In routed cities, platform competition can collapse: a game-theoretic model shows that when autonomous routing providers compete for users and prioritize market share, a weakly preferred provider captures the market in equilibrium, implying monopoly outcomes under realistic-feeling assumptions.

Main Finding

The paper introduces "competitive mediator games" — a generalization of (coarse) correlated equilibria to settings with multiple strategic mediators — and studies their equilibria in the context of future autonomous routing-and-driving (ARAD) markets. In anonymous congestion/routing games with market-share–maximizing mediators, the author proves that (in a generic setting) any competitive mediator equilibrium is a monopoly whenever one mediator is weakly preferred by all users. The framework formalizes how mediator objectives and strategy spaces shape market structure and social welfare in algorithmic mediated markets.

Key Points

• New model: Competitive mediator games extend mediated/multi-signal games by allowing several strategic mediators whose strategies (correlated recommendation distributions) are announced before users choose whether to delegate.
• User model: Each user can choose (i) an action directly (a route) or (ii) one of several mediators. Utility when using a mediator depends on the eventual routed action and on a mediator-specific discount factor (γ) capturing e.g. perceived value of time in AVs or interface preferences.
• Mediator model: Mediators are strategic players whose utility depends on which users choose them and on resulting actions (e.g., market-share maximization, or α·market-share + β·system welfare). Mediator strategies must specify recommendations for any subset of potential users (essential for equilibrium analysis).
• Equilibrium notion: Competitive mediated equilibria generalize correlated and coarse correlated equilibria to multiple competing mediators and both finite and non-atomic (continuum) player populations. These equilibria are relevant because they can be limits of no-regret multi-agent learning.
• Main theoretical result: In anonymous congestion/routing games, when mediators maximize market share and one mediator is weakly preferred by every user, all competitive mediator equilibria collapse to monopolies (one mediator gets the entire market).
• Mechanism-design insight: If mediator revenue is purely market-share-based, mediators may adopt undesirable strategies that harm welfare. Incorporating overall user welfare into mediator objectives can preserve competition and discourage such strategies (as suggested by prior work [28] and discussed in the paper).
• Application context: The model is applied to ARAD markets, with discussion of different market forms (nascent, intermediate, mature), routing/driving bundling, conspicuity and efficiency of AVs, and distributions of user attitudes (enthusiasts vs skeptics). Several example mediator strategies are enumerated (fixed assignment, per-user personalization, randomized splits, etc.).
• Normative concerns: The paper raises policy questions about allowing ARAD, number of competing providers, market rules to incentivize welfare-improving behavior and to curb algorithmic anti-competitive practices.

Data & Methods

• Methods: The paper is theoretical and analytical. It develops new game-theoretic definitions and equilibrium concepts (competitive mediator games, competitive mediated equilibria) for both finite-player and non-atomic (infinite) settings, and proves structural results about equilibria in anonymous congestion/routing games.
• Modeling assumptions:
  • Players (users) choose either direct actions (routes) or mediators; if a mediator is chosen, the mediator draws a recommendation and implements an action.
  • User utility functions depend on ultimate actions and on the mediator used (capturing differing perceived value of travel time and interface preferences). In routing models: uf_i = γ^f_i · τ_route(q_route) where τ is delay as a function of route load.
  • Mediator utilities are specified exogenously (e.g., Uf = market share; or Uf = α·market-share + β·sum_a q_a τ_a(q_a)).
  • Mediator strategies must be defined for every possible subset of users who may opt in (not just for the realized set).
  • Analysis focuses on anonymous congestion/routing games (delay depends only on route loads), and considers generic settings (to avoid degenerate ties).
• Proof techniques: Game-theoretic equilibrium constructions and comparative statics in nonatomic/anonymous congestion frameworks; existence and structural characterization arguments showing monopoly outcomes under the weak-preference + market-share objective assumptions.
• Empirical data: None used. The paper uses illustrative examples and stylized parameterizations (e.g., discount factor distributions) to motivate modeling choices.
• Related theory: Connects to correlated/coarse correlated equilibria and to algorithmic no-regret learning (as mediators’ strategies and resulting equilibria can be learning limits).

Implications for AI Economics

• Market design matters: The mediator revenue objective (market-share vs. incorporating welfare) strongly determines market structure and social outcomes. Pure market-share incentives can lead to monopolization even with heterogeneous user preferences, calling for careful regulator design of payments or rules.
• Algorithmic competition risks: When strategic algorithms act as mediators and can commit to recommendation strategies, they can produce anti-competitive market dynamics (effective monopolies, coordinated patterns) without explicit collusion. Antitrust frameworks that focus on price alone may miss such risks in algorithm-driven service markets.
• Welfare vs. competition tradeoffs: Adding system-welfare components to mediator payoffs can mitigate undesirable strategies and preserve meaningful competition, suggesting policy levers (subsidies, platform requirements, or procurement contracts) to align providers’ incentives with public objectives.
• Design of mediated markets: Regulators and platform designers should require transparency about mediator strategies (or constrain feasible strategy classes), ensure diversity of offerings, and consider mechanisms that tie provider revenue to social metrics (e.g., congestion externalities) to avoid single-provider lock-in.
• Broader relevance: The competitive mediator framework applies beyond ARAD to other AI-mediated markets (e.g., recommendation platforms, algorithmic pricing, routing on networks), highlighting the need to model mediators as strategic economic agents with complex objectives.
• Research directions: Further work should analyze dynamic and out-of-equilibrium learning processes, robustness to richer mediator objectives, empirical calibration with user preference distributions and real traffic models, and mechanism-design interventions that prevent monopolization while preserving innovation incentives.
• Policy takeaway: Antitrust and regulation need to adapt to algorithmic mediators — monitoring objective functions and the ability to commit to recommendation distributions is crucial to predict and steer market outcomes in AI-dominated service markets.