Does the “Made with AI” label hurt engagement? What a Journal of Consumer Research study found
A million TikTok posts and eight experiments agree that the label costs you likes. The reason is not what anyone assumed, and it points at a fix that costs nothing at all: change what the label says.
Posts carrying TikTok’s AI label collected 7% to 8% fewer likes at the same reach. Not because people distrust AI. Because the label says the creator did not try.
The whole setup says effort. Three words under the post can undo it. Photo: Daniel Romero on Unsplash.
The short answer
Yes, and for a reason almost nobody guesses. Across TikTok's first nine months of labelling and eight preregistered experiments, posts carrying an AI disclosure collected fewer likes at the same reach. It is not that people think the content is worse, and it is not a general dislike of AI. The label reads as a signal that the creator did not put much in, which weakens the one-sided bond viewers feel toward them, and people like posts by people they feel close to. Change the label to name a specific tool and the penalty disappears.
Key takeaways
Labelled posts received 7% to 8% fewer likes than the same creator's unlabelled posts at the same number of views.
The penalty was worse on still images, at around 12%, than on video, at around 8%.
Only about half the people in each experiment noticed the label at all. Among those who did, the effect was roughly twice as large.
By the authors' own back-of-envelope sum, each extra percentage point of labelled posts costs TikTok about 382 million likes a month.
What the researchers did
In September 2023 TikTok became the first big platform in the United States to put an AI label on posts. That
gave Stephan Carney, Ignacio Riveros and Stephanie Tully, all at USC Marshall, a natural before and after to
work with, and they used it properly.
First the field data: every post from roughly ten thousand randomly chosen American TikTokers for three months
before the policy and six months after, with the engagement each one collected and whether it carried the
label. Then, because that can only tell you what happened and not why, eight preregistered experiments, a
content analysis of four hundred videos stripped of their interface, and a linguistic sweep of nearly two
million comments.
Their hypothesis was unusual. Most people assume an AI label works by raising suspicion: this might be fake,
this might be low quality, I do not trust it. The authors proposed something else, borrowed from the research
on why people like each other. We judge a relationship partly by how much effort the other person appears to
put in. A viewer and a creator have exactly that kind of one-sided bond, and a label announcing AI is read as
a creator who did not bother.
Field data
1,135,817 TikTok posts from 8,650 creators, spanning the policy launch
Comments analysed
About 1.9 million, across 56,720 posts
Experiments
Eight preregistered studies, 3,396 participants in all
Content checked
400 videos transcribed, scored for brightness, sharpness, motion and theme, with the interface stripped off
Outcome measured
Likes, comments and shares on the platform; likelihood of tapping like in the experiments
What happened on TikTok
Comparing posts by the same creator in the same month, and holding views constant, posts carrying the label
collected less of everything. Fewer likes, and less combined engagement once comments and shares were added
in.
Holding views constant matters more than it sounds. Labelled posts were also shown to far fewer people, which
could be the algorithm or could be the creator posting different sorts of thing. By conditioning on views the
authors are asking a narrower and more useful question: of the people who did see this, how many tapped like.
The answer is fewer.
One finding should worry anyone running influencer budgets. The effect did not belong to small accounts. The
authors split the data by follower count, into accounts under a thousand followers, accounts up to ten
thousand, and accounts above that, then split it again by how many views a creator had been getting before the
policy existed. The penalty turned up across almost every group, and there was no significant interaction with
audience size. Whatever this is, it happens to the mid-tier creators brands actually hire.
It is worth being honest about the raw numbers behind all this. In the post-policy window the dataset held
819,227 unlabelled posts and only 2,749 labelled ones. The label was rare, and the people applying it
sometimes chose to.
It is not the quality, and it is not AI
Three separate attempts were made to kill the obvious explanations, and all three failed to save them.
The first took 200 pairs of videos from the same creators, one labelled and one not, stripped off the TikTok
interface and ran them through transcription, language analysis and computer vision: brightness, colourfulness,
sharpness, motion, scene cuts, topic. Feeding all of that into the model did not make the label's effect go
away.
The second showed those same videos, without any label visible, to 697 students who use TikTok heavily, for
5,277 ratings. Perceived quality came out at 3.04 for the unlabelled videos and 2.98 for the labelled ones, a
difference so small it barely registers, and once it was added to the model it predicted nothing while the
label kept predicting. Creativity ratings did not differ at all. Here is the detail I find most telling:
viewers guessed AI had been used in 36% of the labelled videos and 8% of the unlabelled ones, and the videos
people believed were AI did not lose engagement. Believing is not the same as being told.
