Bite-size: When people think an ad was made with AI, they trust it less and value the product less. Belief does the damage: the same ad rates lower with an AI label, and so do ads people wrongly suspect are AI. Better AI won’t fix this, because people already can’t tell the difference. What’s growing is suspicion.
Full-size:
Last month my five-year-old stood in the hair-care aisle looking at a box of color hairspray for Halloween. Nobody had mentioned AI. She pointed at the woman on the package and said the picture was “that looks like AI.” Then she asked, “Is it actually going to work?”
She had no way to know. In a 2024 study presented at the Vision Sciences Society, children aged 5 to 12 and adults tried to sort real faces from AI-generated ones. Both groups scored about 47%, which is basically a coin flip. The only difference was that kids said “AI” more often. My daughter guessed. The guess alone made her doubt the product.
That is the entire effect of what’s happening, and it’s been measured many times in more ways than one.
Kirk and Givi (2025) ran seven preregistered experiments. When people believed an emotional marketing message was written by AI, they recommended the brand less and felt less loyal to it. The message was the same, it was only the belief changed. People found the AI-labeled message less authentic, and it left them feeling that something was off.
Bellaiche and colleagues (2023) showed people the same artworks, labeled “human-made” or “AI-made.” The AI label lowered every rating. It barely changed how beautiful people thought the art was, but sharply lowered how meaningful and valuable they thought it was.
The Nuremberg Institute for Market Decisions ran two versions of this test. In the first, with 1,000 respondents each in the US, UK and Germany, an identical product ad was described either as a photo or as AI-generated. The AI version was rated lower, especially on emotional measures, and fewer people wanted to click through. In the second, German participants saw six ads labeled as AI- or human-made. The AI label made the ads seem less authentic and less useful, and lowered willingness to buy. The penalty was smaller for high-tech products and larger for everyday ones (like hairspray).
People already discounted ads
This is a new version of an old habit. Calfee and Ringold reviewed six decades of surveys going back to the 1930s. In every decade, about 70% of consumers said advertising is often untruthful. They kept using ads to decide what to buy anyway.
Friestad and Wright (1994) explained how skepticism and purchase coexist. The moment people notice they’re being sold to, they stop judging the message and start judging the seller’s motive. Campbell and Kirmani (2000) showed a single cue is enough to set this off.
An AI image is that cue, and it works at the most basic level. A product picture is a claim. The woman on the hairspray box is saying “this is what it does to your hair.” If she isn’t real, the claim isn't either. My daughter didn’t need to know anything about AI to get there. She’d heard enough adults call things fake to have a category for it, and five-year-olds already doubt sources they’ve learned not to trust. A fake picture read as a fake promise, and a fake promise made her ask whether the product works.
For the adults in the room, there’s a second cost. Kirmani and Wright (1989) found that when an ad looks expensive, people assume the company expects the product to earn the money back. Expensive advertising reads as confidence in the product. Ambler and Hollier (2004) summarized it as “the waste in advertising is the part that works.” An AI image signals that nothing was spent, so the brand loses the one signal that used to offset the old discount.
Koning and Voorveld (2025) confirmed the trigger directly. They put an “AI-generated” label on an Instagram ad, and trust in both the ad and the brand fell, through the same persuasion-knowledge mechanism. The “moral disgust” in Kirk and Givi’s data is the feeling the manipulation research has described for thirty years. An AI tell is a new cue for an old reaction.
“The models will get better”
The usual response is that AI output will improve until nobody can tell. I believe we’re already there yet the penalty remains.
Nightingale and Farid showed in 2022 that people could not tell AI-generated faces from photographs. They rated the fake faces as more trustworthy. Four years and several model generations later, the spring 2026 Yale Youth Poll found that 18- to 34-year-olds, the most confident group, identified AI headshots about half the time.
Music shows the same pattern. Ansani and colleagues (2025) replayed real human piano performances on a self-playing piano. In one video, a pianist mimed along. In the other, the bench was empty, and viewers were told an AI was playing. The audio was identical. Viewers rated the “AI” version lower on liking, engagement, emotion, and quality. Then 79% said they had noticed differences, such as timing that was “too precise.” There were no differences. The label created them. In a separate blind test in 2025, Deezer, the Paris-based streaming platform, commissioned across eight countries, 97% of people could not pick the AI tracks.
So people can’t detect AI, and they still penalize it. The penalty comes from what they believe, not from what they can see. Improving the output doesn’t change the belief.
