Can a label stop a deepfaked ad from fooling people? What a Journal of Advertising study found
Swap a celebrity’s face into an ad and viewers cannot tell. Three experiments tested what you have to write underneath before the penny drops, and the usual wording is not it.
A plain “this ad is deepfaked” label changed nothing. Adding one sentence about what that means cut perceived truthfulness from 3.33 to 2.41, without costing a single sale.
A mouth saying words the face never said. Photo: Altin Ferreira on Unsplash.
The short answer
Only if the label finishes the thought. A plain warning that an ad has been deepfaked left viewers exactly where they started: they still rated the ad as realistic, still liked it the same, still clicked through at the same rate. Adding one sentence that drew the conclusion for them — this is not a real scene — pulled perceived truthfulness down from 3.33 to 2.41 on a seven-point scale. And it did that without denting choice of the advertised product, which is the part that makes it usable in practice.
Key takeaways
Undisclosed, a deepfaked ad was indistinguishable from the real one: viewers rated it 3.61 for looking synthetic, against 3.73 for the genuine ad.
The usual style of label moved nothing at all: realism 4.85 with it against 4.84 without, and no change in liking or click-through.
Spelling out the consequence worked: perceived truthfulness fell from 3.33 with no label to 2.41, and 2.68 when a line about being misled was added.
Across all three studies, the share choosing the advertised product never moved, so transparency here cost nothing.
What the researchers did
A deepfake puts a face into footage it was never in. Malgorzata Karpinska-Krakowiak (University of Lodz) and
Martin Eisend (European University Viadrina and Copenhagen Business School) set out to test whether the labels
the industry reaches for protect anybody at all.
They built their own deepfaked ads. In the first, a perfume commercial starring Charlize Theron was altered so
that Margot Robbie appeared to be the one on the catwalk, selling an invented brand. In the second, a food
delivery ad fronted by Kim Kardashian had her face replaced with a stranger's. Then they attached different
wording under each ad and put it in front of fresh samples of British adults.
The clever part is what they measured. Earlier research on retouched photographs asked whether a label makes an
image look less realistic, and the answer is almost always no. The authors split that idea in two: how real a
thing looks, and whether people believe it actually happened. A deepfake is designed to be the first without
being the second, so the second is the only one a label can hope to shift.
Studies
Three online experiments, with fresh samples each time
People
121, then 308, then 300 UK adults recruited through Prolific
Ads
A deepfaked perfume ad, a deepfaked food delivery ad and a deepfaked chocolate ad, all for invented brands
Labels tested
None, a plain statement of the swap, that statement plus what it means for the ad, and that plus what it means for the viewer
Outcomes
How realistic the ad looked, whether it was believed to be truthful, liking of the ad, and whether people clicked to buy
The fake is invisible until you say so
First the uncomfortable baseline. When the deepfaked perfume ad carried no label, viewers rated how synthetic
it looked at 3.61. The untouched original scored 3.73. In other words, people were no more suspicious of the
doctored ad than of the real one, and if anything slightly less. Only when a disclaimer appeared did the rating
jump to 5.71.
So detection is not happening on its own. Whatever protection exists has to come from the words underneath.
Then the bad news about those words. The label used in this first study was the sort you see in the wild: a
byline explaining that the post contains synthetically generated images, with some elements swapped in from
other sources. It made no difference to anything that mattered. Realism came out at 4.85 with the label and
4.84 without. Liking of the ad, and the decision to click through and buy, were equally untouched.
Finish the sentence
The second experiment is where the paper earns its keep. Four versions of the same deepfaked delivery ad went
out. One had nothing underneath. One had the plain statement: the face of the original actress has been swapped
with somebody else's. A third added a consequence for the ad itself — so it does not portray reality, it is
fake. A fourth added a consequence for the viewer — the goal is to mislead you into thinking you are watching
reality.
Belief in the ad's truthfulness barely moved between no label and the plain one, 3.33 against 3.20. Both of the
versions that carried a consequence pulled it down sharply, to 2.41 and 2.68, and they were not meaningfully
different from each other. The third study repeated the pattern with a different product.
How realistic the ad looked never shifted in any condition, exactly as predicted. A well-made fake looks well
made whatever you write beneath it. What you can change is whether people think it happened.
Did viewers believe the ad showed something real? Mean score by label
Did viewers believe the ad showed something real? Mean score by label
Group
Value (points)
No label
3.33
Plain statement of the swap
3.2
Plus what it means for the ad
2.41
Plus what it means for you
2.68
Study 2: 308 UK adults, perceived factuality on a seven-point scale; the two highlighted labels differ from the first two but not from each other.
It does not cost you the sale
Here is the objection every marketer raises, and the answer the authors anticipated. If a disclosure teaches
people the ad is untrue, surely it wrecks the ad.
It did not. In the second study, the share of people who clicked to buy ran from 13.5% with no label to 17.9%
with the plain one, 15.8% and 15.0% with the two stronger ones. Statistically that is a flat line. The third
study, with a chocolate brand, produced the same flat line at a higher level.
Underneath, two forces were cancelling out: believing the ad less pushed liking down and buying with it, while
the disclosure itself nudged choice up a little, perhaps because candour reads as confidence. The authors treat
that cancellation as the point rather than a disappointment. A label that informs people without hurting the
campaign is one a brand will actually agree to run.
