also seeks perspectives across political and economic viewpoints and that building trust requires explaining the evidence in person, not only publishing it. He names Acemoglu as an economist whose interpretation of the data he would especially like to hear.

Finger asks how GDPval turns professions into tests. Chatterji explains that earlier studies listed work tasks and asked whether AI could perform them. He rejects converting the share of tasks exposed into a direct probability that the worker will lose the job.

He recalls his father getting a computer as an economics professor: the machine changed how he performed tasks, but he used it to run more models and write more papers. Chatterji says a job’s tasks can change as people find new ways to reach an objective and as demand changes.

GDPval can show how a model performs particular tasks, such as making a spreadsheet or slides, but Chatterji says the job itself is a moving target. Hoffman asks whether the occupational task database O*NET is current enough for the AI era.

Chatterji wants investment in better measurement of work and AI, including internationally. He suggests that early labor-market signals might appear in Philippine call centers or Indian business-process outsourcing, not only in U.S. employment data.

The team tries to communicate measures of how people use AI at work, a view he says complements official job statistics. He says the challenge is to make technical findings understandable without losing the accuracy on which credibility depends.

Asked what surprised him most at OpenAI, Chatterji says the speed of capability growth exceeded his expectations. Finger observes that people who tried an older system may still judge AI by an outdated experience; Chatterji distinguishes earlier chatbots from reasoning models and newer agents.

He says AI already helps economics research by suggesting ideas, finding literature and checking code or specifications. Framing a useful question and building a convincing empirical argument still require judgment; he continues to write his own papers.

As a possible early productivity indicator, Chatterji follows revenue per employee at a set of firms he considers exposed to AI. He is looking for signals of change before official statistics can fully capture it.

He argues that organizations mediate technology adoption and deserve more attention in both economics and machine learning. Capable AI does not automatically produce an economic outcome; businesses must decide how to use it.

Asked for a 2030 unemployment forecast, he says he hopes for 5% or less, while acknowledging that adverse disruption would require policies to support transitions. He estimates daily AI use among American workers could rise from roughly 40% to 80% by then, including use embedded in other software.

He would hire economist John Kenneth Galbraith for the ability to connect economic history with technological change and politics. Looking to 2036, he expects some agentic organizations to be possible, while noting open questions about legal liability and ownership.

The hosts thank Chatterji and close the conversation.