AI Disclosure Labels Don't Hurt Ad Performance. Format Is the Real Decision

What the AI Labeling Impact Study Tested

AI disclosure labels do not hurt ad performance. A 900-person study from MediaScience found no decline in brand choice, ad memory, recognition, attitude or liking across four AI disclosure label formats tested against an unlabeled control. The open question is no longer whether to disclose, but which format actually informs the viewer.

MediaScience ran the research jointly with the Ehrenberg-Bass Institute for Marketing Science at Adelaide University and MediaPet, its own AI video platform, polling 900 U.S. respondents. The experiment put four disclosure formats against an unlabeled control: a text label in the first three seconds, a delayed text label from seconds four to six, a text label for the entire duration of the ad, and an icon displayed for the full duration. The four formats were chosen to mirror the disclosure frameworks legislators in the U.S. and the EU were actively weighing, which is what makes the numbers usable rather than theoretical.

The design detail that runs through every finding below: respondents watched well-made ads. The study answers a narrow, practical question - does adding a disclosure change how people respond to a good ad - and the answer, across every metric tracked, is no. What it does change is whether viewers register that AI was involved at all, and that is where format takes over.

The question the study answers has been sitting in every brief since AI creative started shipping at volume. The working assumption in most agencies was that disclosure is a tax: tell viewers an ad was made with AI and something - recall, liking, trust - gives. Nobody had measured it properly, so the assumption hardened into policy, and policy is hard to argue with.

No Performance Penalty for AI Disclosure Labels

Across brand choice, ad memory, brand recognition, brand attitude, ad liking and perceived production quality, none of the four labelled conditions differed significantly from the unlabeled control. On unaided brand recall the labelled groups ran slightly ahead of it: the control and icon groups each recorded 54%, while the delayed and continuous-text groups reached 61% and the early-text group 60%. Brand recognition stayed inside a five-point band, from 87% for continuous text to 92% for early text.

MediaScience CEO Duane Varan's summary is the line to keep: "if the creative is good, disclosure does not hurt it. Advertisers do not need to be afraid of the label." The label is neutral; the work is not. The {{link}} between what advertisers assume and what audiences actually feel has been documented for two years - this study costs the label itself, and the bill comes to zero.

The finding matters because the objection it retires has been doing real work in briefs. Teams have delayed launches to argue about label placement, quietly stripped AI credits from end cards, and treated disclosure as a creative risk to be managed rather than a requirement to be designed. A neutral label removes that excuse and puts the burden back where it belongs: on the spot.

The perception gap between what advertisers assume and what audiences actually feel has been documented for two years - this study costs the label itself, and the bill comes to zero.

The Icon Paradox: Preferred by 13 Points, Least Effective at the Job

The most counterintuitive result is about format. Asked which disclosure style they preferred, respondents chose the visual icon over the other styles by 13 percentage points. Yet the icon was the least effective format at the one job a disclosure has - making viewers aware the ad used AI. Continuous text labeling produced 49% AI awareness, early text 46%, delayed text 40%. The icon reached 38%, barely above the 36% recorded by the unlabeled control.

The same pattern shows up in brand response. Positive brand attitude sat at 44% in the icon group against 51% for both the control and the early-text group, and ad liking ran 56% for icons versus 70% for both short-duration text conditions. A disclosure that viewers prefer but do not parse is a compliance artefact, not a disclosure. Some of that work is moving upstream, where {{link}} attach labels, invisible watermarks and moderation at the moment of creation - but provenance metadata is not the viewer-facing label this study measured.

There is a production lesson buried in the paradox: the cleanest-looking option is the one most likely to fail silently. An icon that does not register leaves the brand with the cost of compliance and none of the clarity, and nobody notices until a regulator or a journalist asks what the mark was supposed to communicate.

Some of that work is moving upstream, where tool-native disclosure safeguards attach labels, invisible watermarks and moderation at the moment of creation - but provenance metadata is not the viewer-facing label this study measured.

Where Audiences Actually Want AI Disclosure Labels

The study also mapped when people believe disclosure is warranted, and the gradient is steep. Simulated humans topped the list at 60%, followed by animals at 46%, product placement at 45% and voices at 45%. Sensitivity falls away from there: animation 41%, background 35%, script writing 33%, music 28%, text and subtitles 24%, translation and dubbing 23%, and colouring and lighting last at 21%.

The practical read is to segment by synthetic-human risk rather than by tool usage. A synthetic performer in the frame and a colourist's AI-assisted pass are not the same disclosure problem, and audiences say so themselves. YouTube's rule draws the same line: creators must disclose when they use AI to meaningfully alter or generate photorealistic content, including making a real person appear to say or do something they did not do, while non-realistic or minor edits do not require it.

That distinction should shape the brief, not just the end card. If the spot's AI involvement is invisible - a grade, a cleanup, a music bed - the disclosure expectation is low and the effort belongs elsewhere. If a synthetic human carries the message, disclosure is the single most expected element in the cut, and the format decision above becomes the whole game.

The Compliance Calendar Turned Disclosure Into a Formatting Decision

New York's AI Transparency in Advertising law took effect in June 2026 and requires specific disclosures when advertisements include what the statute calls "synthetic performers", with fines of US$1,000 for a first violation and US$5,000 for each one after. The EU AI Act's transparency rules came into effect in August 2026: providers of generative AI have to ensure that AI-generated content is identifiable, and certain AI-generated content should be clearly and visibly labelled.

None of this makes disclosure optional. It makes it decidable. The teams that get ahead are the ones treating the label as part of the creative's trust mechanics - the way sponsored tags and privacy notices became standard furniture rather than apologies - instead of a required evil retrofitted at delivery.

The obligations themselves are settled - {{link}} moved from soft guidance to enforceable law in 2026 - so what remains is operational: which format, where in the cut, and who signs off. That is a different finding from the collapse in {{link}} documented this year: audiences are skipping more AI-made creative overall, yet the label on a good ad costs nothing by itself. Brands are already contracting for visibility into how their agencies use AI, and {{link}} is now a deal term rather than a courtesy.

The obligations themselves are settled - synthetic media advertising regulation moved from soft guidance to enforceable law in 2026 - so what remains is operational: which format, where in the cut, and who signs off.

That is a different finding from the collapse in AI video ad tolerance documented this year: audiences are skipping more AI-made creative overall, yet the label on a good ad costs nothing by itself.

Brands are already contracting for visibility into how their agencies use AI, and agency AI transparency is now a deal term rather than a courtesy.

What to Do on the Next AI Spot

Brief the label into the first three seconds. Early text was the only condition at or near the top of every measure - 46% AI awareness, 92% brand recognition, 51% positive brand attitude and 70% ad liking - and it is the cheapest option to produce: one supers, locked before the final render, with no icon-licensing debate attached.

Design the disclosure early rather than retrofitting it at delivery. The study's own framing is that disclosure has become a design constraint inside creative, not a legal line item, and constraints are cheaper to honour in the brief than in the conform. If the house style standardises on icons because they look cleaner, plan the reinforcement - a caption, an end card, a voice line - because the awareness data says the mark alone will not carry the meaning.

Then keep testing. These results come from a controlled setting with a single exposure per respondent; your category, frequency and audience are still variables. Run the label the way you run any other element with a job to do: as a testable unit with a read on awareness and brand response, not a legal footnote bolted to the end card. Set the read at the intervals you already use for hook tests - an early signal on awareness, a longer read on brand attitude, because that is where the icon condition lagged. The performance objection is dead. The format decision is yours.