The most honest AI label is not always the longest one.

A controlled study of AI disclosures in news writing with 40 participants tested one-line and detailed disclosures. The detailed version reduced reader trust and subscription intent; the one-line label did not produce the same penalty. Both versions increased source-checking behavior. That is a small sample, so it is a signal rather than a universal rule. It is still a revealing signal.

It would be tempting to turn the result into a slogan: say less about AI. The wider evidence refuses that conclusion. In a 393-person experiment on AI-assisted fact-checking, disclosure did not reduce message, author, or source credibility. Detailed explanations of procedure and human oversight improved author and source credibility. A separate study of AI-generated sports news found that ethics policies and supervision could strengthen perceived trustworthiness and financial value.

I keep coming back to the gap between what a disclosure says and what an audience hears. The institution describes a production process. The audience hears a claim about the product.

A label is a theory of the product

Disclosure looks like metadata. In practice, it is part of the message.

A detailed label can signal care, supervision, and candor. It can also read like a process memo attached to work because the promised expertise has gone missing. A short label can feel clear and proportionate, or evasive. Word count does not settle the meaning.

The news study makes that distinction unusually visible. Readers checked sources more often even as detailed disclosure lowered trust and subscription intent. Checking is behavior, not a confession of distrust. It can indicate curiosity, caution, or skepticism. The study does not tell us that every click meant the same thing.

Trust, credibility, source-checking, subscription intent, and purchase intent should not be compressed into a single score. Credibility is an assessment of a message, author, or source. Trust is broader and often relational. Subscription and purchase intent are stated commercial preferences, not completed transactions. They can move together, but this evidence shows why we should not assume that they will.

The paradox softens once those outcomes are separated. Transparency can make AI involvement more salient. Salience can prompt scrutiny. Scrutiny can coexist with lower trust, higher credibility in a supervised process, or no change in willingness to buy. Audiences are not voting on AI in the abstract. They are deciding what its presence means here.

The buyer is judging what AI replaced

Research on generative AI disclosures in service advertising offers the most commercially useful clue. The effect changed with the kind of service being sold. Intangible services, where the offer rests on expertise, care, or identity, faced a stronger trust penalty when people inferred that AI had replaced the human substance of the offer.

That inference matters more than the mere presence of a model.

If AI accelerates research, creates variants, translates material, or helps an expert explore more options, the underlying promise may be intact or improved. If the label suggests that the expertise itself was removed, the same efficiency can look like a cheaper substitute. The artifact may be polished in both cases. The audience’s theory of its value is different.

This is not an argument for preserving manual work as theater. Frontier models should do more where they improve speed, coverage, personalization, accessibility, or creative range. But “made with AI” is now too blunt to explain any of that. It names a tool category while leaving the material role unanswered.

My read is that many disclosures answer the institution’s question, “Did we disclose?”, while the customer is asking, “What am I buying?” The second question is the one that shapes trust and purchase intent.

For marketers and publishers, that creates a new narrative task. The label should explain what AI added without accidentally announcing that the scarce ingredient has vanished. When the product is information, expertise, taste, or care, production method is inseparable from perceived value.

Supervision changes the meaning of automation

The fact-checking study provides a clean counterweight. Detailed disclosure improved author and source credibility when it clarified procedure and human oversight. In that setting, AI could be interpreted as an instrument inside an accountable method, rather than a substitute for accountability.

The sports-reporting research points in the same direction. An ethics policy and visible supervision can make AI use feel managed rather than vacant. These studies do not cancel the news-writing result, and their measures are not identical. Together, they suggest a useful mechanism: audiences respond to whether disclosure signals substitution or supervision.

That pattern is still an inference, not a settled law. But it gives builders a better design variable than disclosure volume.

A strong disclosure makes three things legible: the material AI intervention, the direction behind it, and the person or institution willing to own the result. Direction does not require a person to rewrite every sentence. Accountability does require someone to stand behind the decision to publish, recommend, or sell.

This framing lets a company be openly AI-native. It can say that models expanded the work while remaining precise about where judgment and responsibility sit.

Disclosure is becoming a creative primitive

As models move through ideation, research, drafting, editing, localization, personalization, and versioning, a binary AI badge carries less information. “AI was used” could describe a translated caption, a synthetic image, a first draft, the central analysis, or an entire adaptive customer experience. Those are not the same intervention.

A reader-centered study of news disclosure design points toward more useful interfaces. Participants proposed detail on demand, proportional views of AI involvement, and outlet-level signals. A ratio may imply more precision than a fluid workflow can support. The more promising idea is layering.

The first layer can state the material role in plain language. A second can explain process, oversight, and correction practices for readers who want it. Machine-readable provenance can sit beneath both. A concise label might say that AI assisted with translation and first-pass summarization while editors reported, verified, and approved the article. If AI generated the central analysis, the label should say that instead.

The point is not the sample copy. It is the architecture. Disclosure can be designed around audience need rather than institutional anxiety.

Industry practice is beginning to move this way. The IAB’s AI transparency framework favors a functional, risk-based model over blanket labeling when AI has a material effect. Google’s advertising disclosures combine post-click information in My Ad Center with machine-readable provenance such as SynthID and C2PA for assets made in its tools. Neither example proves what audiences will trust. They do show that disclosure can be a layered product surface rather than a warning sticker.

This extends two arguments already in this archive. Trust has to be designed into the interface, and authenticity has become a deliberate design choice. The additional point here is narrower and more commercial: disclosure is not information placed around the creative product. It helps determine what the audience believes the product is.

That makes it a creative primitive. It belongs in the brief, the customer journey, and the value proposition, rather than being confined to final approval. Done well, it can show that AI increased range or responsiveness while preserving the meaning of the offer. Done poorly, it makes low-cost production the headline and asks the audience to guess what, if anything, remains accountable.

Say what changed in the promise

The strongest disclosure is proportional. It says enough to prevent a false impression about material authorship or expertise, without burying the customer in process exhaust. Restraint here is precision, not concealment.

Start with the thing being sold. If the promise is speed, breadth, personalization, or continuous adaptation, visible AI use may strengthen it. If the promise is expert judgment, care, identity, or editorial authorship, the disclosure must make clear how AI contributed and who still stands behind the consequential choices.

This is the opportunity hidden inside the trust tension. Better disclosure does not merely reduce reputational risk. It gives AI-native products a clearer language for explaining their advantage. The organization can be candid about automation without pretending that every use has the same meaning.

As frontier systems take on more of the creative and analytical process, the disclosure question will become more specific, not less. The useful label will not answer every imaginable process question on first contact. It will identify the intervention that changed the offer, make accountability visible, and let interested audiences inspect the rest.

The honest question is no longer simply, “How much AI was involved?” Before publishing the label, ask: What part of the promise changed, and who is willing to stand behind what remains?

Sources and further reading