The competitive worry I hear most often now isn't that AI will produce bad work. It's that it will produce acceptable work for everyone at once, and that a category full of acceptable work is a category where nothing stands out. Marketers have a phrase for the destination, which is mediocrity at scale. What's been missing is evidence about whether it's actually happening.
A preprint posted at the end of July is the most useful thing I've read on it. Dong and Yakura ran a preregistered creative-metaphor experiment with first- and second-language English speakers across three conditions: no AI, AI generating the ideas, and AI refining ideas the participant had already produced. The finding worth carrying around is that AI ideation compressed the diversity of the group's output while raising how good any individual's work was rated. AI refinement preserved the diversity. Same tool, two ways of using it, opposite effects on variety.
It's a preprint and it's a metaphor task rather than a campaign, so I'd hold it loosely. But the distinction it draws is the one that matters commercially. If you ask the model what to say, you converge with everyone who asked the same question. If you decide what to say and ask the model to sharpen it, you don't. That's a workflow difference, not a tooling difference, and it's invisible on a budget line.
The practitioner data points the same way, with the usual caveats about who paid for it. WARC and TikTok surveyed 400 marketers involved in creative production this May. Eighty-eight percent report higher creative volume since adopting generative AI. Forty-five percent report a significant improvement in quality. Forty percent name over-reliance on generic styles as their main limitation. TikTok commissioned the study and has an interest in the conclusion, so treat it as a description of how marketers feel rather than a measurement of what's true.
DoubleVerify's global study in late July, across 22,000 consumers, found that 56 percent can't consistently identify AI-generated content, while 42 percent say low-quality AI advertising damages their perception of a brand and 40 percent respond positively to polished AI ads. DoubleVerify sells verification, so again, directional. The pattern across both is that consumers aren't reacting to AI as such. They're reacting to work that's obviously cheap, which they've always done.
That's the part I'd hold onto. The threat isn't the technology, it's the temptation. Volume went up 88 percent and quality went up for fewer than half. What that describes is an industry that used a productivity gain to make more things rather than better ones, which is a choice, and a familiar one. The same thing happened with desktop publishing and with programmatic.
Where it gets genuinely hard is that distinctiveness has always been expensive and now the cheap alternative is much better than it used to be. A generic ad in 1998 looked generic. A generic ad now looks fine. When the floor rises, the gap between adequate and distinctive narrows visually while mattering more commercially, and that's an argument no marketer wins easily in a budget meeting.
The defensible position, I think, is that the input has to be something the model doesn't have. A model trained on the public internet knows what your category says. It doesn't know what your customers told you on the phone last week, what your service team hears at the same point in every onboarding, or what your founder believes that the category disagrees with. Those are the raw materials that don't converge, and they're the ones most organizations are worst at collecting.
So the practical version is unexciting. Use the tools on the back half of the process, where the study says the diversity survives. Spend the time you save on getting closer to something the model can't already read. That's not a strategy anyone will call visionary. It's the one the evidence currently supports.
Sources
- Dong & Yakura, preregistered study on AI ideation, refinement and creative diversity (arXiv preprint, July 2026): arxiv.org
- The New Creative Advantage (WARC with TikTok, reported July 2026): ppc.land
- Poor quality AI content puts brand trust at risk (DoubleVerify, July 2026): globenewswire.com