The Transparency Tax

Every AI product now ships with a sentence written by someone in legal, and almost nobody in the building is measuring what it costs. That was survivable while the sentence was optional. It stopped being optional on 2 August 2026, and the research on what disclosure does to conversion is considerably less comfortable than the compliance decks suggest.

What does the disclosure actually cost?

The cleanest evidence is a field experiment published in Marketing Science in 2019. A fintech firm ran outbound loan-renewal calls to 6,255 customers, randomizing whether the caller was a chatbot or a human, and whether the bot said what it was.

Undisclosed, the chatbot converted at 23.7%. Proficient human agents converted at 25.1%, so the difference was not statistically meaningful, and the bot was roughly 4 times more effective than inexperienced staff. Disclose the bot's identity before the conversation and the purchase rate fell to 4.8%. Calls that opened with the disclosure lasted about 10 seconds against nearly 64 for human agents. People heard "AI" and hung up.

The mechanism matters more than the number. The disclosed bot was the same bot. It had already been shown to perform at human parity. What changed was that customers who knew rated it as less knowledgeable and less empathetic, which is a judgment about the category rather than about the system in front of them.

The mitigation just became unavailable

That study also found what fixes it, and the fix is timing. Disclosure delivered after the conversation recovered conversion to 11.0%. Disclosure delivered after the customer had already decided recovered it to 23.2%, with no meaningful gap from human agents. The cost was never the fact. It was the fact arriving first.

Article 50 of the EU AI Act requires that people interacting directly with an AI system be informed they are dealing with AI, with a carve-out for cases where that is already obvious from context to a reasonably well-informed person. AI-generated or manipulated content has to carry machine-readable provenance marking. The European Commission's enforcement powers became applicable on 2 August 2026, and Article 50 breaches sit in the tier that runs to 15 million euros or 3% of worldwide annual turnover, whichever is higher.

Two details get missed. The duty falls on deployers, so a company that builds nothing and integrates someone else's model still carries it. And when the Digital Omnibus pushed the high-risk conformity deadlines out to December 2027, it left Article 50 and the general-purpose AI dates untouched, which is why a lot of roadmaps filed this under 2027 and were wrong. Where exactly the obvious-from-context carve-out lands for a given product is a question for counsel, not for a growth team.

Why does a true statement cost anything?

Because the penalty is about legitimacy, not accuracy. Across 13 experiments published in Organizational Behavior and Human Decision Processes, actors who disclosed using AI were trusted less than those who did not, and the effect traced to reduced perceptions of legitimacy rather than to doubts about the output. Trust from students dropped 16% when they learned a professor had used AI for grading.

The uncomfortable part is what did not help. The researchers tried gentler framings, saying AI was used only for proofreading, or that a human had reviewed the output. Trust still declined. It declined whether the audience already knew AI was involved, and whether the disclosure was voluntary or legally required. Copy alone does not buy the penalty back.

1 finding cuts the other way and is worth more than the rest combined: being exposed is worse than disclosing. When a third party revealed the AI use, trust fell further than when the actor disclosed it themselves. So the transparency tax is the price of the cheaper option, which is a very different thing from a cost you can avoid.

More disclosure does not fix it either

The instinct after reading the above is to over-explain: a longer notice, a capability breakdown, a confidence score, a link to the model card. A 2025 web experiment with 491 participants across 2 different interaction types found that this backfires in a specific shape. The relationship between transparency and intention to use is an inverted U. Moderate transparency raised trust and certainty. Excessive transparency produced cognitive overload and heightened scrutiny, and adoption fell.

Which puts the design target somewhere neither the legal team nor the growth team would pick on their own. Minimal disclosure fails the regulation. Maximal disclosure fails the user. The version that works is calibrated, and calibration is an empirical question about your specific product, which means somebody has to test it.

The honest counterweight

79.7% is the worst case the mechanism can produce, not a forecast for your funnel. That experiment ran in 2019, on cold outbound calls, to consumers who at the time had almost no experience of a competent AI on the phone. Public familiarity has moved a long way since.

The trust research supports discounting it too. The penalty was attenuated, though not eliminated, among evaluators with favorable attitudes toward the technology and those who believed AI to be accurate. Both of those populations have grown. And in support specifically, satisfaction tends to track whether the problem got solved rather than who solved it, which suggests the tax is heaviest at the front of the funnel and lighter after the product has done something useful.

The direction is well evidenced. The magnitude is yours to measure, and the fact that nobody has measured it is the actual problem.

What growth owns here

The sentence is required. Everything around it is a product decision, and 4 of them belong to growth.

Where it sits. The obligation attaches to the interaction, which is not the same as the first screen of your marketing site or a line in the terms. Mapping which surfaces actually trigger it, and which are already obvious enough to fall in the carve-out, is scoping work that decides how many times a user meets this sentence.

What it says. Not because good copy recovers the trust penalty, since the evidence says it does not, but because a disclosure that also sets an accurate expectation prevents the second and larger failure. A user told they are talking to an AI, and shown what it can and cannot do, is being handed the honest version of the trial I described in The AI Trust Gap. A user told only the first half meets the same disappointment with less warning.

What follows it. A visible route to a human is what keeps disclosure-driven abandonment from being terminal. In the call experiment, the disclosed condition did not produce a worse conversation. It produced a 10-second one.

Whether it is measured at all. This is the gap. Disclosure arrived as a compliance artifact, so it entered the product without a funnel event attached, and no dashboard currently reports what happens at it. Instrument the step: drop-off at the disclosure surface, session length after it, completion rate for users who passed it, and repeat rate. That is ordinary funnel work on a surface every single user sees, which is exactly the profile of the changes I argued compound in The Compound Effect of Small Funnel Fixes. Defining what counts as passing it is the part that needs care, for the reasons in Metrics That Move Teams.

The provenance requirement is a separate build. Machine-readable marking on generated output is a release-path artifact, which puts it next to the eval suite rather than next to the privacy policy, and I made that case in Evals Are the New QA.

The step that is doing a job

Growth spent 20 years removing steps. Agentic checkout was the reminder that some steps are carrying the argument rather than blocking it, which is what I worked through in Friction Was Doing a Job. Disclosure is the inverse case and a stranger one: a step nobody chose, that removes something, and that cannot be deleted.

So it gets designed instead. Legal will tell you the sentence is mandatory, and that is the entire scope of their answer. Where it appears, how it reads, what sits underneath it, and what it costs are product questions on the first thing a user learns about your software, and most teams shipped it without anyone checking the number.