The third read the comments. If people thought labelled posts were worse, the comments should be more
negative. They were slightly less negative. If people were worried about being deceived, there should be more
language about risk or right and wrong. There was not. What did shift was the balance of pronouns: comments on
labelled posts used relatively fewer personal words like I, you and she, and more impersonal ones like it and
that. People were talking about the content rather than to the person who made it.
Finally, the AI aversion story. In one experiment 56% of participants said they liked AI and fewer than a
quarter said they disliked it, the effect appeared anyway, and neither their attitude to AI nor how much they
used it changed the size of it.
What the label actually says
So what is left? Effort.
Shown the same video with and without a label, people rated the creator as having put in substantially less
work when the label was there. Their sense of connection to that creator fell with it, from 4.01 to 3.80
overall, and from 4.01 to 3.43 among the people who actually noticed the label. A formal test of the chain,
label to perceived effort to connection to engagement, came out significant, and once effort and connection
were in the model the label had nothing left to explain on its own.
That is the whole mechanism, and it is a social one rather than a technological one. Nobody is deciding the
video is fake. They are deciding the person behind it did not show up, and a like is a small social act you
perform for someone you feel something for.
How much effort people thought the creator had put in, out of 7
How much effort people thought the creator had put in, out of 7
Group
Value (out of 7)
With an AI label
4.98
Same video, no label
5.9
Study 5: 385 participants shown an identical video with and without the disclosure.
Owning up does not help
The intuitive fix is honesty with context: say it yourself, in your own voice, and explain that AI only did
something trivial. One experiment tested exactly that, with 565 people and a photograph of a Japanese
breakfast by a window, where AI had added falling snow. The caption version read like a creator being cheerful
about it: check out the snow effect I added.
It made no difference whatsoever. Disclosing it yourself scored the same as being labelled by the platform,
and having both was no worse than either. Every version carrying any admission of AI landed well below the
version with none.
Read that against the mechanism and it makes sense. The caption was not reassuring anyone about deception,
because deception was never the issue. It was confirming the thing that costs you: a creator who reached for
a tool instead of doing it.
How likely people were to like the post, by who admitted the AI
How likely people were to like the post, by who admitted the AI
Group
Value (out of 7)
Creator said so in the caption
3.27
Both
3.31
Platform label only
3.34
No disclosure anywhere
4.46
Study 4: 565 participants, one photograph with AI-added snow, disclosure varied two ways.
The wording is the whole lever
Here is the finding worth the price of admission. In the last experiment, 198 people saw the same posts with
three different labels: none, a generic AI disclosure, and one naming a specific tool, Photoshop neural
filters. Photoshop is AI. It can generate and alter an image as thoroughly as anything else. But a pretest
confirmed people read it as requiring more work.
The generic label cost engagement. The specific one cost nothing measurable at all, landing level with no
label. Sense of connection to the creator followed the same shape. And because the actual content was shuffled
between the three label conditions, the content itself had no effect in the model. The words under the post
were doing all of it.
That hands platforms a design decision rather than a trade-off. YouTube Shorts has already started naming
tools in its disclosures rather than announcing AI in the abstract. On this evidence that choice keeps the
transparency and drops most of the penalty.
There is also a warning in here for policy, and the authors make it plainly. If a label reliably costs
creators reach and money, creators will find ways around it, which defeats the point of having one. A
disclosure regime that punishes honesty does not produce honesty. My own reading, beyond what they tested: the
same logic should apply to any provenance label a brand is now being asked to carry, from AI notices to
sponsorship tags. What the label says about the person behind the work is doing more than the fact of the
label.
How likely people were to like the post, by what the label said
How likely people were to like the post, by what the label said
Group
Value (out of 7)
"Made with AI"
3.17
"Made with Photoshop Neural Filters"
3.85
No label at all
3.98
Study 6: 198 participants saw all three, with the posts shuffled between labels.
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
Check what is already labelled. Open your last 30 posts and see which ones carry an AI label. Platforms add one on their own when you use their filters, captions or voice effects. Most people do not know it is there.
Keep AI off the posts where you are the point. List the posts where a person is the reason anyone watches: a face talking, a founder answering a question, a reply to a comment. Use AI on the other ones.
2
SetBuild it
Name the tool, not the technology. When you can write the words yourself, say which tool you used and what it did. "Edited with Photoshop neural filters" costs you nothing. "Made with AI" costs you likes.
