Alex Imas is a professor of behavioral science, economics, and applied AI at Chicago Booth and Director of AGI Economics at Google DeepMind. His research spans decisions under risk, discrimination, AI adoption, and markets in which people delegate choices to AI agents. — DeepMind Institute — Economic Policy for AGI.

Visual summary of operating lessons from Alex Imas.

Part 1: Scarcity and the Relational Sector

  1. On the future of scarcity: Imas argues that if AI makes many commodities inexpensive to produce, scarcity could shift toward goods and services whose value depends on human participation. He presents this as a prospective framework, not a measured inevitability. — What Will Be Scarce?.
  2. On human-centric industries: He calls teaching, nursing, therapy, hospitality and other person-to-person services examples of a possible relational sector, where a human being part of the service can itself matter to buyers. — What Will Be Scarce?.
  3. On the premium for exclusivity: In experiments with Kristóf Madarász, randomly excluding others from access to an otherwise identical good roughly doubled willingness to pay in some treatments; this is experimental evidence, not a universal pricing rule. — What Will Be Scarce?.
  4. On the human art premium: In research Imas describes with Graelin Mandel, exclusivity raised bids for human-made artwork by 44%, compared with a smaller 21% increase for AI-generated art in that study. — What Will Be Scarce?.
  5. On the value of personal connection: For some buyers, knowing that a particular person made a good can be part of what they value, even apart from the good’s functional qualities. — What Will Be Scarce?.
  6. On the limits of objective automation: Imas uses Starbucks as an example of a company reversing some standardization and restoring barista staffing and hospitality after customers responded to a less personal store experience. — What Will Be Scarce?.
  7. On AI usage beyond work: The ATLAS study reports that Gemini usage spans categories accounting for about 98% of Americans’ waking time; that figure describes coverage of activity categories, not the share of time Americans actually spend using AI. — Google AI & Economy ATLAS.

Part 2: Automation, Job Exposure, and Labor

  1. On the definition of a job: Imas describes a job as a bundle of tasks: automating some components does not by itself imply that the entire role disappears. — Dwarkesh Patel Interview.
  2. On the true meaning of AI exposure: When AI performs some tasks in a job, workers may be able to focus on remaining tasks; whether that raises employment or pay depends partly on demand for the output. — Dwarkesh Patel Interview.
  3. On the O-ring model of automation: In his O-ring example, automating nine of ten tasks can leave the remaining human task as a bottleneck, but additional hiring depends on how strongly demand responds to cheaper output. — Dwarkesh Patel Interview.
  4. On the absence of an immediate jobs apocalypse: In the 2026 interview, Imas said aggregate labor data did not yet show a broad white-collar collapse from AI, while acknowledging possible weaker hiring among junior software developers. This is a dated observation, not a forecast. — Dwarkesh Patel Interview.
  5. On historical precedents: Imas recounts Ricardo’s concern that machinery would displace existing jobs; those particular jobs did disappear, yet later work emerged and aggregate employment did not vanish. — Dwarkesh Patel Interview.
  6. On consumer demand elasticity: Whether task automation expands or contracts a field’s employment depends in part on demand elasticity: how much more of the cheaper output people will buy. — Dwarkesh Patel Interview.
  7. On collaboration over replacement: In a study of Gemini interactions, Imas and coauthors found work-related usage spread across many occupations but predominantly collaborative, with relatively limited end-to-end task automation in the observed data. — Google AI & Economy ATLAS.

Part 3: Behavioral Anomalies in Economics

  1. On the realization effect: Imas found opposite risk responses depending on whether a prior loss had been realized: participants became more risk-averse after realized losses but took more risk while losses remained on paper. — The Realization Effect.
  2. On dynamic inconsistency: In brokerage data and experiments, participants planned to keep taking risk after gains and stop after losses, but often cut gains early and chased losses when making subsequent decisions. — Dynamic Inconsistency in Risky Choice.
  3. On the value of commitment: The same research found that offering participants commitment to their initial strategy increased willingness to take risk. — Dynamic Inconsistency in Risky Choice.
  4. On the dynamics of discrimination: In a platform field experiment, accounts presented as women faced initial disadvantage, but positive evaluation histories reversed the pattern; the authors interpret the reversal as evidence of biased beliefs in that setting. — The Dynamics of Discrimination.
  5. On the shift in behavioral economics: Imas says behavioral economics moved beyond early laboratory demonstrations when field studies showed related anomalies among experienced people facing consequential decisions. — Think Better with Alex Imas.
  6. On missing factors in mental models: Imas argues that observed choices may reflect a distorted mental representation of the options, not just underlying preferences; missing or misunderstood features can change what a choice reveals. — Justified Posteriors Transcript.

Part 4: AI in the Firm: Adoption and Systemic Risks

  1. On FOMO-driven layoffs: Imas warns that if markets begin treating layoffs as a signal that firms are adopting AI, companies may feel pressure to cut jobs to keep up even when the underlying productivity case is weak. He presents this as a risk, not an observed general pattern. — Dwarkesh Patel Interview.
  2. On the cascade effect: He describes a possible coordination cascade in which firms imitate visible AI-related layoffs; the point is conditional on that market narrative taking hold. — Dwarkesh Patel Interview.
  3. On the problem with individual economic forecasts: Imas cautions that individual economists’ forecasts of large labor transitions have often been poor and that current experts disagree substantially about AI’s effects. — Dwarkesh Patel Interview.
  4. On leveraging the wisdom of the crowd: He favors prediction markets and aggregated forecasts over relying on a single economist’s labor-market prediction, while treating the future effects of AI as uncertain. — Dwarkesh Patel Interview.
  5. On missing data: Imas says forecasting AI employment effects requires better evidence on demand for newly cheap goods and on which human-involved services people will continue to value. — Dwarkesh Patel Interview.
  6. On shallow adoption despite broad reach: ATLAS finds observed AI usage across occupations representing just over 88% of US employment, but shallow task penetration in many of those occupations. — Google AI & Economy ATLAS.
  7. On recursive AI and exponential growth: Imas argues that AI improving AI research could challenge economic models built around diminishing returns, though recursive gains remain a scenario rather than an established growth path. — FeedForward Podcast.

