Labeling content as AI-generated or AI-enhanced reduces user engagement, particularly for emotional posts; delaying disclosure helps only for AI-enhanced content, not fully for AI-generated pieces.

Main Finding

Labeling visual social-media content as AI-generated or AI-enhanced reduces both affective and behavioral engagement relative to labeling it human-created, with the effect strongest for emotional content. Delaying disclosure (late vs. early) increases affective engagement for AI-enhanced content but does not restore engagement for fully AI-generated content.

Key Points

• Experimental manipulations: level of AI involvement (human-created, AI-enhanced, AI-generated), content type (emotional vs. rational), and disclosure timing (early vs. late).
• Two online experiments on Prolific: study 1 (n = 325), study 2 (n = 371); participants mean age = 35; 55% female.
• Outcome measures: affective engagement (attitudinal responses) and behavioral engagement (intended actions toward the profile/content).
• Main pattern:
  • AI labels (both enhanced and generated) reduced affective and behavioral engagement versus human-created labels.
  • Reduction was larger for emotional content than for rational content.
  • Late disclosure mitigated the negative effect only for AI-enhanced content’s affective engagement; it did not help for AI-generated content.
• Interpreted as evidence of algorithmic aversion in the context of generative AI, especially when content appeals to emotions.

Data & Methods

• Design: two preregistered online randomized experiments (between-subjects manipulations).
• Sample: participants recruited via Prolific; combined N = 696 across studies; demographic snapshot provided (mean age 35, 55% female).
• Stimuli: Instagram-style profiles presenting visual content; labels indicated whether content was human-created, AI-enhanced, or AI-generated.
• Factors:
  • AI involvement: three levels (human, AI-enhanced, AI-generated).
  • Content type: emotional vs. rational appeals.
  • Disclosure timing: early (label shown upfront) vs. late (label shown after initial exposure).
• Outcomes: self-reported affective engagement (e.g., liking, warmth, trust) and behavioral engagement (e.g., intentions to like/follow/share — reported intentions rather than real platform behavior).
• Limitations: online convenience samples, artificial profile exposure, self-reported engagement (not observed real-world behavior), single platform and visual-content focus.

Implications for AI Economics

• Creator incentives and monetization:
  • Labeling reduces engagement and thus may lower creators’ returns (ad revenue, sponsorships, follower growth) for AI-labeled content.
  • Creators face a trade-off: use generative tools to reduce production costs but risk lower engagement if labeled.
  • Concealment incentives: negative engagement effects could encourage creators to obscure AI involvement, raising enforcement and verification challenges.
• Platform design and revenue:
  • Platform-level labeling policies can materially affect user engagement metrics and downstream monetization (ads, subscriptions).
  • Timing of disclosure matters: allowing or engineering later disclosure for AI-enhanced content could partially preserve engagement while maintaining some transparency—though this raises ethical and regulatory questions.
• Regulation and policy trade-offs:
  • Mandatory labeling increases transparency but may impose negative economic externalities (reduced engagement, lower creator incomes).
  • Policymakers should weigh consumer-protection goals against potential dampening of creator and platform revenues; nuanced rules (e.g., distinguishing levels of AI involvement, or disclosure timing/format standards) may be more efficient than blunt mandates.
• Market dynamics and content composition:
  • If emotional content suffers more from AI labels, creators may shift toward rational or informational content formats, changing the overall content mix and consumer welfare.
  • Platforms could see changes in supply-side behavior, with differential adoption of generative tools across creator types and genres.
• Research and modeling priorities:
  • Need for economic models of incentives under disclosure regimes (dynamic entry, signaling, enforcement costs).
  • Empirical follow-up: natural-field experiments measuring actual engagement and revenue impacts, heterogeneity across user segments, and long-run behavioral adjustments.
• Practical recommendations:
  • Platforms and policymakers should consider graduated labeling (distinguish AI-enhanced vs AI-generated) and experiment with disclosure timing/format to balance transparency with engagement.
  • Invest in verifiable audit mechanisms to prevent concealment while reducing blunt engagement losses (e.g., reputational certification, contextual disclosures).