What is changing is how much people suspect. In a May 2026 Center for Media Engagement experiment, only 14% of readers thought an article had no AI involvement. That included readers of the original, fully human-written version. Older readers assumed more than half of it was machine-written. The more AI readers assumed, the lower they rated the piece on accuracy, trust, and enjoyment. Disclosure labels went unnoticed and changed nothing. When people suspect everything, every brand pays the discount, even on work humans wrote.
The generation growing up with it
This isn’t only an adult problem. IAB’s January 2026 survey found 39% of Gen Z feel negative about AI ads, nearly double the 20% of Millennials. The piano study found the penalty was largest among people who already disliked AI. So the youngest adult consumers are both the most hostile and the harshest judges. The generation behind them is forming its instincts now.
Those instincts are learned. Brink, Gray and Wellman (2019) found that children under about nine don’t find human-like robots creepy. The discomfort appears around age nine, once kids understand what a mind is and can notice something imitating one. My daughter is five. Her suspicion didn’t come from the image. It came from hearing adults call things fake, and from a culture that already treats “made by a computer” as “cut a corner.”
That suspicion won’t stay limited to pictures. The same reasoning will apply to copy that sounds like every other brand, to a tone that fits everyone and belongs to no one, and to insights no one actually lived. Once “AI-ish” is a category in a consumer’s head, it gets applied to anything that triggers it, right or wrong.
What to do
Use AI for factual and functional work. Kirk and Givi found almost no penalty on factual messages. Specs, store hours, comparison tables, routine emails. Use whatever works.
Keep real people on emotional and heritage work. That’s where the penalty is largest and where the brand’s meaning lives. The woman on the hairspray box is doing emotional work. She needs to be real.
Spend where people can see it. The cost signal only works if the cost is visible. A real location, a real face, a production the audience can tell was expensive. That spending is part of the message.
When you disclose, say what the humans did. The penalty shrinks when AI only edited human work, and people assume full automation unless told otherwise. “AI-assisted” reads as “AI-made.” “Shot in Austin, retouched with AI” doesn’t.
Test for suspicion, not detection. Your concept test asks whether the image looks good. Add one question: “Who do you think made this?” If many people say “a computer,” you’ll pay the discount whether they’re right or not.
Sources
Ambler, T., & Hollier, E. A. (2004). The waste in advertising is the part that works. Journal of Advertising Research, 44(4).
Ansani, A., Koehler, F., Giombini, L., Hämäläinen, M., Meng, C., Marini, M., & Saarikallio, S. (2025). AI performer bias: Listeners like music less when they think it was performed by an AI. Empirical Studies of the Arts, 43(2), 1137–1161.
Bellaiche, L., et al. (2023). Humans versus AI: Whether and why we prefer human-created compared to AI-created artwork. Cognitive Research: Principles and Implications, 8, 42.
Brink, K. A., Gray, K., & Wellman, H. M. (2019). Creepiness creeps in: Uncanny valley feelings are acquired in childhood. Child Development, 90(4).
Buder, F., & Unfried, M. Transparency without trust. NIM Insights Research Magazine, 7.
Calfee, J. E., & Ringold, D. J. (1994). The 70% majority: Enduring consumer beliefs about advertising. Journal of Public Policy & Marketing, 13(2), 228–238.
Campbell, M. C., & Kirmani, A. (2000). Consumers’ use of persuasion knowledge. Journal of Consumer Research, 27(1).
Center for Media Engagement (May 2026), via INMA.
Chien, S. H.-L., Liu, P.-T., Ma, Y.-T., & Chang, C.-K. (2024). Like adults, children cannot distinguish between AI-synthesized faces and real ones but exhibit a response bias toward synthesized faces. Poster, Vision Sciences Society; abstract in Journal of Vision.
Deezer & Ipsos (November 12, 2025). Survey of 9,000 adults in eight countries.
Friestad, M., & Wright, P. (1994). The Persuasion Knowledge Model. Journal of Consumer Research, 21(1).
IAB & Sonata Insights (January 2026). The AI Ad Gap Widens.
Kirk, C. P., & Givi, J. (2025). The AI-authorship effect. Journal of Business Research, 186.
Kirmani, A., & Wright, P. (1989). Money talks: Perceived advertising expense and expected product quality. Journal of Consumer Research, 16(3).
Koning, B., & Voorveld, H. A. M. (2025). Disclaimer! This content is AI-generated. Journal of Interactive Advertising, 25(3).
Nightingale, S. J., & Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. PNAS, 119(8).
Yale Youth Poll (spring 2026), via The Nation.
