How brands are rewriting their AI ad policy
In August the industry's own standard stopped asking for a label on everything. A 31.5% fall in click-through is why.
Through most of 2026 the internal policy on AI in advertising was the same at a lot of brands: label it, to be safe. In August the industry's own standard stopped recommending that. The IAB published version two of its AI Transparency and Disclosure Framework, seven months after the first, and the change is not a tightening. It is a narrowing — from labelling AI use to labelling only the AI use that could mislead someone.
There is a number behind the change, and it is the reason this is a commercial document rather than an ethical one.
Disclosure has a measured cost now
The framework cites research from NYU Stern finding that telling consumers an ad was made with generative AI cut its click‑through rate by 31.5 per cent. Worth being careful with that figure: the reporting we have attributes it to NYU Stern without naming the authors or the sample, so treat it as a direction with a magnitude attached rather than a settled constant. But it is the first time the industry has put a price on a label, and a third is not a rounding error.
31.5%
Reported fall in click‑through when an ad is disclosed as AI‑made. Cited by the IAB framework to NYU Stern research.
Put that next to the label‑everything policy and the problem is obvious. A brand that discloses uniformly is paying that penalty on work where nobody was at risk of being misled — a colour‑corrected product shot, a translated headline, an upscaled export. It is a real cost incurred for no disclosure benefit, and it was being incurred out of caution rather than on advice.
The test is about the viewer, not the tool
The framework replaces the blanket rule with a consumer‑impact test. Disclosure is called for where AI involvement could mislead a reasonable consumer about what is authentic, factual or human‑created. Where it could not, it is not.
The question is no longer whether a model touched the work. It is whether the audience was led to believe something untrue.
Readers of this site will recognise that test, because it is the one Singapore's own advertising code already applies — disclose where not disclosing would mislead. The IAB has arrived at the same shape by a different route, which is mildly reassuring if you work across both. We have written separately on where the APAC mandates actually bite, which is law rather than voluntary standard and does not bend to a consumer‑impact argument.
The two lists
The useful half of the framework is the second list, because it is the first time anyone has written down in a citable document what does not need a label.
Warrants disclosure. Images and video generated from prompts, even after human editing, unless the output is obviously non‑realistic. Synthetic voices of deceased people saying things they never said — even with estate authorisation. Synthetic voices of living people depicting events that never happened, in a way a viewer could take as authentic. Digital twins of deceased people, in any use. Digital twins of living people in fabricated events or locations. Photorealistic AI influencers in paid creative. Conversational agents that simulate human interaction, which carry an AI‑powered label rather than a disclosure.
Does not warrant disclosure. Text and copy, including headlines, product descriptions and translations. Routine editing — colour correction, lighting, blemish removal. Technical work — upscaling, noise reduction, format conversion. Stylised or fantastical imagery. Animated mascots, and incidental AI figures in crowds or backgrounds. Authorised voice clones in scripted commercial content. Generic synthetic voiceover where the speaker's identity does not matter. AI‑generated background music, EQ and pitch correction.
Read the second list as a production document and a pattern appears. The line is drawn at photorealism plus a person plus an implied claim about reality. Everything that is clearly a construction, or clearly a technical step, or clearly not about a human being, sits on the safe side of it. That is a workable rule, and it is close to the one a careful studio was already using.
The clause about dynamic creative
This is the part that will catch brands running personalised or assembled campaigns, and it is easy to miss.
If any component that exists before the campaign launches triggers disclosure, the assembled creative carries the label. Not that component — the whole assembly. So one photorealistic AI‑generated background in a library of forty modules puts a label on every variant built with it. And for components generated at serve time, where nobody can review each output, the advertiser is expected to set disclosure rules at the system level in advance.
Which makes this an architecture decision taken at build time, not a legal review taken at the end. If the module library is mixed, the whole campaign discloses. Keeping the photoreal‑human modules in a separate campaign from the rest is the cheap version of that problem, and it has to be decided before anything is made.
What disclosure does not fix
The framework is explicit on a point brands keep getting backwards: AI‑generated false claims remain prohibited under existing consumer protection law, and a disclosure does not cure them.
A label is not a disclaimer. Saying a film was made with AI does not license a product claim the product cannot support, and it does not retrospectively authorise a likeness nobody consented to. The rules that applied to a false claim in a filmed ad apply unchanged. All the framework governs is whether the audience is told how the thing was made.
Why the standard moved, and where it is going
Voluntary industry standards usually move because statute is about to. The regulatory pressure behind this version is real and it is not in one place: Article 50 of the EU AI Act, California's SB 942, New York's synthetic performer law, South Korea's AI Basic Act. Those have different triggers and different remedies, and a voluntary framework is partly an attempt to give the industry one shape to comply with rather than four.
It is worth saying what that means for a brand sitting outside the US and EU. The IAB framework is not law anywhere. It is a defensible position to have adopted — which, when a complaint lands, is most of what a written policy is for. The mandates in China and Korea are not satisfied by it.
What we would put in a brand's policy
Three lines, and they are shorter than the policy most brands currently have.
Disclose where a real person's authenticity is in play. Photoreal humans, voices, twins, testimonials. That is where the framework, the EU, Singapore's code and ordinary decency all agree, and it is where the risk actually lives.
Stop disclosing the technical steps. Grading, upscaling, translation and background music are on the published does‑not‑warrant list. A brand paying a click‑through penalty on those is paying for nothing.
Decide the dynamic‑creative rule at build time. One triggering module labels the whole assembly, so the module library's structure is the policy. That is a decision for whoever builds the system, taken before the first variant exists.
And the test underneath all of it, which has not changed and will outlast this version of the framework: if the audience were told exactly how this was made, does the work still stand up? If it does, disclosure is paperwork. If it does not, the disclosure found a problem that was already there.
Where this came from
The framework is the IAB's AI Transparency and Disclosure Framework V2, announced in August 2026, with the full document published as a PDF; version one appeared in January 2026. The disclosure lists and the dynamic‑creative clause above are as reported by PPC Land, which is also our source for the 31.5 per cent figure and its attribution to NYU Stern. We have not seen the underlying study, and have said so above rather than letting the number travel unqualified. Checked November 2026. A voluntary framework on its second version in seven months will have a third.