Part 5: Rethinking Research, Policy, and Identity

  1. On the speed of AI deployment: Imas distinguishes the technical possibility of automating a task from the slower process of organizational adoption; a much faster, broader transition could still strain worker adjustment. — Think Better with Alex Imas.
  2. On the universal basic ETF: Imas considers broader capital ownership alongside income-support policies for an automated economy, but cautions that building universal capital takes time and would not immediately help workers displaced within months. — Dwarkesh Patel Interview.
  3. On academic publishing delays: Imas says his journal papers can take seven or eight years and involve extensive preemptive response to expected referees, a pace ill-suited to some fast-moving AI questions. He describes this as his own research process. — Justified Posteriors Transcript.
  4. On open science and Substack: He treats his Substack writing as a complement to academic papers: a place to share developing arguments and technical notes earlier, without treating fear of being scooped as decisive. — Justified Posteriors Transcript.
  5. On unconventional career paths: Imas says Bob Dylan’s music helped pull him away from medical school; during the ensuing period he painted in New York, and a radio interview with Richard Thaler later introduced him to behavioral economics. — Justified Posteriors Transcript.

Part 6: Behavioral Economics That Survived the Lab

  1. On the "confused subject" hypothesis: Imas recalls the “confused subject” objection: economists argued that low-stakes student experiments might not predict the behavior of experienced market participants. — Think Better with Alex Imas.
  2. On transitioning to field data: Imas points to field evidence, including professional golfers making consequential decisions, as a reason behavioral anomalies became harder to dismiss as laboratory artifacts. — Think Better with Alex Imas.
  3. On robust market anomalies: Imas says the endowment effect was studied first with inexpensive objects and has since been examined in consequential settings such as housing and options trading. — Behavioral Scientist Interview.
  4. On updating foundational research: Imas and Thaler revisited classic anomalies with replications and field evidence; he also notes that some published effects, particularly in markets, did not persist. — Think Better with Alex Imas.

Part 7: Behavioral Exploitation and Policy

  1. On corporate exploitation: Imas argues that private firms may use behavioral insights to extract value from consumers, while much academic and policy work focuses on beneficial nudges. — Behavioral Scientist Interview.
  2. On regulatory hurdles: He says regulating exploitation of biases is difficult because proving consumer harm and choosing a welfare standard for intervention are both challenging. — Behavioral Scientist Interview.
  3. On the Pareto criterion: Imas questions whether a strict Pareto test, requiring everyone to benefit, can set too high a bar when policy helps consumers but reduces a firm’s gains from a behavioral bias. — Behavioral Scientist Interview.

Part 8: Agentic Markets and Delegated Choice

  1. On agentic bargaining: In an experimental marketplace, buyer and seller agents instructed by humans produced at least as much outcome dispersion as human-to-human negotiations, rather than uniform results. — Agentic Interactions.
  2. On the power of the principal: The agentic-market experiments found that the identity and traits of the person writing an agent’s instructions can strongly predict the outcome, sometimes more than shared code or information. — Agentic Interactions.
  3. On machine fluency: The authors identify variation in “machine fluency,” or how effectively a person instructs an AI delegate to pursue their objectives, as a possible new source of unequal outcomes. — Agentic Interactions.
  4. On the persistence of behavioral frictions: Personality-related behavioral frictions reappeared in the agents’ market interactions, showing that delegation did not erase all differences among human principals in this experiment. — Agentic Interactions.
  5. On demographic echoes: Although agents did not receive a principal’s demographic details, their outcomes still varied with principal characteristics, including gender, in the experimental setting. — Agentic Interactions.
  6. On auditing bias: Imas suggests AI behavior could be easier to audit at scale than human discrimination in some settings, while noting that model behavior and its causes still require careful testing. — Justified Posteriors Transcript.
  7. On specification hazards: The authors warn that agents may pursue what a principal wrote rather than what they meant, creating information asymmetries and specification hazards in delegated markets. — Agentic Interactions.

Part 10: Technology Shocks and Labor Transitions

  1. On early automation fears: Imas uses Ricardo’s nineteenth-century machinery debate to show that automation can eliminate familiar jobs without proving that total employment must permanently collapse. — Dwarkesh Patel Interview.
  2. On the pace of structural change: Imas notes that the move away from telephone-operator work took roughly two decades; organizations and institutions can slow adoption, while a faster and broader shock would be harder to absorb. — Think Better with Alex Imas.

Part 11: Growth and Welfare After AI

  1. On infinite variety: Imas acknowledges that AI could generate such a wide variety of new machine-produced goods that consumer spending on them could remain large even if people still value human services. — Dwarkesh Patel Interview.
  2. On divergent welfare: Imas argues that under some AI scenarios, measured economic output and human welfare could diverge; higher or lower GDP alone would not settle whether people are better off. — Justified Posteriors Transcript.