📈 economics

When Is Self-Disclosure Optimal? Incentives and Governance of AI-Generated Content

This paper develops a formal model demonstrating that mandatory self-disclosure of AI-generated content is economically optimal only at intermediate levels of AI capability and cost savings, as advanced AI capabilities eventually render strict disclosure regimes counterproductive by suppressing high-quality content and reducing creator surplus.

Original authors: Juan Wu, Zhe, Zhang, Amit Mehra

Published 2026-01-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Juan Wu, Zhe, Zhang, Amit Mehra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive digital town square (like YouTube or Instagram) where people come to share stories, art, and videos. For years, everyone made these things by hand. But now, a new, super-fast robot assistant (Generative AI) has arrived. This robot can make content much cheaper and faster, and sometimes it's even better than what humans make.

However, the town square's managers (the platforms) are worried. The townspeople (viewers) generally prefer human-made stories because they feel more "real" and trustworthy. If they find out a story was made by a robot, they might lose interest. But if the robot is really good, the townspeople might actually enjoy it anyway.

The managers have a big problem: How do they handle the robot content?

This paper explores two main ways the managers can run the town:

The Two Rules of the Town

The "Silent Detective" Rule (Non-Disclosure): The managers try to use a magic scanner to spot robot-made content. If the scanner catches it, they label it. If the scanner misses it, the content stays hidden. The creators don't have to say anything; they just hope the scanner doesn't catch them.
The "Honesty Box" Rule (Self-Disclosure): The managers tell creators, "You must raise your hand and say, 'I used a robot to make this!'" If you don't raise your hand and the scanner catches you later, you get a fine (or your content gets hidden). If you do raise your hand, you get no fine, but the townspeople might still be a little skeptical because they know a robot helped.

The Big Discovery: It Depends on the Robot's Skill

The paper's main finding is that there is no single "best" rule. The right choice depends entirely on how good the robot is and how much money it saves the creators.

Think of the robot's quality as a dial that can be turned up or down.

Scenario A: The Robot is Clunky (Low Quality)
If the robot makes bad stuff, the managers should not force people to label it. Why? Because if you force labels, nobody will use the robot at all. Since the robot makes bad stuff anyway, that's fine. But if you don't label it, the robot might still be used by some people, but the bad content gets filtered out naturally because people don't like it.
- Analogy: If a new food truck sells burnt toast, you don't need a sign saying "Burnt Toast." People won't buy it anyway. Forcing a sign just stops the few people who might have tried it, but it doesn't help the town.
Scenario B: The Robot is a Masterpiece (High Quality)
If the robot is amazing and makes better stuff than humans, the managers should stop forcing labels. Why? Because if you force a label, the townspeople will see "Made by Robot" and skip it, even though it's great. This hurts the creators and the platform. The managers should just let the robot do its thing without the "robot" label getting in the way.
- Analogy: If a robot chef makes the best pizza in the world, putting a giant "ROBOT CHEF" sign on the menu might scare people away. It's better to let the pizza speak for itself.
Scenario C: The Robot is "Just Okay" (Intermediate Quality)
This is the sweet spot. The robot is good enough to be useful and save money, but not so perfect that it scares everyone away. Here, the Honesty Box (Self-Disclosure) is the best strategy.
- How it works: It acts like a filter. Creators who make "robot stuff" that is sensitive to trust (like news or serious advice) will choose to label it so they aren't caught lying. Creators who make fun stuff (like memes) might hide it. This sorting helps the platform keep the "good" robot content while filtering out the "risky" robot content that people hate.

The Cost of Honesty

The paper also points out a tricky side effect. While forcing creators to be honest makes the town square more transparent, it actually hurts the creators' wallets.

The Platform Wins: The platform makes more money because they can manage trust better.
The Creators Lose: Creators end up paying a "trust tax." They either lose money by labeling their content (because people click less) or they risk fines by hiding it. Even when the policy is good for the platform, it often takes value away from the people making the content.

The "Enforcement" Dance

The paper also looks at how strictly the managers should punish people who lie.

When the robot is new/weak: The managers should be very strict. "If you use the robot and don't say so, you're out!" This stops people from using bad robot tools secretly.
When the robot is very advanced: The managers should stop punishing people so hard. If the robot is amazing, punishing people for using it (even secretly) just stops the best content from being seen. The managers should let the creators decide whether to be honest or not, because the content is so good it doesn't matter as much.

The Bottom Line

The paper concludes that disclosure is a tool, not a magic wand.

It works best when the technology is in the middle—good enough to be useful, but not so perfect that it needs no rules.
It fails when the technology is too bad (it just stops people from trying) or too good (it stops the best content from being seen).
Most importantly, while transparency sounds nice, it often shifts the burden of "trust" onto the creators, making them poorer even as the platform gets richer.

In short: Don't force a "Robot Made This" sign if the robot is terrible (nobody cares) or if the robot is a genius (people will love it anyway). Only force the sign when the robot is "pretty good," because that's when the sign helps sort the wheat from the chaff.

Technical Summary: When Is Self-Disclosure Optimal? Incentives and Governance of AI-Generated Content

1. Problem Statement

The proliferation of Generative AI (Gen-AI) on digital platforms presents a governance dilemma: while AI reduces production costs and scales content output, it raises concerns regarding authenticity, credibility, and user trust. Platforms currently rely on two primary mechanisms to manage this:

Algorithmic Detection: Imperfect and costly systems that identify AI content without creator input.
Self-Disclosure Policies: Mandates requiring creators to label AI usage, backed by penalties for non-compliance.

