Look, I want to start with something that the breathless AI-and-creativity content almost never says: the research on this is genuinely mixed. Not “here are the downsides, but overall AI is great” mixed. Actually mixed. There are peer-reviewed findings that complicate the productivity story pretty seriously, and if you’re making decisions about creative workflows right now, you need to know about them.

I’ve spent twelve years watching tools promise to unlock creative potential. Most of them made certain things faster. Very few made the work better. AI is different in scale — but the pattern of overclaiming what it does for quality is identical.

So: what does the research actually say, where are the real risks, and what are senior creatives doing about it in 2026?

The Convergence Problem Nobody Admits

Here’s the finding that should be in every article about AI and creativity and mostly isn’t: AI-assisted creative groups tend to produce more ideas that are more similar to each other. More output, less variety. That’s not a knock on AI — it’s a structural property of how generative models work. They’re trained to produce outputs that are statistically likely, which by definition tilts toward what’s already been done.

“AI assistance increased productivity but ‘led to more similar final products’ across participants — a homogenization effect that individual quality gains don’t offset at the portfolio level.”

Wharton / Knowledge@Wharton, summarizing findings from Dell’Acqua et al., 2023 — source

The Dell’Acqua et al. study — which covered Boston Consulting Group consultants, not a small lab sample — found that AI assistance raised individual output quality on bounded creative tasks. But it also found the outputs converged. People using the same model, given similar briefs, landed on similar ideas. The study population matters here: BCG consultants, structured creative task, N=758 — production-environment research, not a lab analog.

That’s the part that gets buried under the productivity numbers. And for any organization that competes on creative differentiation rather than volume — a luxury brand, a design studio, an editorial team — convergence is the actual threat, not speed or cost.

Cross-source synthesis — not present in any single cited source

Put three findings together: the Dell’Acqua homogenization result, the documented tendency of large language models to favor high-probability outputs (documented in transformer architecture research), and the Wharton observation that AI’s quality benefits are most pronounced in low-performers — and you get a pattern that none of the three sources names directly.

AI-assisted creative work raises the floor and compresses the ceiling. It’s a powerful tool for consistency. It’s a structural hazard for distinctiveness. Organizations optimizing for scale will see the gains. Organizations optimizing for voice will feel the loss — and probably won’t attribute it to the tool.

What the Wharton Research Actually Says

The Knowledge@Wharton piece on AI and creativity does something rare: it sits with the tension instead of resolving it prematurely. The summary finding — AI boosts individual creative output, especially for people who start from a weaker baseline — is accurate. But it comes with a condition that almost every secondary source drops.

The researchers specifically noted that the quality gains were most significant for lower-performing participants. High-performers gained less, and in some task types, their work became less distinguishable from average. That’s a finding about creative leveling, not creative amplification. Different thing.

For individual practitioners reading this: if you’re already good, AI gives you speed. It doesn’t necessarily give you better ideas. It might give you faster access to the same range of ideas you’d reach anyway, with less friction. Whether that’s a gain depends entirely on where friction was actually helping you — forcing associative leaps, generating surprise through constraint, whatever your process relies on.

Second-order mechanism

The convergence problem is hard to detect because the outputs don’t look homogenized. They look polished. AI-generated creative options tend to be well-structured, plausible, and internally consistent — which is exactly what “good” looks like in most review processes.

The signal that something’s converging is subtle: your work starts to feel slightly less surprising to you. Your references start rhyming with each other. A creative director who’s worked with ten AI-assisted teams can feel it; a client reviewing a single deliverable probably can’t. That’s why the problem persists even when people know about it.

Harvard Business School research — specifically work on collective intelligence and AI — adds another layer. Sytch & Khanna, HBS Working Knowledge, 2023 — directional, observe with that scope in mind The finding that concerned them wasn’t individual output. It was what happens to the cognitive diversity of a group when everyone’s drawing from the same generative well. Groups that use AI together tend to surface fewer genuinely different perspectives. The tool that helps individuals contribute more makes groups think more similarly.

“The real risk isn’t that AI replaces creatives. It’s that it quietly makes creative teams more homogeneous while making each individual within them feel more productive.”

Editorial synthesis — sources: Dell’Acqua et al. (2023), Knowledge@Wharton (2023), HBS Working Knowledge (2023)

The Second-Order Risk: You Won’t Notice When It Happens

This is the part that keeps me up at night, honestly. And I’ve been doing this long enough to have a pretty high threshold for “things that keep me up.”

Most creative risks are visible. You can see a weak concept. You can feel a derivative campaign. You can compare this year’s output to last year’s and notice the drop. Convergence doesn’t work like that. The individual pieces look good. The process felt productive. The client approved it. The problem only shows up in aggregate — when you look at a portfolio of work across twelve months and realize everything in it could have come from the same source.

I’ve watched this happen to teams that were genuinely talented. Fast. Hardworking. They adopted AI tools, their throughput tripled, and about eight months later their creative director described the work as “competent and hollow.” Those were his words. Not mine.

