Mandatory labeling of AI-generated content raises transparency but cuts creator surplus and can stifle high-quality outputs; platforms should move from strict policing to lighter screening as AI capabilities increase.

Main Finding

Disclosure requirements for AI-generated content are not universally welfare-improving. Mandating self-disclosure (with imperfect enforcement) is privately optimal for a platform only in an intermediate region where AI content quality and AI cost savings are neither too low nor too high. Outside that region, disclosure either needlessly suppresses adoption (when AI quality is low) or destroys valuable high-quality AI output (when AI becomes technologically advanced). As AI capability increases, the platform’s optimal enforcement policy transitions from strict deterrence → partial screening → deregulation. Disclosure raises transparency and can raise platform profit in some parameter regions, but it reliably reduces aggregate creator surplus and can lower aggregate content quality when AI is very good.

Key Points

• Model objects and mechanisms
  • Continuum of heterogeneous creators each producing one content unit; platform monetizes through engagement and retains commission r.
  • Creators choose human (cost c, quality normalized to 1) or AI production (cost δc with δ∈(0,1); AI quality v∈(0, v̄), which can be below or above human quality).
  • Platform detection accuracy is β ∈ (0,1); self-disclosure regime requires creators to declare AI use, with imperfect enforcement (penalty p when detected non-disclosure).
  • Viewers discount labeled AI content by factor f and penalize creators whose non-disclosure is detected by a trust discount k (captures algorithm aversion / credibility losses).
• Two regimes compared
  • Non-disclosure (N): platform labels only via its detection algorithm; creators not required to self-report.
  • Self-disclosure (D): creators must disclose AI use; they can conceal but face detection risk and penalties; platform chooses enforcement intensity (penalty p, detection effort).
• Main equilibrium insights
  • Disclosure reallocates AI usage across creator types rather than only changing the aggregate adoption margin: it changes strategic incentives to adopt, disclose, or conceal.
  • Disclosure is optimal (for the platform) only when AI quality v and AI cost advantage δ are at intermediate levels: it preserves trust without severely distorting productive efficiency.
  • If AI quality v is low: disclosure suppresses low-value AI adoption but does not improve content value — net welfare losses from reduced production and engagement.
  • If AI quality v is high: disclosure imposes credibility discounts and enforcement distortions that reduce efficient adoption of high-quality AI content; disclosure can lower aggregate content quality and creator surplus.
  • Enforcement policy evolves with technology: as v increases, the platform’s best response moves from strict deterrence (high penalties/detection) to partial screening (targeted enforcement) to eventual deregulation (relax disclosure), reflecting the diminishing need to police AI as it becomes high quality.
• Distributional effects
  • Disclosure increases platform transparency and can increase platform profit in some parameter regions by mitigating trust externalities.
  • Aggregate creator surplus declines under disclosure even when disclosure is privately optimal for the platform — creators bear compliance and credibility costs.
• Role of imperfect detection and trust frictions
  • Imperfect detection (β<1) creates incentives to conceal and makes enforcement endogenous; the trust penalty k magnifies the cost of being detected and therefore shapes concealment behavior.
  • Viewer discount f for labeled AI content is a key transmission mechanism: higher f (stronger viewer aversion) increases the costs of disclosure and strengthens the case for platform intervention in some regimes.

Data & Methods

• Type of analysis
  • Formal theoretical model / mechanism design-style analysis (no primary empirical dataset).
  • Comparative statics and equilibrium characterization across policy regimes and parameter values; platform chooses enforcement (penalty/detection intensity) to maximize profit.
• Core model ingredients and parameters
  • Creator choices: human (cost c, quality = 1) vs AI (cost δc, quality v).
  • Platform: revenue sharing fraction r; detection algorithm with accuracy β.
  • Viewer responses: multiplicative credibility discount f for labeled AI content; additional trust discount k when non-disclosure is detected.
  • Penalties p for detected non-disclosure; enforcement intensity is endogenous.
• Analytical approach
  • Solve creator best responses under each regime (adopt AI or not; disclose or conceal when required).
  • Determine equilibrium composition of content types and compute platform profit, creator surplus, and aggregate content quality.
  • Conduct comparative statics on v (AI quality), δ (cost advantage), β (detection accuracy), f and k (viewer/ trust frictions), and platform parameters (r, p).
  • Characterize the platform’s optimal policy as a function of technological maturity (v) and other parameters.
• Robustness & model assumptions
  • One-period, reduced-form engagement-to-revenue mapping (quality → engagement → revenue).
  • Continuum of creators and reduced-form heterogeneity (cost/ability heterogeneity abstracted in model specifics).
  • Imperfect detection modeled as single parameter β; viewer responses summarized by two discount parameters f and k.
  • Limitations: abstracts from multi-period dynamics, platform competition, endogenous viewer learning, and micro-foundations of engagement formation.

Implications for AI Economics

• For platform governance design
  • Disclosure is a strategic policy instrument, not a pure transparency good. Platforms should calibrate disclosure and enforcement based on AI capability and cost structure rather than applying blanket mandates.
  • Investment trade-offs: when detection is costly and imperfect (low β), strict disclosure enforcement may be ineffective and costly—platforms may prefer partial screening or opt for non-disclosure with detector-improvements or alternative tools (watermarking, provenance).
  • Dynamic policy: platforms should plan for evolving rules as Gen-AI improves — relaxing disclosure/penalties as v rises to avoid suppressing valuable AI output.
• For creator incentives and market structure
  • Disclosure regimes shift surplus from creators to platforms and change which creators adopt AI; this can affect the distribution of earnings, entry incentives, and content specialization (e.g., novices vs experts).
  • In contexts where AI augments quality (v>1), heavy disclosure/enforcement can reduce innovation and productive complementarities between humans and AI.
• For regulation and public policy
  • Policymakers should recognize disclosure’s ambiguous welfare effects: while transparency is valuable, mandatory self-disclosure combined with penalties can reduce overall content value and creator welfare when AI is high-quality.
  • Tailored, context-specific rules (domain sensitivity: news, health, political content vs entertainment) are preferable to one-size-fits-all labeling mandates.
  • Support for technical provenance (watermarks, cryptographic signatures) and improvements in detection accuracy may reduce the trade-offs identified.
• Directions for future research
  • Multi-period models with learning about AI quality, reputation dynamics, and endogenous investment in detection/watermarking.
  • Platform competition: how competing platforms’ disclosure policies interact and affect creator migration and social welfare.
  • Empirical validation: field experiments or platform data to measure parameters (f, k, β, δ, v) and test the model’s predicted regime boundaries and welfare implications.