Share who clicked to buy the advertised product, by label
Share who clicked to buy the advertised product, by label
Group
Value (%)
No label
13.5
Plain statement of the swap
17.9
Plus what it means for the ad
15.8
Plus what it means for you
15.0
Study 2: the differences are not statistically significant, which is the finding that matters here.
Who the label leaves behind
The third study added the catch. The stronger labels worked best on people who expected advertising to tell the
truth in the first place. On people who assume ads are embellished anyway, the warning landed on ground that
was already conceded and changed little.
That is an awkward result for policy. The consumers best served by transparency rules are the trusting ones;
the cynical are left where they were, which is not the same as being protected. Nor does scepticism help you
spot a deepfake, as the first study showed.
Reaching past what these experiments tested, the lesson generalises to every label now being bolted onto
synthetic media: the AI tags on social platforms, the small print under a cloned voice in a radio spot, the
badge on a virtual influencer's profile. Naming the technique tells people what was done. It does not tell them
what to do with that knowledge. The sentence that changes behaviour is the one after the disclosure, and almost
nobody writes it.
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
List every asset with a synthetic person in it. Go through the campaign and mark anything where a face, voice or body is generated, swapped, de-aged or cloned, including AI stock models and archive footage of people who did not film the ad. That list is what needs a label.
Agree the wording before production, not at trafficking. Settle one standard sentence with legal and brand now, so nobody is drafting it the night before launch and quietly shortening it to a technical term.
2
SetBuild it
Write two lines: what changed, then what it means. Name the alteration, then draw the conclusion for the viewer: the person in this film was generated, so the scene never happened. Keep both lines under 20 words in total and out of technical language.
Put it on the creative itself. Place the line directly beneath the post, burn it into the first and last three seconds of video, and read it aloud in audio. A profile bio, a campaign landing page or the terms will not be seen.
3
MarketLaunch and measure
Launch it on one campaign and watch conversion weekly. Run the labelled version on a single live campaign for a month and track conversion against the previous period, so the internal argument is settled with your own numbers rather than in the abstract.
Test whether they understood, not whether they noticed. Ask 100 people from the target audience one question: did this show something that really happened? If a meaningful share still say yes, the wording is not doing its job, so rewrite the second line and retest.
What to expect
No change in how real it looks. Perceived realism held steady across every label in every study, so do not judge a disclosure by whether the ad still looks convincing.
A shift in belief, not in behaviour. The wording moved what people thought was true; clicking to buy stayed flat.
Diminishing returns on words. The consequence for the ad did the work; adding the personal warning on top was no better.
Smaller effects on sceptics. Audiences who already assume advertising is embellished respond least to being told it is.
Where it does not work
Invented brands, single exposures. Participants saw one still social post for a fictitious product, not a campaign from a brand they have feelings about.
British samples only, and one of them all women. The first study ran with 121 women aged 18 to 60; the later two were mixed but still UK-based Prolific panels.
Belief in these ads was low to start with. Perceived truthfulness sat below the midpoint everywhere, so the labels were pushing an already sceptical audience further.
Clicks in a survey, not purchases. Choice was a two-button decision inside the questionnaire, with no money involved and no time to reconsider.
Questions marketers ask
Can people tell when an ad has been deepfaked?
No. With no label attached, a deepfaked perfume ad was rated no more synthetic-looking than the genuine ad it was built from, and viewers liked it just as much (Karpinska-Krakowiak and Eisend, 2025, Journal of Advertising).
Does labelling an ad as AI-generated protect consumers?
Not in the form usually used. A byline explaining that images were synthetically generated left perceived realism, ad liking and click-through unchanged. Only labels that stated the consequence of the manipulation reduced belief in the ad.
What should a deepfake disclosure actually say?
Two things: what was altered, and what follows from it. The version that worked named the face swap and then stated plainly that the scene is not real.
Will disclosing a deepfake hurt the campaign?
On this evidence, no. Purchase choice was statistically unchanged across every disclosure condition in all three experiments, even though the stronger labels lowered how truthful the ad seemed.
Why does a disclosure not make the ad look less realistic?
Because realism and truth are separate judgments. Production quality is visible and a label cannot argue with it; whether the thing really happened is a belief, and that is what the wording can change.
An analysis of 391 virtual influencers on X found four distinct ways they write, separated mainly by which pronouns they lean on. Posts addressed to the audience made the character feel more human, and that feeling, rather than the wording itself, carried credibility and intention to act. Whether the influencer was a person or a construct mattered less than who the sentences were pointed at.
People expect to enjoy AI-written music less than human-written music, and then, having listened, rate it about the same. The penalty survives in wordier forms such as writing, but not in sound. It does not translate into acceptance either: listeners still replayed and shared it less, enjoying something they would rather not stand behind.
Source
Karpinska-Krakowiak, M., & Eisend, M. (2025). Realistic portrayals of untrue information: The effects of deepfaked ads and different types of disclosures. Journal of Advertising, 54(3), 432-442. https://doi.org/10.1080/00913367.2024.2306415
This brief is written for practitioners from the published article. Open access (CC BY 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
Viewers who worked out the actor was AI rated the ad 2.83 against 4.59. Being told afterwards hurt far less, and priming people made them twice as likely to notice.
Placing ads by both look and topic of the page lifted predicted attention by up to 10%; matching by category or text alone did not. Evidence from 3,531 ads.
People used AI search for 54% of advice searches but 6% of shopping searches, and clicked in 12% of AI searches vs 59% on Google. What it means for search ads.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.