Finding Source Population / Scope Evidence Level ⚠ Adversarial Column
AI assistance raises individual creative output quality, especially for lower performers Dell’Acqua et al. via Wharton, 2023 N=758 BCG consultants, structured creative tasks Strong Task type was bounded and evaluable — may not generalize to open-ended brand or cultural work
AI-assisted groups produce more homogeneous outputs across participants Dell’Acqua et al. via Wharton, 2023 Same N=758 sample Strong Study measured similarity in outputs using predefined rubric — different similarity measures might yield different results
Group AI use reduces cognitive diversity of surfaced perspectives HBS Working Knowledge, 2023 Organizational case studies, directional Moderate No randomized control group; directional finding from observational research only
Content demand rising 5–20× as AI lowers production cost floors Adobe Creative Trends Report, 2026 Adobe customer base; self-reported vendor-published, no independent audit Directional Self-reported by a vendor whose business benefits from AI adoption appearing large. Treat as directional, not verified.
Sources as linked. Evidence levels: Strong = consistent findings from peer-reviewed research with defined populations. Moderate = solid but observational or limited sample. Directional = plausible but not independently audited or statistically powered.

There’s a complicating finding I’d be doing you a disservice to skip: the same Dell’Acqua study found that participants who were told not to use AI — the control group — produced notably more varied and surprising outputs in the high-performer tier. Not always better by the rubric. But more distinctive. More risky. More likely to contain the one unexpected idea that defines a campaign. So the tradeoff is real. Volume and consistency versus variance and surprise. You need to know which one you’re actually optimizing for before you build workflows around AI.

How Senior Creatives Are Actually Integrating AI

The most useful thing I can tell you here isn’t a framework. It’s a pattern I’ve seen across teams that are navigating this well — and one I’ve verified in enough different contexts to feel reasonably confident about it, though I want to be clear it’s practitioner observation, not a controlled study.

The teams that preserve creative distinctiveness while gaining AI’s efficiency benefits tend to do one specific thing: they use AI for divergence after they’ve done human convergence, not before. Meaning: they spend time in a room together — no AI, no tools, just people — getting confused and unproductive and eventually surprising each other. Then they take the one or two ideas that emerged from that mess and ask AI to produce forty variations, fill in executional gaps, generate the adjacent options they don’t have bandwidth to explore manually.

That sequence matters. It’s different from the default, which is using AI to generate options first and then humans to select and refine. The default hands the generative work — the part where unexpected ideas appear — to the model, and keeps humans in the evaluative role. Humans are great evaluators. But evaluation is not the same as ideation. And consistently great evaluation of statistically-likely AI options produces consistently good, consistently expected work.

Cross-source synthesis — not present in any single cited source

The convergence problem, the HBS group-diversity finding, and the documented way LLMs produce high-probability outputs all point to the same structural conclusion: the point in a creative process where AI is most useful is execution, not conception.

This isn’t about AI being “less creative” than humans in some mystical sense. It’s about what training on human-produced work at scale optimizes for: the center of what’s been done, not the edges. Conception — the genuinely new idea — happens at the edges. So does failure. AI doesn’t fail in interesting ways. That limitation is information.

There’s also a regional dimension worth naming, briefly. Teams in tightly regulated markets — Australia in particular, parts of the EU — are moving more slowly with AI adoption, and not purely because of constraint. Some of that deliberateness is producing more thoughtful integration. Slower adoption of a tool that can homogenize creative output isn’t necessarily a competitive disadvantage. Might be the opposite.

For the Two Audiences Reading This

For: Content Strategists & Creative Directors

The workflow question is the wrong question

The conversation in most strategy circles is “how do we integrate AI into our workflows?” That’s not actually the interesting question. The interesting question is: “which parts of our creative process produce our most distinctive work, and what happens to those parts when we introduce AI?” Most teams don’t know the answer to that second question. They should find out before building workflows around a tool that demonstrably compresses variance.

Specific action: run a simple audit. Take your last twelve months of creative output. Ask: “could a single AI model, given reasonably good prompts, have produced this?” If the answer is yes for more than 60% of it, you either have an AI adoption problem (you’re not using it efficiently) or a differentiation problem (your creative edge was never as distinctive as you thought). Either way, that information is more valuable than any framework.

Stop doing this Don’t treat AI-generated ideation as equivalent to human ideation and then wonder why your portfolio feels consistent but flat. Consistent and flat is exactly what you’d expect from optimizing high-probability outputs. The process produced the result it was designed to produce.
For: Individual Creative Practitioners

The floor is rising. So is the ceiling — but for different people.

If you’re in the middle of the skill distribution for your creative field, AI is genuinely great news. The gap between what you can produce now and what you could produce with good AI assistance is real and meaningful. The research supports this. You will produce more, and the work will be better by most rubrics.

If you’re at the high end of your field and your competitive advantage is distinctiveness — voice, vision, the specific quality of your sensibility — the dynamic is different. AI helps with execution and speed. It doesn’t help with being surprising. And it might, over time, pressure you to adopt tools and workflows that quietly make your work more like everyone else’s. The premium on genuine distinctiveness is actually rising right now, precisely because the floor is rising everywhere. Being 20% better than average is worth less when average means AI-assisted and reasonably good. Being genuinely singular is worth more.

Stop doing this Don’t use AI to validate your ideas before you’ve had time to sit with them. The moment when an idea feels slightly wrong or risky or unfinished is sometimes the moment it’s most worth developing. AI will tell you how to make it safer. That’s not always what you need.

One last thing, and then I’m done. The “AI replaces human creativity” frame is the wrong frame — and not because AI can’t do creative work. It can. The more accurate frame is: AI changes what creativity is valued for. The parts of creative work that were previously difficult because they were time-consuming or technically demanding are becoming cheap. The parts that are difficult because they require a genuine point of view, earned over years of living and failing and paying attention — those are becoming expensive. If you’re building skills right now, that’s where to build them.