A trade body measured what honesty costs, then wrote a rule to avoid paying it

How the IAB put a price on ai labelling, then wrote around it#

On 18 August 2026 the Interactive Advertising Bureau published version 2 of its AI Transparency and Disclosure Framework, the industry's most detailed attempt yet at ai labelling rules, replacing guidance first issued on 16 January 2026. The IAB's own framework page dates it to that day and frames it as a risk-based, materiality-driven approach balancing transparency against what it calls disclosure fatigue.

Page eleven is where the argument lives. The framework document states that an NYU Stern study found telling consumers an advertisement was made with generative AI cut its click-through rate by 31.5 per cent. It then draws the conclusion everything else rests on: that cost is warranted where AI use could mislead, and where it could not, ai labelling penalises the advertiser without protecting anybody.

Read the sequence slowly, because it is unusual. A trade body commissioned an evidence base, found a measured price for candour, then wrote a standard limiting how often its members pay it. PPC Land's read of the document lands in the same place. That is not automatically wrong. It is a conflict worth naming out loud rather than admiring quietly.

Nobody should pretend that number is trivial. Nearly a third of your clicks is a real budget line, and any agency telling a client to shrug it off has never defended a media plan in a boardroom. The awkward part is not the finding. It is the conclusion the framework drew from it.

Voluntary, by its own admission#

The framework states plainly that it carries no enforcement authority and exists to help the industry self-regulate. That sentence sits unhedged in the executive summary, and it is the most important line in the document for anyone weighing how much to lean on the rest. A standard with no teeth is still useful. It is not a defence.

It also arrives sixteen days after two binding regimes became applicable on the same day, 2 August 2026. Either way the sequencing matters, because the framework is not writing on a blank page. It is writing over ai disclosure laws that already reach advertising creative, and in two named places its thresholds do not agree with them.

The number holds up. The citation is not in the document#

A figure carried only by trade press is a lead, not evidence, so folkfox went hunting for the study behind the 31.5 per cent. The first find was about the framework rather than the research: across all 42 pages the document carries no footnotes, no endnote list and not a single URL. Every statistic in it, including the one the whole disclosure threshold rests on, is attributed by name and nothing else.

The study is real, and it is findable if you go looking. NYU Stern's own research page carries the finding verbatim: telling consumers that an advertisement was made using generative AI reduced its effectiveness, causing a 31.5 per cent decrease in click-through rates. The paper is The Impact of Visual Generative AI on Advertising Effectiveness, by Hyesoo Lee and Anindya Ghose at NYU Stern with Vilma Todri and Panagiotis Adamopoulos at Emory.

The method matters as much as the magnitude. The researchers tested three production routes, human-made, human-made then modified by generative AI, and fully AI-generated, across a laboratory experiment and a live field study on the Google Display Network. Reporting on the paper places the penalty in the field test specifically, against an unlabelled human control.

One caveat belongs in the copy rather than a footnote. The working paper is listed on SSRN, and its full text was not publicly retrievable at the time of writing, so folkfox could not confirm sample sizes at source. The 31.5 per cent is verified to the institution that produced it. The number of participants behind it is not.

That 37-point gap between what advertisers assume and what younger consumers report is the framework's own number, and by its own account it has widened since 2024. It sits four pages from the 31.5 per cent, pointing in precisely the opposite direction, and only one of the two ends up shaping a rule.

The evidence that got quieter treatment#

The framework reports that 73 per cent of Gen Z and Millennial consumers said knowing an advertisement was AI-made would increase or not affect their likelihood to purchase, leaving 27 per cent less likely. It also reports clear disclosure as the third-highest driver of attention to AI-made advertising, behind only high-quality visuals and humour.

Meanwhile Pew Research Center found 76 per cent of US adults say telling AI-made pictures, video and text from human-made is extremely or very important, from 5,023 adults surveyed in June 2025. The framework cites that too, filed under industry pain points.

So the document holds two truths at once: labels cost clicks, and audiences want them. Faced with that tension, the framework built its ai labelling thresholds around the cost, and Digiday has been asking who gets to decide that for months. Neither choice is neutral, and pretending otherwise is how a standard loses the trust it was written to earn.

Where binding ai disclosure laws and voluntary ai labelling part company#

Here is the part that should worry any brand treating this framework as ai labelling cover. The IAB itself names two cases where its thresholds sit below EU law, and it names them in the document rather than leaving anybody to find the gap in an enforcement letter. Credit where it is due. The exposure is still real.

Both cases involve commercial endorsement. The framework treats authorised voice clones of a living person, and authorised digital twins of a living person not placed in fabricated events, as carrying no materiality risk, reasoning that a consumer watching a celebrity endorsement knows they are seeing paid messaging. Under that logic, no label is needed.

