AI-generated ads can look human and still underperform.

Picture a shop owner opening an AI ad tool at nine at night, typing in a few lines about the business, and getting back a finished-looking ad in under a minute: a headline, a photo, a call to action, all stitched together and ready to post. It reads fine. It looks like an ad. The question that matters isn't whether it looks like an ad. It's whether it was built to work like one.

That gap, between looking right and working, is where most of the research on AI-generated advertising now sits.

Looking authentic and performing are not the same test.

A study covered by Harvard Business Review found that AI-generated ads underperformed human-made ones even in cases where people looking at the ads couldn't tell which was which. The authors put it plainly: "looking real enough to make consumers believe they're seeing human-designed ads is not the same as working well."

That's the key finding, because it rules out the easy explanation. If an AI ad looked fake, you'd expect it to flop, and the fix would be obvious: make it look more human. But this study separates the two questions entirely. An ad can clear the authenticity bar and still fail the performance bar. Something else is going wrong, underneath the surface, that polish doesn't fix.

Looking real enough to pass is a much lower bar than being built to do the job.

A separate large report, led by Columbia Business School's Oded Netzer with researchers from Harvard, the Technical University of Munich, and Carnegie Mellon, found that AI-generated ads matched human creative overall. But within that report, data from the ad platform Taboola showed something more specific: AI ads outperformed human-made ones on click-through rate specifically when the AI ad appeared authentically human and prioritized visual trust, according to Columbia Business School's own notice on the research. The two findings aren't a contradiction, they're a layer: AI holds its own broadly, and wins outright on one measurable factor, looking authentically human. Read alongside the HBR piece, the pattern sharpens: AI tools are genuinely competitive at the surface layer, the part a viewer sees and judges in a glance. The open question is what happens underneath that surface, in the parts of an ad a person only evaluates with real attention: whether your regular customers recognize themselves in the ad, whether it's actually about what your shop does, whether it tells them to do one clear thing. No AI tool can answer those questions for you, since they live below the surface a glance can judge, and Knowing shows whether the weak spot is the first glance, the read or the click.

AI is good at variation. It isn't good at knowing which variation is right.

A 2026 study by Justin Matthews, Angelique Nairn, and Daniel Fastnedge put five professional creative practitioners through a ten-day exercise: building a fictional cereal brand using a stack of generative AI tools, including Gemini, DALL-E, Firefly, Midjourney, and Runway. The point wasn't to see if AI could make something. It was to watch, closely, what the experts actually had to do to make the output usable.

Their conclusion has a clean, quotable shape. As the authors report, generative AI tools are "powerful engines of variation but unreliable engines of meaning." Translated: the tool can hand you fifty headline options or twenty versions of a product photo in minutes. What it can't reliably tell you is which of those fifty actually says something true and specific about your business, and which ones are generic filler that happens to be grammatically correct.

The researchers identified three specific jobs the human experts kept doing, over and over, across the ten days. They had to inform the tool's direction before generating anything, anchoring it so the output didn't drift into generic or culturally wrong shorthand. They had to challenge what came back, negotiating workarounds when the output was close but not right. And they had to extend the process, running multiple tools against each other to force out better options than any single one produced alone. None of that work disappears when you add AI to the process. It just moves to a different point in the timeline.

The researchers gave that ongoing labor a name: editing, the continuous work of holding a single, coherent meaning steady across a tool that produces something slightly different every time you ask it. That's a useful word for shop owners, because it names exactly the gap between generating an ad and shipping one. A generative tool doesn't converge on one right answer. It produces a spread of plausible answers, and somebody has to pick.

The style shift is real. Whether it's AI's doing is still an open question.

It would be convenient to point at a specific stylistic tell and say: that's the AI talking, swap it out and you're fixed. Researchers looked for a sign that AI had changed the tone of ad writing: a content analysis of 240 Instagram captions from six consumer goods brands, comparing the period before generative AI tools were common to the period after. They found a real shift: more conversational tone, more emotional appeal, more narrative and engagement-driven framing in the later period. But, as the next section shows, they found the shift matched broader platform trends instead of pointing specifically to AI.

But the authors of that study were careful not to overclaim what caused it. As they put it, "these patterns reflect broader changes in digital advertising communication practices rather than direct evidence of AI-generated content," according to the study. In plain terms: advertising copy got more casual and more emotional across this period for a lot of reasons, platform trends among them, and it's hard to isolate how much of that shift, if any, came specifically from brands using AI to write captions. A tempting story turns out to be a correlation the authors themselves won't vouch for as causation.

That caveat matters for a different reason than it first appears to. It means the tell isn't in the tone of the copy. An AI-written caption and a human-written one can sound equally warm, equally conversational, equally on-trend. The problem the other two studies point to lives somewhere the eye doesn't easily catch: in whether the specific claim is true and checkable, whether the idea fits the brand that's making it, whether the ad is actually built to move someone toward a specific action. Tone is the easy thing to judge. It's also, apparently, not the thing that's breaking.

What this means for the ad you're about to post.

None of this is an argument against using AI to make ads. Five expert practitioners used a stack of generative tools and produced a full campaign in ten days; the tools clearly accelerate the parts of the job that are slow by hand, generating options, producing variations, exploring directions a person might not think to try alone. The question these studies raise isn't whether to use the tools. It's whether the first thing the tool hands back is the thing you should actually publish.

The research says, consistently, no. Not because the output looks wrong. It often looks exactly right, polished enough to pass as human-made, which is precisely what makes it risky to ship on faith. The gap between a fluent draft and a working ad is the same gap the five creative practitioners spent ten days closing by hand: checking the claim is specific, checking the idea matches the brand, checking the ask is clear. That's editing, and the research suggests it's still required, whether the first draft came from a junior copywriter or a language model.

The decision a business owner actually faces isn't AI or no AI. It's whether something, or someone, reviews the ad against what the brand actually says and what the campaign is actually supposed to do, before it goes live rather than after the results come in.