Three thousand people were shown advertisements and could not reliably say which ones a machine had made. The advertisements the machine made still performed worse.
Both halves of that sentence come from the same study, and holding them together is the useful part. The thing that failed was invisible to the people it failed on.
How the study was built
Ipsos ran it with Syracuse University’s Newhouse School: 20 advertisements across 10 brands, tested on 3,000 respondents in the United States. The brands are ordinary and recognisable. Cheerios, Chewy, Febreze, Fiat, H&M, Old Navy, Herbal Essences, Ray-Ban Meta, TurboTax and Visa.
The design is what makes it worth reading rather than another headline about AI content.
They took human-made advertisements produced before 2021, so no AI tool could have been involved in making them. Then they paired each one with a fully AI-generated version built from the same strategic brief. The brief is the document an agency actually works from, which means both versions were asked to solve an identical problem with the same information.
That is a fair comparison, and fair comparisons in this area are rare.
Nobody could spot it
Only 13 per cent of viewers who saw an AI-generated advertisement were at least somewhat confident it was made by AI.
The control is what makes that number mean something: the same 13 per cent wrongly suspected the human-made advertisements were machine work. Forty per cent of all viewers were uncertain either way.
So on the question people usually ask, whether you can tell, the answer is no. Not “not always”. The suspicion rate was identical whether or not there was anything to suspect, which is what guessing looks like.
It did not matter
Against Ipsos’s benchmark, the human-made advertisements over-indexed by 11 points on average. The AI-made versions under-indexed by five.
The measures were purchase intent and brand recall. Not whether people enjoyed the advertisement, not whether they rated it as well made, not whether they could tell who made it. Whether it worked.
Sixteen points of benchmark separate the two groups, on outcomes, produced by work that the audience could not distinguish.
Where the gap opens
The study found AI did best on simple, product-driven briefs. It fell away when the brief called for storytelling, emotion or a point of view.
That is a usable line, and it generalises past advertising. The further a piece of work sits from stating a fact, and the closer it sits to making somebody feel something or believe something, the wider the distance between looking finished and being effective.
Product descriptions, specification summaries, straightforward explanations: narrow gap. Anything that has to persuade, reassure, or establish that you understand somebody’s situation: wide gap, and invisible.
The finding that is not about advertising
Here is the management consequence, and it is the reason this study is worth an hour of your attention even if you never commission an advertisement.
The risk everybody prepares for is obviously bad output. Wrong facts, clumsy writing, an image with too many fingers. That risk is largely handled, because somebody notices. It is the failure mode every review process was built to catch.
The risk nobody has a process for is output that looks good enough to approve.
Think about how work gets signed off in your organisation. Somebody senior reads it and forms a judgement. That judgement is the instrument, and it is the only instrument in the room. This study is 3,000 people establishing that the instrument does not detect the problem, because the problem is not visible in the artefact. It shows up later, in whether the thing worked.
Approval and effectiveness have come apart. Only one of them is observable in the meeting.
What that means in practice
Two things follow, and neither is “stop using AI for this”.
The first is that looking right stops being evidence. If a piece of work is meant to change what somebody does, the review question cannot be “is this good”. It has to be “how will we know whether this worked”, answered before the thing ships, with a measure attached.
The second is that the sorting matters more than the quality control. Work that states facts is where the tool is genuinely strong and the gap barely exists. Work that has to carry a point of view is where the gap is widest and hardest to see. Deciding which pile a task belongs in is now a real editorial judgement, and it is one a non-specialist can make.
Why this is not an argument about taste
There is a version of this finding that gets used badly, and it is worth heading off.
It is not that AI-made work is tasteless, or soulless, or that human creativity is irreplaceable. Those claims are unfalsifiable and this study does not test them. It tested whether people could tell, and they could not, which is the opposite of the taste argument.
What it found is narrower and more awkward. The work was indistinguishable and less effective at producing the behaviour it existed to produce. Something is being lost that does not show up in how the work looks, and the audience cannot articulate it either, because they could not identify which was which in the first place.
That is a harder finding to act on than “it looks worse”, and a more honest one. It puts the burden on measurement rather than on judgement, which is uncomfortable, because judgement is free and measurement is not.
Read the dates and the credits
Three things to be straight about before any of these numbers get repeated.
The research is from 18 May 2026. The article that brought it back into view is from September. The findings are four months old and should be described that way rather than as new research.
That article was written by Adam Peruta, one of the study’s own authors, so the write-up and the evidence come from the same place. It is a fair account, and it is not independent coverage.
Ipsos is a commercial research firm, a co-author of the study, and the benchmark the advertisements were indexed against is Ipsos’s own. Attribute the figures to Ipsos and Newhouse rather than presenting them as neutral. That does not make them wrong. It makes them attributable, and this audience checks.
One more, because it is circulating: there is a separate paper about AI disclosure labels with a 31 per cent figure in it. Different study, different question. Do not merge them.
The one-line version
Three thousand people could not tell. It sold less anyway. Whatever your review process is currently measuring, it is not that.