A simple transparency dial on who you see can reshape strategic thinking: when platforms hide global information people overestimate others’ sophistication and escalate competition, while full visibility restores cognitive-calibrated behavior and stabilizes coordination.

Main Finding

The paper introduces the Connected Minds model: a unified behavioral framework showing that network-mediated observation (parameterized by a locality/transparency parameter p ∈ (0,1]) systematically distorts agents’ beliefs about the distribution of opponents’ cognitive sophistication. p continuously interpolates between Level-k (p → 0, local/insular views) and Cognitive Hierarchy (p = 1, global/accurate views). Network opacity induces a persistent “Sophisticated Bias” — agents overestimate opponents’ depth of reasoning — and, depending on game structure, tuning p can either increase aggregate effort (Escalation Principle in games with strategic complements) or improve stability and reduce variance (Transparency Reversal in coordination games). Even as cognitive depth grows, topology can generate persistent miscalibration (Poisson-Shift Convergence).

Key Points

• Connected Minds augments Cognitive Hierarchy (CH) by making an agent’s belief about lower-level types a network-weighted truncation: g_k(h) ∝ p^{k−h} f(h) for h < k, with normalization. p captures effective reach/visibility.
• Interpretations of p:
  • Cognitive homophily (stochastic block / geometric decay of links by cognitive distance).
  • Algorithmic ranking / attention (rank-biased decay in recommender feeds).
  • Information diffusion (random-walk with restart / conductivity of information).
• Limits:
  • p → 0: belief mass concentrates on type k−1 → reproduces Level-k behavior (local echo-chamber).
  • p = 1: uniform weights → recovers standard CH (accurate conditional beliefs).
• Theoretical properties:
  • Poisson-Shift Convergence Theorem: for Poisson-distributed types, perceived mean sophistication shifts to τ/p; hence high individual sophistication does not eliminate topology-induced bias.
  • Monotone comparative statics: changing p induces monotone shifts in beliefs while preserving regularity such as log-concavity of belief distributions.
  • Sophisticated Bias: network opacity leads agents to systematically overestimate opponents’ reasoning depth.
• Welfare/mechanism-design results:
  • Escalation Principle: in games with strategic complements (e.g., effort, adoption), reducing p (more opacity) can increase aggregate effort because agents compete against a (perceived) higher-sophistication population.
  • Transparency Reversal: in coordination games (e.g., beauty contest), increasing p (more transparency) is needed to reduce belief variance and stabilize outcomes; opacity raises payoff dispersion and instability.
• Microfoundations: the exponential kernel p^{k−h} is derived from plausible structural models (homophilous link probabilities, rank-biased attention, random-walk information flow).
• Modeling assumptions: level-0 uniform randomization, agents are overconfident (assign zero mass to types ≥ k), p ∈ (0,1], truncation to support {0,...,k−1}. p is treated as a reduced-form policy/engineering parameter but admits structural grounding.

Data & Methods

• This is a theoretical / analytic paper; no empirical dataset is used.
• Methods employed:
  • Formal model building: embedding CH/Level-k into a network-observation framework via kernel weights.
  • Analytical derivations and comparative statics for general f(h) distributions (examples with geometric and Poisson).
  • Measure-theoretic characterization to show preservation of distributional regularity (e.g., log-concavity) under biasing.
  • Asymptotic analysis (Poisson-Shift Convergence) to show topology-driven miscalibration persists as k → ∞.
  • Mechanism-design analysis: a “Cognitive Designer” chooses p (globally or personalized p_k) to optimize welfare objectives under different game structures.
  • Microfoundational constructions: stochastic block models, rank-biased attention models, and random-walk diffusion used to justify the exponential kernel.
• Limitations noted by authors:
  • p is modeled reduced-form; empirical calibration / identification left for future work.
  • Agents truncate beliefs at k (consistent with CH axioms); dynamic learning and time-evolving networks are not the primary focus here.

Implications for AI Economics

• Platform design and algorithmic transparency are active levers with predictable, structure-dependent effects:
  • Recommender systems and ranking algorithms (which effectively set p via rank decay and diversity settings) can alter equilibrium behavior in strategic environments. Designers can increase or decrease aggregate activity or stabilize outcomes by adjusting visibility.
  • In settings where aggregate effort is desirable (markets with strategic complements, adoption cascades, investment/effort contests), reducing transparency (lower p) may intentionally induce competition and raise total effort — but potentially at the cost of increased inequality, instability, or welfare dispersion.
  • In coordination/forecasting environments (prediction markets, consensus tasks, beauty-contest–type settings), high transparency (higher p) reduces variance and payoff inequality; opacity can destabilize coordination and worsen aggregate accuracy.
• Policy and regulation:
  • Algorithmic transparency rules are not unambiguously welfare-improving: optimal disclosure depends on the game-theoretic environment. Regulators should consider the function of the platform (e.g., promoting adoption vs. stabilizing coordination) before mandating transparency levels.
  • Personalization of visibility (p_k) offers a targeted policy tool: platforms could tune visibility by segment to achieve nuanced welfare objectives (e.g., more visibility to low-sophistication users in coordination tasks).
• Empirical predictions and tests:
  • Observed behavior that looks like Level-k vs CH should systematically vary with measures of network homophily or recommender diversity (i.e., estimated p).
  • Experimental manipulations of feed-ranking diversity or local neighborhood sampling should shift depth-of-reasoning measures and aggregate outcomes in predicted directions (escalation under complements; stabilization under coordination).
  • Calibration: estimate p from observed action distributions using CH-style structural inference extended to include p as a parameter (e.g., fit p such that g_k(h) explains observed actions).
• Broader applications in AI economics:
  • Multi-agent systems and marketplaces where agents’ beliefs are shaped by algorithmic information flows (crowdfunding, gig platforms, social trading) should account for topology-induced belief bias when predicting adoption, bidding, or effort.
  • Strategic manipulation risk: platforms or adversarial actors could exploit p to bias markets or collective behavior (e.g., artificially creating echo chambers to increase competition or engagement).
• Research directions:
  • Empirical validation: lab and field experiments to estimate p and test the Poisson-Shift prediction.
  • Dynamics and learning: extend the model to time-evolving networks and repeated interaction where beliefs update endogenously.
  • Heterogeneous/strategic platforms: study endogenous choice of p by competing platforms or profit-maximizing recommenders.
  • Welfare trade-offs and regulation: formalize social welfare functions reflecting trade-offs between aggregate output, inequality, and stability; design constraints for platform policy.

Summary takeaway: network transparency (captured parsimoniously by p) is a cognitive-design lever with predictable, game-structure–dependent effects. For AI economists, this links algorithmic ranking and information architecture explicitly to collective strategic outcomes, generating both prescriptive tools and cautionary implications about the power of platform-mediated visibility.