Three small words on an ad: "generated by AI."
For ChatGPT they helped. For AXA they hurt.
Same label. Same three words. Opposite effect on purchase.
Everyone Is Starting to Label, Few Know the Price
More brands now mark their AI images. Some because they have to. Some because legal feels safer with it. Almost nobody asks the question that matters for the P&L: what is the price we pay, or the gain we earn, when we use AI and label it that way?
The discussion usually stays on the level of ethics, regulation and "what will people think?" All fair. But "what will people think?" is exactly the kind of question people answer badly. Ask them and they give you the reasonable answer. Whether they buy is decided somewhere else, in the gut, in milliseconds.
So we measured it there.
The Test: Two Ads, One Label
As part of our "Do Pictures Sell?" study, which SUPRA ran with SightEffect and image expert Per Kasch, we put the line "generated by AI" on two ads and looked at what it did to purchase. (The wider study showed about 1,000 people in Germany, in September 2026, versions of ads for about ten seconds each. The effects below are reported results from that research.)
The two ads were a ChatGPT ad and the AXA pension ad from the same study. Two very different products, so the label met two very different customers' minds.
ChatGPT: the label made the picture fit the brand
On the ChatGPT ad, the label made the picture fit the brand better. And for ChatGPT, brand fit is one of the two picture qualities that move purchase. So the label helped.
It makes sense once you say it out loud. A product built on AI shows an AI-generated image and says so. The picture and the promise point in the same direction. If anything, it is rather strange that a brand like that would not showcase what its own model can do.
AXA: the label made the picture feel less honest
On the AXA ad, the same label made the picture feel less honest. And for a retirement product, honesty is one of the two qualities that move purchase. So the label hurt.
The AI label reduces subconscious honesty associations. For an insurance brand, that is a disaster. Nobody hands over 30 years of savings to a brand whose picture feels a little bit fake.
That is no opinion of mine. It is a causally validated impact, measured in this study, in this setting.
The Real Lesson: Labels Don't Have Effects. Perceptions Do.
It would be easy to read this as "AI labels are good" or "AI labels are bad." Both readings are wrong, and both are dangerous, because both give you a rule where you need a measurement.
What actually happened is simple. The label changed one perception, and each category had a different perception that decides the sale:
Same label, different driver
- ChatGPT: the label raised brand fit. Brand fit is one of the two picture qualities that move purchase, so purchase went up.
- AXA: the label lowered honesty. Honesty is one of the two qualities that move purchase for a retirement product, so purchase went down.
- The mechanism is the same. A visual element changes a gut-level perception, and the perception either is or isn't a driver of buying in your category.
This is the same pattern we saw in the rest of the study, where a space-suit picture won on "interesting" and lost on trust. If you want the full story of how one picture can be worth millions, read Do Pictures Sell? The €5 Million Difference Between Look and Feel.
Why You Can't Settle This by Debate
Imagine the meeting. Legal says label everything. Creative says the label cheapens the work. Marketing says customers don't care. All three sound reasonable. All three are guesses about a gut reaction nobody in the room has measured.
And customers can't help you out. If you ask people whether an AI label would change their willingness to buy an insurance product, they will give you a sensible, socially acceptable answer. The honesty penalty happens below that level. That is the say-do gap in a new costume, and it is why so much customer feedback about AI in advertising will mislead you. For the underlying logic, see why customers can't tell you why they buy.
The alternative is to measure the subconscious associations the label triggers, then use Causal AI to infer which of them drive the choice in your category. That is what deep implicit research does. It is also hard to fool, because people can't talk their way around a reaction they don't know they have.
What to Do Before You Label (or Don't)
If you use AI imagery, and more and more brands do, here is a short sequence worth running before the legal text goes on the picture.
First: find your two drivers. For every category, a small number of picture qualities decide the purchase. For ChatGPT it was brand fit, for a retirement product it was honesty. Know which two are yours before you touch the image.
Second: test the label as its own variable. Same picture with and without the label, everything else unchanged. If you test the AI picture against an old photo and the label at the same time, you will never know which one did the damage or the good.
Third: decide with the number, not the mood. If the label is required, and sometimes it is, the result tells you something different: you need a picture that carries your driver without relying on the label. A required label that costs honesty is a creative brief, not a verdict.
My summary is short: AI use is a brand decision with a measurable price, and most brands are making it blind. I would rather not make a decision like that on a hunch, and I don't think you should either.
In my book THE TOP 5%, I describe how the brands that grow stopped guessing at exactly these moments. To see where your own decisions still rest on guesses, the AI Diagnostic is the shortcut.
The AI label paradox: frequently asked questions
Does labelling an image 'generated by AI' hurt advertising performance?
It depends on the category. In SUPRA's Do Pictures Sell? research with Per Kasch, the same label had opposite effects on two ads. On a ChatGPT ad it improved brand fit, which moves purchase for that brand, so it helped. On an AXA retirement ad it reduced honesty associations, which move purchase for that product, so it hurt. The label itself has no fixed effect; the perception it changes does.
Why did the AI label help ChatGPT but hurt AXA?
The label changed a different gut-level perception in each case, and each category has different perceptions that decide the sale. For ChatGPT, the label made the picture fit the brand better, and brand fit is one of two picture qualities that move purchase. For AXA, the label made the picture feel less honest, and honesty is one of two qualities that move purchase for a retirement product.
Should brands label AI-generated images?
Where labelling is legally required, brands have no choice, and the question becomes how to build a picture that still carries the key purchase driver. Where it is optional, the decision should rest on measurement, not on mood. Test the same picture with and without the label and see which gut-level perceptions change, and whether those perceptions drive purchase in your category.
How can you measure what an AI label does to purchase?
Show matched versions of an ad, with and without the label and everything else unchanged, to a sizeable sample for a few seconds. Capture gut-level reactions rather than stated opinions, then use Causal AI to infer which perceptions, such as brand fit or honesty, drive purchase propensity. SUPRA's deep implicit research follows this logic and avoids the say-do gap of asking people directly.
Dr. Frank Buckler advises leadership teams on growth decisions that depend on why customers choose. He founded SUPRA, holds a PhD in marketing and was among the first to apply causal AI to marketing questions. His latest book is THE TOP 5%.