These ideas are just example topics. I have deliberately not put in concrete project suggestions since I want you to think up a project idea. A good project should combine one or two anchor papers with a small model, simulation, experiment, or original example that helps explain what is going on. You do not need to use expensive APIs. It is fine to use small open models, publicly available outputs, synthetic agents, or simplified simulations, as long as the connection to the underlying economic or game-theoretic model is clear.
LLMs are becoming an interface through which users discover products, services, papers, restaurants, and software tools. This raises a natural auction-design question: if several firms want to appear in an answer, what exactly are they bidding for? A slot, a mention, a favorable summary, a citation, or some notion of prominence?
Classical anchors: Vickrey's second-price auction; Myerson's optimal auction design; generalized second-price auctions for sponsored search.
Recent AI-facing papers:
AI systems depend heavily on data, but data is unusual as an economic good: it can be copied, combined, and reused, and its value may depend on what other data is available. This makes data markets a useful place to revisit questions about exchange, fairness, and equilibrium.
Classical anchors: Arrow's discussion of information as an economic good; Shapley and Scarf's housing market model for exchange and the core.
Recent AI-facing papers:
Many AI services sell access through menus: free tiers, premium tiers, larger context windows, faster inference, better models, or usage limits. This connects to classical nonlinear pricing and price discrimination, but the product being priced is model quality or model access rather than a physical good.
Classical anchors: Mussa and Rosen on monopoly and product quality; Wilson's book on nonlinear pricing.
Recent AI-facing papers:
RLHF and preference tuning often aggregate feedback from many people. This is a social choice problem in disguise: different users may disagree, and there may be no single ranking or reward model that represents everyone well. The classical impossibility theorems are useful not because they make the problem hopeless, but because they tell us what kinds of tradeoffs to expect.
Classical anchors: Arrow's impossibility theorem; Gibbard-Satterthwaite on strategic manipulation.
Recent AI-facing papers:
Position: Social Choice Should Guide AI Alignment in Dealing with Diverse Human Feedback
Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?
An LLM that gives advice is not only answering questions. It may also be choosing what to reveal, what to emphasize, and how to frame uncertainty. This connects naturally to Bayesian persuasion: a sender has information, sends a signal, and a receiver updates their belief before taking an action. The new AI question is what happens when the sender is an AI system or AI provider whose objective may not be the same as the user's.
Classical anchors: Crawford and Sobel on strategic information transmission; Kamenica and Gentzkow on Bayesian persuasion.
Recent AI-facing papers:
Personalization Aids Pluralistic Alignment Under Competition
Friend or Foe: Delegating to an AI Whose Alignment is Unknown
Classical bargaining theory gives a clean benchmark for negotiation: a feasible set, a disagreement point, and a rule for selecting a reasonable agreement. AI negotiation asks what happens when one or both parties are represented by an LLM, or when an LLM acts as a mediator.
Classical anchors: Nash's bargaining solution; Rubinstein's alternating-offers bargaining model. Kalai-Smorodinsky bargaining is a useful alternative fairness benchmark.
Recent AI-facing papers:
How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis
Fine-Tuning Games: Bargaining and Adaptation for General-Purpose Models
LLM agents can act as buyers, sellers, bidders, competitors, or advisors. This creates a bridge between classical market models and experiments with artificial agents: do LLM agents behave like equilibrium players, boundedly rational players, cooperative agents, or something else?
Classical anchors: Cournot competition; Bertrand competition; basic first-price and second-price auction theory.
Recent AI-facing papers and resources:
Many allocation and matching problems begin by asking people what they want. LLMs could help translate natural-language preferences into rankings, utilities, or constraints. That sounds helpful, but it also creates new questions about ambiguity, strategic behavior, and whether the elicited preferences really match the person's intent.
Classical anchors: Gale and Shapley on stable matching; Shapley and Scarf on exchange.