Article 50(4) of the EU AI Act contains no such carve-out. It requires deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake to disclose that the content has been artificially generated or manipulated. Its only meaningful exemption covers evidently artistic, creative, satirical or fictional work, and a celebrity endorsement is none of those things. The Official Journal text of Regulation (EU) 2024/1689 is the binding version.

Every date on that chart comes from the framework's own regulatory section: Taiwan's advertising-specific fraud rules from November 2024, China's labelling measures on 1 September 2025, South Korea's AI Basic Act on 22 January 2026, India's amended IT Rules in February 2026, New York's synthetic performer law on 9 June 2026, then two on a single day, and folkfox verified those last two at source rather than on trust.

The first is Article 50, which the framework states applies to advertisers as deployers and took effect on 2 August 2026, carrying fines of up to 15 million euros or 3 per cent of worldwide annual turnover. The second is Californian.

That reaction, posted the morning after publication, is the honest version of the industry's case and it deserves a straight answer rather than a sneer. Blanket ai labelling really does dilute the signal, and a label on everything teaches an audience to see nothing. The framework's reasoning there is sound. Soundness in principle just does not survive a regulator who wrote a different rule.

The Californian gap, and why it is not the shield it looks like#

California's AI Transparency Act, SB 942 as amended, became operative on 2 August 2026. AB 853's own text made that change, amending section 22757.6 to read that the chapter becomes operative on 2 August 2026, having previously been set for 1 January.

The statute itself binds covered providers, meaning generative AI systems with over one million monthly visitors or users that are publicly accessible in the state. It does not bind advertisers. It makes the invisible latent disclosure mandatory and the visible consumer-facing label merely an option the provider must offer, with a civil penalty of 5,000 dollars per violation and each day counted separately.

So a brand running Californian creative through a covered tool gets provenance metadata automatically and no visible label unless somebody chooses one. Legal analysis of the commencement reads it the same way. The framework is candid that this leaves a gap and positions itself as what closes it, a fair claim as far as it goes.

It goes exactly as far as the border. An advertiser following the framework's endorsement exemption is compliant in California, exposed in the European Union, and, per the same annex, exposed again in South Korea and India. The scent trail runs cold at the first jurisdiction that wrote its rule differently.

A working method for ai labelling when three rulebooks disagree, and one is not law#

The implementation advice is the sensible half of the document, and it costs nothing to adopt. It asks organisations to designate an AI Disclosure Lead within 60 days, establish governance in the first six months, pilot through month twelve, and scale automated flagging within 24 months. It is explicit that this roadmap is not a compliance schedule, and that binding deadlines apply regardless of where you sit on it.

That caveat is the one to pin above the desk. A 24-month roadmap is a governance plan. Article 50 was applicable on 2 August 2026 and did not wait for your phase two. Reading the roadmap as a grace period misreads a document that goes out of its way to say otherwise.

None of this is exotic work. It is the same discipline behind folkfox's brand strategy and paid social practice: decide the position before the platform decides it for you. Regulated categories learned that in the thicket years ago, which is why the same instincts prowl through the healthcare and fintech work.

The teams who handle this badly will be the ones treating an ai labelling decision as a legal ticket. It is a creative decision with a legal floor, and the people who own it sit in the studio.

What generative ai advertising costs once trust is the budget line, not the click#

Type do ads have to disclose if they use ai? into a search bar this week and you will find a dozen answers that are confidently, jurisdictionally wrong. The honest one depends on where the impression serves, what the AI actually did, and whether anybody at your end wrote it down. Unsatisfying, and true.

It is worth saying what the 31.5 per cent does and does not prove. It proves ai labelling depresses clicks in a display environment where nobody expected one. It does not prove a label depresses trust, revenue or the willingness to buy again, and the framework's own 73 per cent figure suggests the longer arc runs the other way. A click is the cheapest thing you will measure this year.

The practical brief is short. Find every market your creative reaches, apply the strictest rule among them, name one person who owns the call, and record the reasoning per asset rather than per campaign. Then stop optimising for the click a label costs and start measuring what it buys, which is the second purchase, not the first.

This trail is one folkfox has followed before, through content marketing built to survive a screenshot and paid search built to survive an audit. The pattern holds. Brands that write their own ai labelling position early spend two years adjusting it. Brands that wait for a rule spend the same two years explaining themselves, and explaining is expensive in a way disclosure never was.

One last thing, for anyone tempted to wait for version 3. The framework's own annex describes a map that grew from one binding regime to seven in under two years, and calls it fragmented and accelerating. Sitting still in that undergrowth is a decision too, and the only one nobody has to write down.

Frequently asked questions#