The core economic question addressed is under what conditions a platform should mandate self-disclosure versus relying solely on detection. Specifically, the paper investigates how disclosure regimes reshape creator incentives (adoption, concealment, and truthful labeling), affect platform profitability, and alter aggregate content quality and creator surplus.

2. Methodology

The authors develop a formal game-theoretic model involving a continuum of heterogeneous creators and a platform. The model captures the strategic interaction between platform policy, creator behavior, and viewer perception.

Model Components

Agents:
- Platform: Chooses the moderation regime (Non-Disclosure vs. Mandatory Self-Disclosure) and sets penalties ( $p$ ) for non-compliance. The platform retains a commission rate $r$ .
- Creators: Heterogeneous in their sensitivity to audience credibility (modeled via a credibility parameter $f$ ). They choose between Human-Generated ( $N_{AI}$ ) and AI-Generated ($AI$) content. Under disclosure, they further choose to truthfully disclose or conceal AI usage.
- Viewers: Derive utility from content quality but apply a credibility discount ( $f$ ) to content labeled as AI. If a creator conceals AI usage and is detected, viewers apply an additional trust penalty ( $k$ ).
Technologies:
- Human Content: Cost $c$ , normalized quality 1.
- AI Content: Reduced cost $\delta c$ (where $\delta \in (0,1)$ ) and variable quality $v$ .
Detection: The platform's algorithm detects AI content with probability $\beta$ (imperfect detection).
Regimes Compared:
1. Non-Disclosure ( $N$ ): Creators are not required to label; the platform labels detected AI content.
2. Mandatory Self-Disclosure ( $D$ ): Creators must label. Non-disclosure triggers penalties ( $p$ ) and trust penalties ( $k$ ) if detected.

Analytical Approach

The authors solve for the Subgame Perfect Equilibrium using backward induction:

Viewer Response: Determine engagement based on labels and detection outcomes.
Creator Strategy: Creators maximize expected utility given the regime, penalties, and their specific credibility parameter $f$ . This leads to cutoff-based segmentation of creators.
Platform Optimization: The platform selects the regime and penalty level to maximize profit, comparing outcomes across different ranges of AI quality ( $v$ ) and cost efficiency ( $\delta$ ).

3. Key Contributions

The paper contributes to the literature on platform governance and AI-enabled production in three specific ways:

Endogenous Incentive Mechanism: Unlike prior work treating disclosure as a passive transparency requirement, this model frames disclosure as a strategic instrument that actively shapes creator decisions regarding adoption, concealment, and content positioning.
Integration of Trust Frictions: The model uniquely integrates imperfect detection, viewer discounting of AI, and heterogeneous creator credibility to analyze how trust externalities interact with technological efficiency.
Unified Economic Perspective: It provides a comparative analysis of platform policy design and creator behavior, extending information disclosure models to the Gen-AI context to show how disclosure reallocates AI usage across creator types.

4. Key Results

Optimal Disclosure Conditions

Disclosure is not universally optimal. The platform adopts a mandatory self-disclosure regime only when both the intrinsic value of AI content ( $v$ ) and its cost-saving advantage (low $\delta$ ) are intermediate.

Low AI Quality/Cost Savings: Disclosure suppresses adoption without improving content value; non-disclosure is preferred.
High AI Quality: Disclosure becomes counterproductive. The credibility discounts and enforcement costs distort efficient adoption, reducing aggregate content quality. Non-disclosure is preferred to allow high-value AI to flourish.
Intermediate Region: Disclosure acts as an effective screening device, limiting adoption among authenticity-sensitive creators while preserving aggregate quality.

Evolution of Enforcement Strategy

As AI capability improves, the platform's optimal enforcement strategy evolves:

Strict Deterrence: When AI quality is low, high penalties are used to prevent concealment and maintain transparency.
Partial Screening: When AI quality is moderate, the platform sets intermediate penalties. This trades off screening low-credibility creators against retaining high-value AI content, allowing some strategic concealment to persist.
Deregulation: When AI quality is sufficiently high, the optimal penalty drops to zero ( $p^*=0$ ). The platform allows creators to sort endogenously between concealment and disclosure, as penalties would unnecessarily reduce the realized value of superior AI content.

Impact on Stakeholders

Platform Profit: Can increase under disclosure in the intermediate region by mitigating trust externalities. However, profit exhibits non-monotonic behavior (U-shaped or W-shaped) relative to AI quality depending on cost savings.
Creator Surplus: Self-disclosure reduces aggregate creator surplus ( $CS_D < CS_N$ ). While it increases transparency, it shifts compliance and credibility costs onto creators.
Transparency vs. Quality:
- Disclosure strictly increases transparency ( $T_D > T_N$ ).
- Disclosure improves aggregate content quality only when AI is technologically inferior to human creation ( $v < 1$ ).
- When AI is superior ( $v > 1$ ), disclosure reduces aggregate content quality by suppressing the adoption of high-quality AI content due to credibility discounts.

5. Significance and Claims

The paper concludes that disclosure is a strategic governance instrument rather than a neutral transparency mechanism. Its effectiveness is contingent on the technological maturity of AI and the magnitude of trust frictions.

Strategic Reallocation: Disclosure does not merely shift adoption margins; it fundamentally changes the composition of content by reallocating production technologies across creator types.
Trade-offs: There is a fundamental trade-off between transparency and productive efficiency. While disclosure reduces informational opacity, it can distort efficient AI adoption, potentially suppressing high-quality content and reducing creator incentives.
Policy Implication: The results suggest that a "one-size-fits-all" disclosure mandate may be suboptimal. Platforms should calibrate disclosure and enforcement intensity based on the specific quality and cost-efficiency characteristics of the AI tools in use, recognizing that transparency gains may come at the expense of ecosystem value and creator welfare in advanced AI regimes.