AI generated ad creative has two findings attached to it that are worth knowing together, because separately each is misleading. It lifted click-through rates by up to 19% in one study. Disclosing the AI involvement reduced effectiveness by up to 31.5% in the same work.
Both come from research rather than from vendor marketing, which puts this topic in a different category from most of this cluster.
This is for you if you are deciding whether generative tools belong in your creative pipeline.
What the first study measured
Lee, Todri, Adamopoulos and Ghose, working across NYU Stern and Emory, published an analysis in October 2025 finding generative AI creative lifted click-through rates by up to 19%.
The second half of the same work is the part that gets omitted. Disclosing that AI was involved reduced effectiveness by up to 31.5%.
Read those together and the finding is not that AI creative works better. It is that AI creative works better when nobody says so, which is a considerably more awkward result.
Neither number should be quoted alone. Doing so produces either an unjustified endorsement or an unjustified warning, and both misrepresent what was measured.
What the larger analysis found about AI generated ad creative
A separate study by Exner, Hartmann, Netzer and Zhang, across TU Munich, Columbia and Harvard, covered more than two million ad-day observations, 7,074 advertisers and over 16 billion impressions.
Its findings sharpen the picture. Ads perceived as human-made achieved 0.79% click-through. Ads perceived as AI achieved 0.62%. Human-made ads sat between 0.55% and 0.67%.
Roughly 45% of the AI-produced ads were rated as likely human-made.
And there was no significant difference in conversion between AI and human creative.
That last point matters most. Whatever is happening is happening at the attention stage rather than the persuasion stage, so a click-through lift does not automatically become a revenue lift.
Reading the two studies together
The consistent thread is perception rather than origin.
AI creative that reads as human performs best. AI creative that reads as AI performs worst, below human creative. And a large share of AI output does read as human, which is why the average looks favourable.
So the useful framing is not whether to use these tools. It is whether the output announces itself.
That maps onto everything else in this cluster. Adverts that look like adverts get filtered. Adverts that look machine-produced appear to get filtered too, and by a similar mechanism.
The tool is not the variable. The tell is.
Why disclosure is not a performance decision
The 31.5% figure will tempt somebody to conclude they should not disclose, and that conclusion does not follow.
Whether you must disclose AI involvement is a legal and regulatory question that varies by market, by industry and by what claim you are making. Some jurisdictions and platforms have requirements. Some client contracts do.
A measured performance cost does not change an obligation. It tells you what the obligation costs, which is different.
So take advice on what you are required to do, and treat the research as information about the price rather than as an argument for avoiding it. Presenting it any other way would be recommending a course of action on the basis that it performs better when concealed, which is not a recommendation I would make.
The real argument for AI generated ad creative
Volume, and it holds without either study.
If generative tools let you produce more distinct angles for the same budget, that is the case. Angle count is normally the binding constraint on account performance, and roughly one in twenty creatives becomes a genuine winner, so more attempts is how you find them.
That argument does not depend on a click-through lift, does not depend on perception, and does not require anybody to be uncomfortable about disclosure.
It also sets the right expectation. You are buying supply, not quality. The tool does not know your customers, has not read your sales calls and cannot invent the argument. It can produce many executions of an argument you supply.
Which means the scarce input is unchanged. You still have to know what to say.
That reframing also settles a question people find difficult. Whether AI generated ad creative is somehow illegitimate is not really the issue, because nobody asks whether a designer used a particular filter. What matters is whether the advert says something true, specific and worth reading, and that judgement sits with you regardless of what produced the pixels.
The risk in AI generated ad creative nobody prices in
Sameness, and it is structural rather than a quality problem.
Everybody has access to similar tools, prompts them with similar briefs, and produces output drawn from similar training. The result is a feed where a growing share of adverts share a visual and verbal register.
That erodes exactly what you were buying. An advert that looks like every other advert gets categorised and skipped, which is the failure mode this whole cluster is about.
So the tools solve supply and can worsen distinctiveness. Those pull in opposite directions and the balance is worth watching rather than assuming.
The practical defence is the same as everywhere else. Bring a specific argument, drawn from your own customers, that nobody else’s prompt contains.
How to use these tools in a real pipeline
Three rules that follow from the evidence.
Use them for volume and variation, not for the argument. Supply the angle yourself, from voice of customer material, and let the tool produce executions.
Check whether the output announces itself. If a piece reads as machine-produced, the research suggests it performs below human creative rather than above it, so that is a quality gate worth applying.
And measure on conversion rather than click-through, since the larger study found the difference disappearing at that stage. An improvement in clicks that does not reach revenue is not an improvement worth scaling.
What to change this week
Three steps.
If you already use generative tools, check whether your comparison has been on click-through or on cost per result, because the evidence says those may disagree.
Then establish what you are actually required to disclose, in your markets, before the question arrives from somewhere less convenient.
Then check your recent AI-assisted output for the tell. If it reads as machine-produced, that is a fixable quality problem rather than a reason to stop.
Fund the ad set and read what the algorithm chose is the position underneath, creative drives sales covers why the argument matters more than the execution, creative volume covers the supply problem these tools address, and the weekly rhythm I run covers fitting them into a cadence. Getting the landing page to match the advert’s register is web development work. To review whether your creative pipeline is producing variety or repetition, book a teardown.