Advertisers Think 82% Like Their AI Ads. The Real Number Is 45%.

There’s a number most advertising teams have not seen, and it describes them.

Ask ad executives how Gen Z and Millennial consumers feel about AI-generated advertising, and 82% of them say those consumers feel positive about it. Ask the consumers, and 45% do. That is a 37-point gap between an industry’s belief about its audience and the audience’s report of itself, and it comes from IAB and Sonata Insights, who surveyed 505 US Gen Z and Millennial consumers who engage with ads alongside 104 US ad executives at companies spending at least $1 million a year on media.

Two things about that research before it gets used as a cudgel. It was published on January 15 and the fieldwork ran from October 2025 into January 2026, so it isn’t new. And the gap is the finding that survives its own age, because the same team ran a comparable study in 2024 and measured 32 points. The distance keeps moving. Over two years it grew by five points while the industry got more confident and more committed.

More committed is measurable too. In the same window, the share of ad executives reporting they deploy AI in the creative process went from 60% to 83%. The industry roughly doubled down on a technique while its read on the audience drifted further from what the audience says. That describes a feedback loop that’s come loose.

The audience knows what it’s looking at

The comfortable reading here is that consumers say they dislike AI ads and then respond to them anyway. We’ve made that argument ourselves. Last week we covered a survey where 69% of shoppers said they’d let an AI buy for them while 39% check every recommendation against four or more sources, and the honest conclusion was that stated preference is a poor guide to behavior. That caution applies here, and it’s the strongest objection to everything below. People are unreliable narrators of their own reactions.

What makes this case different is that stated preference isn’t the only signal. The other two are behavioral.

The first is detection. Pangram, which builds AI-text detection, scanned 1,002,627 posts across LinkedIn, Medium, Substack, X and Reddit, collected through a browser extension from users who agreed to share what they scrolled past. More than 40% of LinkedIn posts longer than 250 words came back flagged as fully AI-generated. LinkedIn made up about a third of everything scanned and accounted for roughly two thirds of everything flagged. Whatever you think of any single detection call, the shape is hard to argue with: on the platform where B2B marketing lives, a large share of the long-form writing is machine-made, and readers are swimming in it.

The second is platform behavior, and platforms don’t act on stated preference. They act on engagement data they own. On July 30, LinkedIn added a button letting users report a post as “seems like AI slop” and said it would use that signal to decide how far a post travels beyond the author’s own network. Hari Srinivasan, LinkedIn’s chief product officer, put it plainly:

“AI slop is a top priority for all of us. We really care about this. People come to LinkedIn to connect with real people and share their real perspectives, ideas, and expertise.”

In the same announcement LinkedIn retired its own “enhance your post” AI writing feature and replaced it with a tool that proofreads without changing the writer’s voice. That’s the detail worth sitting with. A platform with every commercial reason to keep people generating content withdrew its own generation tool and kept the correction tool. It also said it blocks hundreds of thousands of automated comment attempts a day.

Three independent readings, three different methods, one direction. Consumers report discomfort. Detectors find saturation. The platform is spending engineering effort to push the stuff down. The industry’s confidence is the only line moving the other way.

What the gap is actually made of

The data gets specific about the shape of the disagreement, and it isn’t a general grumble about technology. Consumers were more than twice as likely as executives to associate “manipulative” with brands using AI in advertising, at 20% against 10%, and more than twice as likely to reach for “unethical,” at 16% against 7%. Gen Z runs harder than Millennials: 39% negative against 20%.

Those are brand-attribute words. When your audience reaches for manipulative, the cost doesn’t show up as a worse click-through rate on one campaign. It lands on the thing we’ve argued is the only durable asset left, back when AI turned content, SEO, ads and email into a commodity everyone runs. If execution is no longer an edge and brand is what remains, spending brand trust to save production time is a bad trade made invisible by the fact that nothing on the dashboard reports it.

We have watched the individual version of this happen. When fans caught a single chatbot-sounding phrase in a science video, Hank Green paused his channels and apologized to an audience of 32 million subscriptions. One phrase. The reaction was not proportional to the offense, which is the tell that it was never about the phrase.

The part the industry keeps skipping

Here’s the finding that makes the gap self-inflicted rather than tragic. In the same research, 73% of consumers said knowing an ad used AI would either increase their purchase likelihood or leave it unchanged. 56% want disclosure on AI video. 52% want it when an ad is entirely AI-generated.

Read those together. The audience wants to be told. Nobody in that sample asked the industry to stop using AI, and most say being told wouldn’t cost the brand a sale. Disclosure is available, cheap, and mostly free of downside according to the people you’d be disclosing to. The industry largely hasn’t done it, and in Europe it’s now a legal duty rather than a choice.

So the practical position is narrow and unglamorous. Use AI in production. Disclose it when the output is what the audience sees, rather than when it merely touched the workflow, a distinction Anthropic’s own watermarking makes badly. And stop treating your own comfort with AI creative as evidence of anyone else’s, because the one dataset measuring that specific belief says it’s been wrong for two years and got wronger.

That sits alongside the argument we made on Sunday, that most of what a marketing team runs through a model is routine, internal, and now runs free on a laptop. Read together the two make one point rather than two. Push AI down the stack, into the work nobody sees, where it is cheap and the only thing at stake is your own time. Keep it away from the work everybody sees, where it is expensive in a currency that does not appear on any invoice.

When IAB next runs this study, we expect the gap to be wider than 37 points, not narrower. The rate of change points that way, five points over two years. Deployment is still climbing, disclosure is still rare, and the platforms have started penalizing the output, which means the audience will encounter more AI content while being handed more reasons to notice it. If that gap comes back below 30, the industry will have started listening faster than its own behavior currently suggests, and we will say so.