In my world, the question of AI disclosure shows up in litigation: do consultants need to disclose when they use AI to prepare an expert report?
Courts and professional organizations are increasingly debating the question.
But the question pops up everywhere. Should AI-generated music be labeled? Should doctors have to tell you when they used AI to read your scan? Should a news photo carry a tag if a model touched it? The same three words—“made with AI”—are being proposed as a label across wildly different markets, as if one rule could fit them all.
I spent much of my academic career studying mandatory disclosure markets—restaurant hygiene grade cards in Los Angeles, calorie counts on chain menus. AI-labeling raises the same core question: which economic assumptions hold in a given market? Sometimes, AI disclosure serves no purpose. In other cases, it clearly matters. Where it matters, the deeper issue is whether the market produces disclosure on its own, or requires a mandate.
Where the market will solve it: News images
Start with photographs and videos used in news. The entire value of a news image is its claim to be a faithful record of something that happened. A fabricated image presented as real is not a low-quality news image; it is the opposite of one.
Provenance here is not a bonus feature—it is the product. News organizations compete on credibility and are rewarded for authenticity. We are already seeing moves toward content-provenance standards and trusted-source signaling, suggesting the market will solve this without a mandate.
Where disclosure is irrelevant: Expert reports
Now the opposite case. Do I care whether my accountant used AI to prepare my return? I do not. Not one bit. I care that the return is correct. The work’s origin is not core to what I am buying. I’m buying a verifiable return. If the numbers are wrong, and an audit catches them, it makes no difference whether the error came from a junior associate, a spreadsheet macro, or a language model.
Expert reports, the most scrutinized document I can think of, fall into this category. Opposing counsel and courts scrutinize every detail, and belief comes from surviving that adversarial process. Disclosure of AI use adds nothing, because what matters is already verified through this scrutiny. Provenance is irrelevant to value.
The ambiguous middle: Creative goods
Then there are creative goods, where it gets hard. A study of one of the largest stock-image marketplaces by Samuel Goldberg and Tai Lam of Stanford GSB and UCLA Anderson, respectively, is instructive. In December 2022, the marketplace began allowing labeled AI-generated images while banning AI in certain “editorial” markets (used as a control group). In a difference-in-differences analysis, they find that permitting AI-generated content led to a 136% increase in images, a 47% increase in active artists, and an 82% rise in total sales. Variety and quality both rose, the latter by about 10% overall.
The same policy also led to a 15% decline in non-AI image production, a 29% decline in active non-AI artists, and a 28% decline in non-AI sales. The lowest-quality human artists exited; the survivors differentiated into niches the AI did not crowd. The conclusion: AI expands the market and helps buyers, but it crowds out original human production, with real consequences for copyright and the long-run sustainability of human creative work.
The deeper question is not if labeling mattered, but whether a mandate was necessary or whether the market would have produced disclosure on its own.
The unraveling hypothesis, and where it breaks
The unraveling hypothesis—a fundamental insight in the economics of disclosure—says that under the right conditions, you do not need to mandate disclosure at all. Suppose a seller privately knows some quality, buyers know the seller knows it, disclosure is cheap and verifiable, and everyone agrees on what counts as good. The best type discloses to separate itself from the pack. Once it does, the best of the remaining pool discloses, too. The cascade runs all the way down, because silence is read as “worst remaining,” and no one wants to be assumed the worst. Skeptical inference is the engine. Voluntary disclosure becomes total, and a mandate is redundant.
The elegance of the result is also its weakness: it rests on a stack of assumptions, and the AI-disclosure puzzle is about which assumption fails in which market.
In expert reports, unraveling has nothing to grab onto. There is no agreed-upon ranking—provenance is orthogonal to value—so there is no worst type to flush out. And the dimension that does matter is already verified by adversarial audit.
In news images, the value ranking is clear—real beats fake for everyone—but verifiability is the constraint. Once certification technology exists, unraveling and market disclosure follow. My bet that “the market will solve it” is really a bet that verification arrives and unraveling follows.
That bet is not guaranteed. AI watermarks are technically fragile—easily stripped—and content-provenance standards require near-universal adoption to produce the rational discounting that unraveling needs. If verification proves more difficult than current standards suggest, a mandate may be the only alternative.
In creative goods, two assumptions give way at once. Verification is costly, which is why the platform itself had to build the detection and enforce it. And there is no common ranking: some buyers prefer human-made work, others are indifferent, and some prefer the cheaper, faster, on-trend AI option. When the “bad type” is not bad to everyone, revealing it as AI does not uniformly punish the seller, and the cascade stalls.
The likely market resolution for creative goods is not a mandated warning label slapped on AI work to make it confess. It is a voluntary certification flowing the other way—“human-made,” like “organic,” “non-GMO,” and “handmade.” The producers serving the taste for human work certify, charge a premium, and the market sorts itself. That is unraveling operating on the authenticity dimension, no mandate required—provided someone can verify the claim cheaply. Verification is the hinge in every one of these cases.
Reputation as a substitute for information disclosure
Before Los Angeles ever put a grade card in a restaurant window, a quarter of the county’s restaurants already kept an A-grade hygiene. Nobody made them; reputation motivated them. The clearest evidence came from chains, which scored about 3.7 points higher than independents.
A chain internalizes a shared reputation—a bad meal at one unit tells you something about all of them. Franchisees, who keep their own unit’s profit and free-ride on the chain’s name, scored lower: Burger King’s franchised units ran 4.89 points below its company-owned ones. After the grade card arrived, the gap collapsed to a tenth of a point, and the chain’s hygiene advantage fell by roughly half.
The grade card is informative only for restaurants that lack a reputation. For the one that already has it, the card “provides little additional information.” The mandate did the work that the reputation had been doing.
The restaurant data reframe the expert-report case. Reports survive adversarial audit and carry firm reputation—both substitutes for a disclosure mandate. These verify what buyers care about: correctness of analysis, not how it was produced.
A subtler lesson: mandates not only become redundant for reputation-rich firms but can erode their advantage. Grade cards hit hardest where there was no reputation, shifting value from firms that invested in reputation to those that had not.
This result should give pause to anyone who has spent decades building a reputation, which I take personally. A disclosure mandate for expert work is not the neutral gesture its proponents imagine. It grades each report on a provenance tag rather than the standing of the firm behind it, commoditizing the very thing a reputation-rich firm has spent its existence building.
But the strongest argument against the label is that it fails to track the value that clients are buying: provenance is orthogonal to what matters. The restaurant analogy holds, but in expert work, a mandate would misalign with what buyers care about.
A caution from the field
The field evidence keeps telling me that the larger consequence of a disclosure rule is not what buyers learn, but what it does to the firms and people subject to the rule.
So, when someone proposes that AI use “should be disclosed,” I have stopped treating it as one question. Instead, I consider these three questions:
Does provenance affect what the buyer is buying?
Can the thing that matters be verified in some other way?
And if disclosure does belong in this market, which assumption is broken, and is a mandate the right repair, or is the fix a verification technology that lets voluntary disclosure do the work on its own?
For expert reports, my answer stays the same. The audit already does the work—and so does the reputation. Both verify the one input that stays scarce as AI absorbs the rest—not the production of the analysis, but the judgment behind it. The label flags production; the bottleneck is judgment. The label is a red herring.
The views expressed herein are solely those of the author and do not necessarily represent the views of Cornerstone Research.
I couldn’t agree with this more if I was twins!