AI Didn’t Cause Mass Unemployment—It Slowed Entry-Level Hiring First

On October 29, Google DeepMind published a paper titled The Economics of Intelligence, by Alex Imas, Andrew Koh, and Shane Legg. It offers a counterintuitive judgment: AI is already changing the economy, just not through the indicators the public is watching. Mass unemployment has not appeared. The change is hidden in leaner startups, flatter org charts, and slower hiring for some entry-level roles.

The paper’s core distinction is simple: competence and coordination. Most benchmarks measure a single model’s performance on isolated tasks. Firms, however, compete as teams. Economic transformation is the equilibrium result of organizations competing under supply and demand. A model can outperform every individual on a single task and still lose to a well-structured human team. The organization itself is like a kind of superintelligence.

They had 100 people take a test spanning six domains: math, verbal reasoning, science, patterns, culture, and economics. A person’s performance in one domain does not predict performance in another well. People look more like specialists in different directions. The assumption of interchangeable average-intelligence copies does not hold. Pairing complementary people to solve a specific task performs 32% better than random pairing. AI models are also jagged, but their weak spots are similar. Copying a model adds only capacity, not breadth of ability. So “is a model stronger than a person?” asks the wrong question. The right question is whether AI organizations can outperform human organizations, or hybrid human-AI organizations.

The paper uses 'mosaic intelligence' to describe organizations: complementary experts fill one another’s gaps, and the team’s capability exceeds any individual’s. A coordinator may look mediocre on individual competence alone, but within a team reduces handoff friction and lets the team scale without losing focus, hence a high equilibrium wage.

Two differences between AI and humans are singled out: replicability and coordination. The supply of human experts is fixed, and wages are affected by scarcity. Once a model reaches a certain capability, it can be replicated without limit at compute rental costs, and wages anchor to compute costs. On coordination, human teams are limited by communication bandwidth, span of control, and misaligned incentives. Coase asked why, if organizations can save costs internally, one giant firm does not take over everything. Organizational economics answers with the limits of human coordination. In the 19th century, the telegraph and railroads did not make individual workers stronger; they increased handoff speed and fidelity, reorganized production, and that is how modern firms arose.

AI agents could in theory bypass these limits. Replicas share weights and representations, can transmit complete working states—including codebases, intermediate reasoning, and full context windows—and have no conflicting career incentives. But actual coordination today is poor. OpenAI’s Noam Brown said on the Dwarkesh podcast that getting large numbers of agents to collaborate productively is very hard; multi-agent contributions account for less than 10%.

The paper breaks organizational growth into three curves: single-person work, limited scale, and unbounded scale. Competence sets the level of output per worker; coordination determines how large the organization can grow. Once coordination crosses the supercooperative threshold, each additional member actually raises team productivity, the Coasean ceiling is broken, and firm size is limited only by market size. Competence changes the economy gradually; coordination transforms it abruptly.

Combining the two thresholds, the paper identifies four regimes: minimal AGI, supercompetent AI, supercooperative AI, and ASI. Under minimal AGI, wages change first and organizational restructuring has not yet arrived, so low unemployment does not prove AI has no impact. Under supercompetent AI, firms become more efficient but not larger, and 'abundance without concentration' may occur; wages for routine experts fall, while the value of judgment, orchestration, and designing coordination architectures rises. Under supercooperative AI, large standardized markets suddenly concentrate, wages plunge, and niche and relational sectors are affected. Under ASI, large AI companies dominate deep markets, and niches can no longer shelter human teams.

Today’s labor market corresponds exactly to the position the paper describes: AI’s single-task competence index is rising while coordination lags. Symptoms include leaner startups, flatter org charts, slower entry-level hiring, and stable or rising demand for experienced people who can guide complex work. Mass unemployment and sudden monopolization have not yet appeared.

This yields a judgment: single-model leaderboards will overestimate AI’s ability to replace organizations, and underestimate the abruptness of coordination breakthroughs. The paper recommends adding organization-level benchmarks that test whether multiple agents can decompose unfamiliar problems, share context cleanly, and detect rather than amplify one another’s errors over long horizons. Economic statistics should also track wage spending and compute spending together. Wages and job roles are the earliest signals, ahead of GDP and industry concentration.

The change in entry-level roles deserves separate attention. One commenter under the tweet said that one fewer junior role is quieter than a layoff announcement, but it changes who gets a chance to enter the field. Another commenter worried that junior roles are where people learn the work; after AI takes them over, companies will lack experienced managers ten years from now. The paper does not give a timetable; it gives an observation coordinate: do not stare only at single-model scores. What could truly suddenly change market structure is agents’ ability to coordinate with one another.

Source: original paper, working paper, Rohan Paul's tweet.