Core Uses of AI for Economic Theory (Markus Academy 166-1)

Author/Source: Pietro Ortoleva (Princeton), with Fedor Sandomirskiy — Markus' Academy Ep. 166-1 (1st of 5), hosted by Markus Brunnermeier
Original: https://www.youtube.com/watch?v=kVVggnIaw2Q
Key Ideas
Economic theory is a fixed-point search, not a proof of a conjecture. Mathematicians set out to prove a stated claim. Theorists develop a model whose assumptions yield true statements they can certify — and both the assumptions and the statements are adjustable. You start with a model, try to prove something, discover you need to tweak the assumptions, and iterate until assumptions and statements agree. AI's contribution is to make each loop of that iteration cheap, letting you converge to the fixed point faster.
Seven use cases: (1) sketching and brainstorming models, (2) literature review, (3) suggesting proofs, (4) checking proofs, (5) extensions, microfoundations, and simplifications, (6) general proofreading, (7) simulations.
Sketching is the most underused. Ortoleva's worked example: you suspect political polarization can be generated by rational inattention and costly information acquisition across policy dimensions. Prompt the model for multiple models that generate the intuition, and ask it to judge which is most elegant, which is most fragile and assumption-dependent, and which points toward theorems worth proving. Especially valuable at project start — fixing primitives, timing, solution concepts, and what predictions to aim for.
Two adjacent sketching moves: spinning off variants (change the timing, change the information structure, check which comparative statics survive), and stripping down to the smallest model that still delivers the prediction. Ortoleva invokes a Princeton colleague known for always asking "can you do it with two types and two periods?" — AI answers that question quickly, and there is "a huge price in theory for clean simple models."
On proofs, frontier models are reliable for the kind of proof economists typically write — meaning the proofs are usually correct, not that they are submittable. The prose is notation-heavy, over-long, and introduces extraneous definitions; it needs pruning, and the author retains full responsibility for checking it. Non-frontier models are not reliable end-to-end: with those, decompose the task into steps and have the model do one step at a time.
Three higher-value proof uses than proof generation:
- Attack — an agent playing hostile referee, hunting for weak steps.
- Repair — when the model refuses because the statement is false, ask which assumption would make it true. This is the fixed-point loop run directly: "AI for this is amazing."
- Inspiration — a wrong proof can point to an attractive route you hadn't considered. Both authors independently experienced this: an incorrect AI proof suggested an approach neither had thought of; patching it across several models produced the real proof.
The rabbit-hole risk is real and scales with your own ignorance. Ortoleva lost two entire weekends to a polished-looking proof of a statement he knew to be false — adversarial AI couldn't find the flaw either, and feeding in his counter-example wasn't enough. His warning: "AI is becoming better at fooling you," so proofs in mathematics you don't personally understand — invoking theorems or whole branches you can't check — are dangerous territory.
Extensions, microfoundations, and simplifications are the cheap auxiliary wins. Microfoundations especially: "it's so quick... if you're not doing it you're leaving money on the table."
An editor's worry, from an editor. Ortoleva flags the flip side of cheap extensions: as an editor he is "very worried of receiving a lot of papers that are incredibly long and full of AI-generated extensions." Some extensions genuinely probe robustness; mass-produced ones just inflate manuscripts.
Audience is deliberately broad — graduate students to senior colleagues, occasional chatbot users to people running agents — and extends to mathematicians and anyone doing formal research, including formal political theory.
Summary

The first episode of the Markus Academy theory mini-series is Ortoleva's map of where AI fits into theoretical research. Its value comes from the framing rather than the tool list: economic theory, unlike mathematics, is a search for a fixed point. The theorist does not receive a conjecture and go prove it; they build a model, attempt a result, find the result fails or is uninteresting, adjust the assumptions, and repeat until assumptions and statements are mutually consistent and the whole thing illuminates a mechanism. Every theorist already knows this. The point is that AI does not short-circuit the search — it lowers the cost of each iteration, which is where the gains come from.

That framing determines which use cases matter most. Ortoleva argues sketching is the most underused, and it is the one that most directly attacks iteration cost: hand the model a vague intuition, ask for three minimal models that capture it, and ask it to grade them for elegance, fragility, and which theorem would be worth proving. The same move generates variants (perturb the timing or information structure and see which comparative statics survive) and simplifications (find the smallest model that still delivers the prediction). Proof generation, the use case most people ask about first, turns out to be the less interesting one. Frontier models can produce correct proofs of typical economic models, but the output is verbose and notation-heavy and needs editing; the responsibility for verification stays with the author.

The genuinely high-value proof uses are adversarial or diagnostic. Attack puts the model in the hostile-referee role. Repair exploits the fixed-point structure directly — when the model says a statement is false, the productive question is not "try again" but "what assumption would make this true," which is precisely the move a theorist makes by hand. Inspiration is the most surprising: a wrong proof that suggests an unfamiliar route can be more valuable than a correct one, because it breaks the author out of habitual approaches. Ortoleva balances this with a genuinely cautionary anecdote — two weekends lost to a polished false proof that neither he nor an adversarial model could break, on a statement he already knew was false. His generalized warning is that the danger scales with your own ignorance of the area, which makes AI-generated proofs invoking unfamiliar branches of mathematics the highest-risk case.

Relevance to Economics Research

For theorists, this is the most directly actionable framing available. The seven use cases are a checklist, and the sketching prompt is copy-pasteable. But the fixed-point framing is what generalizes beyond theory: it identifies iteration cost as the thing AI reduces, which is a sharper predictor of where AI helps than task difficulty. Empirical work has the same structure — specification, identification assumption, result, revise — and the same conclusion follows: the gain is in how many candidate specifications you can seriously consider, not in the model doing the econometrics for you.

The repair move deserves attention from empiricists too. Reframing a failure ("this doesn't hold") into a diagnostic ("what would have to be true for it to hold") is exactly what turns a rejected result into a sharpened assumption. And the attack pattern — an agent instructed to be a hostile referee — is the same adversarial-verification discipline that recurs across this wiki, from Cunningham's Referee 2 to multi-agent DiD auditing.

Ortoleva's editorial concern about AI-generated extensions is a concrete instance of the manuscript-inflation problem: when the marginal cost of an extension falls to near zero, the constraint on paper length stops being effort and starts having to be editorial judgment. That is the theory-side version of the supply-shock argument in summaries/cc-series-27-research-vs-publishing.