Put the work back into the post. Show a few seconds of the making: your hands, the rough version, the thing you tried first and binned. That is what the label just took away, so give it back inside the post.
3
MarketLaunch and measure
Count likes for every thousand views. Never compare raw likes, because labelled posts get shown to fewer people. Divide likes by views instead, then compare your labelled posts against your unlabelled ones. Thirty of each is enough.
Never take a label off to fix the numbers. If engagement drops, change the wording and change the post. Do not remove the disclosure. Platforms are getting better at adding it for you, and being caught costs more than the likes ever did.
What to expect
A steady tax, not a collapse. The hit to engagement at a given reach is a few percent. It matters because it compounds across every post and feeds the algorithm, not because any one post dies.
Reach is a separate problem. In the raw data, labelled posts reached a fraction of the audience before anyone chose whether to like them. The study cannot say why, but budget for both effects.
Half your audience never sees the label. That caps the damage today. It also means the penalty grows as platforms make disclosures more prominent.
Specific wording is free to implement. Changing a generic label to a named tool costs nothing, needs no new content, and in this test recovered the whole difference.
Where it does not work
One platform, one policy, nine months. TikTok's first labelling window, in the United States. Other platforms place and word their labels differently, and Instagram buries its one in the description.
Labelled posts were rare and partly self-selected. Creators applied the label themselves when they used outside tools, so the labelled group is not a random sample of AI-assisted content.
The experiments measured intention. Participants said how likely they would be to tap like. Only the TikTok data records what people actually did.
Norms are moving underneath this. The authors note the effect could shrink as AI use becomes ordinary, or grow as the tools get easier and the perceived effort falls further.
Questions marketers ask
Does the "Made with AI" label reduce engagement?
Yes. Holding reach constant, a disclosure knocked a single-digit percentage off a post's likes in TikTok's own data, and eight controlled experiments reproduced the drop with the content held identical (Carney, Riveros & Tully, 2026, Journal of Consumer Research).
Why would a label reduce engagement if people do not mind AI?
Because it is read as a statement about the creator rather than the content. We judge how much someone cares by how much work they appear to put in, and an AI disclosure reads as very little. What falls is the sense of closeness a viewer feels toward that person, and tapping like is partly an expression of it.
Should I just not label AI-generated content?
No. Platforms increasingly apply the label automatically, so the choice is often not yours, and being caught hiding it costs far more than the engagement does. The research points at the wording and the content around the label as the places where you have real room to move.
Does it help to say in the caption that AI handled only a minor detail?
It did not help at all in a test built for exactly that question. Volunteering it in the caption, and naming the small background detail AI had handled, produced the same drop as a bare platform label. What registers is the admission, not how minor you make it sound.
Is there any way to disclose AI without losing engagement?
Naming the specific tool appears to work. Posts labelled as made with Photoshop neural filters performed level with posts carrying no label at all, while the same posts under a generic AI label dropped sharply. Photoshop is no less capable of generating an image; it just reads as something a person operates.
A fashion brand running body-positive advertising with AI-generated plus-size models, and disclosing it, paid for the disclosure. The route ran through a drop in felt human presence and then into perceptions of brand hypocrisy, which is a rough place for a campaign about accepting real bodies to end up. Two things helped: making the cause message concrete rather than vague, and adding a fuller transparency note on top of the bare disclosure. Wording again, rather than whether to tell people.
Journal of the Academy of Marketing Science · 2026
Klostermann, J., Meißner, M., Musalem, A., & Decker, R.
Across Instagram posts for watches and shoes, the posts that pushed the product into the background collected more likes and comments, while the posts that made the product the centre of the picture drove more engagement on the brand's own later posts. The two outcomes pull apart, which means the influencer and the brand paying them are optimising for different pictures. Worth settling which one you are buying before the brief goes out.
Source
Carney, S., Riveros, I., & Tully, S. M. (2026). Made With AI: Consumer Engagement With Social Media Containing AI Disclosures. Journal of Consumer Research. https://doi.org/10.1093/jcr/ucag013
This brief is written for practitioners from the published article. Open access (CC BY-NC 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
A language model guessed the star rating of 28,900 reviews from the text alone. When it missed, human readers missed too, and found the review less helpful.
Thumbnails with a middling amount going on won the clicks at 5.48 out of 7. Middling colour harmony lost them at 3.41. Two dials, pulling opposite ways.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.
