Recognition is a quality-control instrument, not a nicety
In August 2026, researchers at Tilburg University published a systematic review and three-level meta-analysis in Behaviour & Information Technology covering 19 empirical studies and 61 effect sizes on what happens to creative output when humans work with generative AI. The headline finding is modest and precise: a small but statistically significant homogenisation effect across users, robust to sensitivity checks and not explained by publication bias. Individual pieces still read well. The pool of pieces, taken together, narrows.
That is the finding to sit with, because it does not match the way most marketing teams audit their own AI use. Reviews happen one asset at a time. A caption, a landing page, a thought-leadership draft: each is judged on its own merits, usually against a brief, sometimes against a brand guideline. Nothing in that process is designed to notice that the aggregate of this quarter's output looks a little more like the aggregate of every other brand's quarter. The convergence lives above the level anyone inspects.
Where the drift actually happens
The Tilburg authors are careful. They frame the result not as a collapse of individual creativity but as a scale-dependent reorganisation of collective diversity, shaped by task structure, workflows and sociotechnical feedback. Their moderator analysis matters for marketers in particular. Homogenisation is stronger on semantically constrained ideation tasks than on open divergent-thinking tasks. A prompt like "write a LinkedIn post announcing our new report, in our brand voice, for a CMO audience" is exactly that kind of constrained task. The space of defensible answers is narrow. The model is good at finding its centre.
So is the reviewer. Faced with a draft that sits comfortably in the middle of what a sensible version of this post could say, the default verdict is approval. There is nothing visibly wrong. The disagreement a human might have raised, the odd angle a strategist was about to argue for, the sentence that would have made the piece recognisably this brand's rather than any brand's: none of these appear on the page the reviewer is holding.
Why reviews are the control point
The study measures outputs across users, not what happens inside a review meeting. That link is my argument, not the paper's finding. But the mechanism is straightforward enough to state plainly. When creative judgement is invisible in a review, the incentive structure quietly rewrites itself. A strategist who argued for a sharper opening and was met with a shrug learns that the shrug is cheaper than the argument. The next brief, they ship what the model produced, because it passes without conflict. Over a quarter, the team's output converges on whatever the model considers a safe centre, and so does every other team doing the same thing.
Recognition in that setting is not a morale lever. It is the signal that tells the team which specific judgement calls the organisation values enough to defend. "The second paragraph is the brand's" is a quality-control instrument. It names the human decision, it marks the moment the draft stopped looking like everyone else's, and it tells the next reviewer what to protect. Without it, the review process is only checking whether the asset is wrong, never whether it is distinct.
What the evidence does and does not say
Two honest caveats. The reported effect size is small; the Tilburg authors prior reporting describes it as modest, and it should not be inflated into a claim that AI-assisted teams produce interchangeable output. Most brands will still read as themselves most of the time. The risk is directional and cumulative, visible over a year of assets rather than in any single week.
The second caveat is that the underlying studies cover 2022 to early 2026 and mostly involve lab or quasi-field tasks rather than live brand marketing workflows. The meta-analysis tells us the effect is real and robust in controlled conditions. It does not tell us the exact size of the effect in a specific marketing team using a specific stack. That is a measurement job, not a reading job.
What a serious review process does differently
The implication is not to use less AI. The implication is to design the review so that human judgement becomes visible enough to be defended. Three concrete moves follow from the evidence.
- Separate two questions in every review. Is the asset competent? And what specifically in it would not have appeared if a model had drafted it alone? The second question forces the reviewer to name the human decision or admit there was not one.
- Sample output at the aggregate level, not just the asset level. Once a quarter, read fifty pieces side by side and ask whether a competent outsider could tell them apart from a competitor's fifty. The Tilburg finding predicts the uncomfortable answer; measuring it locally turns a general risk into a specific brief.
- Make recognition explicit and specific in the meeting itself. "This paragraph is yours, keep it" is a quality signal. "Great work" is not. The first changes next week's draft. The second does not.
The forward view
The brands that hold a distinct voice over the next two years will not be the ones with the best prompt libraries. They will be the ones whose review process names the human decision often enough that strategists keep making it. Everyone else will produce competent, defensible, slightly converging content, and will discover the cost in the aggregate rather than in any single asset.
The Tilburg meta-analysis does not prescribe any of this. It describes a measurable drift in the collective output of human-AI co-creation and invites the question of what, in a given workflow, counteracts it. For marketing teams, the honest answer is the part of the process no dashboard tracks: whether, when a strategist made a judgement call that the model would have smoothed away, anyone in the room said so out loud.
Source: de Rooij and Biskjaer, "Does Generative AI Make Us Think Alike? A Systematic Review and Meta-Analysis of Homogenisation in Human-AI Co-Creation," Behaviour & Information Technology, 31 August 2026.