When labeling AI art is optional, some creators hide low-quality AI pieces to pass them off as human-made; buyers learn to detect this and punish offenders, cutting their sales but also exposing a decline in average quality as moral hazard rises.

Main Finding

Voluntary AIGC tagging on a major digital-art marketplace generated strategic behavior: about 31% of artists who used AI opportunistically withheld the AIGC tag on low-quality AI images to present them as human-created. Consumers and platform signals, however, tended to detect and penalize this opportunism—reducing sales of opportunistic artists and thereby mitigating adverse selection—while the opportunistic strategy was also associated with lower artwork quality overall, indicating heightened moral hazard.

Key Points

• Dataset and prevalence

  • Final sample: ~3,243 artists and >28,000 artworks posted in the year before and after a voluntary AIGC tag launch (initial raw data: ~85k artworks, ~8k artists).
  • Artist classification after the tag: 588 opportunistic artists (created AIGC but withheld tags), 688 honest AIGC-disclosing artists, 650 original-only artists.
  • Roughly 31% of artists who created AIGC after the tag behaved opportunistically by concealing the AIGC label.

• Strategic tagging behavior

  • Opportunistic artists tended to label higher-quality AIGC honestly and withhold tags on lower-quality AIGC (i.e., selective disclosure).
  • More experienced artists and artists with higher platform traffic were less likely to disclose AIGC (suggesting stronger incentives to withhold when stakes are higher).
  • Honest AIGC creators produce high-quality UGC as well, but their UGC after disclosing AIGC saw almost no sales—evidence of reputational spillover from AIGC labeling.

• Market outcomes: adverse selection and moral hazard

  • Consumers penalized opportunists: opportunistic artists had the lowest sales probability and sale volume among groups (e.g., mean sale probability ~0.002 for opportunists vs. ~0.027 for honest AIGC artists and ~0.006 for original artists).
  • Labeled AIGC from opportunistic artists underperformed labeled AIGC from honest artists—consistent with a trust penalty.
  • Opportunistic behavior correlated with lower artwork quality (measured computationally), implying moral hazard: strategic mislabeling coincided with quality degradation.

• Consumers detect opportunism

  • Evidence suggests buyers infer or detect opportunistic/misleading behavior (partly via observable quality/aesthetic cues), which reduces opportunists’ market success and thus reduces adverse selection pressures despite voluntary disclosure.

Data & Methods

• Platform and policy context

  • Study context: a large digital-art marketplace ("ABC") introduced a voluntary AIGC tag for marketplace uploads (tag is suggested at upload and displayed in artwork metadata).
  • Time window: one year before and one year after the tag introduction (tag introduced Feb 9, 2023).

• AIGC detection (relabeling ground truth)

  • Used multiple image detectors: LGard (GAN artifact detection), UnivFD (generalizable diffusion-model detector), and the Illuminarty web tool (AI probability score).
  • Classification rule: an artwork counted as AIGC if at least two detectors flagged it as AI-generated.

• Objective quality measurement

  • Three automated aesthetic metrics:
    • Color richness: entropy-based measure from RGB histograms.
    • Compositional aesthetics: Alibaba Cloud Visual Intelligence API score.
    • Design complexity: multiscale renormalization structural complexity.
  • These three were combined into a composite quality score using TOPSIS; also retained rank-based quality measures.

• Empirical strategy

  • Artist classification into opportunistic / honest / original based on post-tag behavior.
  • Logistic regressions (and subsample regressions) to analyze determinants of truthful disclosure; key result: higher objective quality strongly increases probability of truthful AIGC disclosure (e.g., Quality topsis large positive coefficient).
  • Staggered difference-in-differences (DiD) with propensity-score matching (PSM) to estimate causal impacts of opportunistic behavior on sales outcomes across UGC and AIGC markets.
  • Additional computational image analysis to corroborate that consumers could detect speculative behavior via visual cues.

Implications for AI Economics

• Endogenous disclosure and signal credibility

  • Voluntary tags function as a low-cost quality signal whose credibility depends on creators’ disclosure choices. Even without perfect detectors, voluntary labels can partially work because consumers infer honesty from observable quality signals; modelling disclosure as endogenous is important for theory and empirical work.

• Adverse selection vs. moral hazard trade-off

  • Voluntary tagging coupled with consumer inference can mitigate adverse selection (buyers penalize detected opportunists), but the opportunity to misreport creates moral-hazard incentives—opportunists produce lower-quality outputs. Models of platform markets with AIGC should incorporate both forces simultaneously.

• Platform policy design

  • Relying solely on voluntary disclosure is insufficient: it incentivizes selective nondisclosure by experienced/high-traffic creators. Platforms should weigh options such as mandatory labeling, automated detection, provenance verification, or targeted monitoring of actors with strong incentives to misreport.
  • Because honest AIGC creators can suffer reputational spillovers that reduce sales of their human-made works, policymakers and platform designers should consider mechanisms that prevent unjust punishment of honest creators (e.g., provenance certificates, verified creator tags).

• Welfare and market structure

  • Even imperfect tags affect sorting between UGC and AIGC markets and buyer beliefs—this alters competition dynamics, price formation, and labor incentives for human creators. Economic analyses of AI adoption in creative markets should account for these second-order effects on supply (entry/exit) and incentives to invest in quality.

• Research directions

  • Develop structural models where disclosure choices, detection technology quality, buyer inference rules, and seller quality investments interact.
  • Evaluate costs/benefits of detection investments and enforcement regimes (mandatory vs voluntary) on welfare, artist incomes, and content quality.
  • Study cross-platform spillovers and how reputational channels propagate when creators operate across multiple marketplaces.

If you want, I can (a) extract the key regression coefficients and treatment-effect estimates into a short table, (b) sketch a simple theoretical model that formalizes the disclosure–adverse-selection–moral-hazard trade-off implied by the paper, or (c) outline policy experiments a platform could run to reduce opportunistic nondisclosure. Which would be most useful?