What Happens to Human Creativity in an AI World?

There is a version of this question that is easy to answer and almost useless. Will AI replace human creativity? No. Next question.

There is a harder version that almost nobody is asking, and it is the one that should worry you: what happens to a market, an industry, or an organisation when every participant is being creative with the same assistant?

The research on this arrived quietly over the past two years and it points somewhere genuinely counterintuitive. AI does not make individuals less creative. In controlled studies, it makes most people measurably more creative. What it does is make everyone creative in the same direction.

That distinction is the entire strategic story, and it has consequences for how you hire, how you run innovation, how you build a brand, and what you should be paying a premium for.

The short answer

AI raises the creative floor and lowers the creative ceiling of the collective. Individual output gets better, faster and more polished. Collective output gets more similar. A 2026 meta-analysis of 19 empirical studies found a small but statistically robust homogenisation effect from generative AI use. The implication for leaders is that in a world of converging output, difference, not quality to become the scarce resource.

The paradox: individual gain, collective loss

Start with the finding that reframes everything.

Researchers at Tilburg University published a systematic review and meta-analysis in 2026 covering 19 empirical studies and 61 effect sizes, drawn from work published between 2022 and early 2026. Their conclusion: generative AI use is associated with a small but statistically significant homogenisation effect (d ≈ 0.33) in creative output. The effect held across sensitivity analyses and was not explained by publication bias. It was strongest in tasks where participants were asked to generate concrete solutions to specific problems to precisely the kind of task that dominates corporate innovation work.

That result did not come from nowhere. It corroborates a pattern that had been building across independent studies:

• Research on AI-assisted creative writing found that stories produced with AI help were rated as individually more creative, while being significantly more similar to one another.

• A separate line of research found that human-written essays contributed roughly two to eight times more to collective semantic diversity than AI-generated essays across three studies.

• The homogenisation effect has now been observed not only in laboratory settings but in real-world creative output: essays, visual artwork, advertising content and academic publications produced before and after the widespread arrival of AI tools.

And then there is the finding that made me stop and re-read it.

The "artificial hivemind"

A team from the University of Washington, Carnegie Mellon University and the Allen Institute for AI examined whether different AI systems at least produce different outputs. If the models diverge, the homogenisation problem is manageable to just use several.

They tested 25 language models across different families and sizes, generating 50 responses each to creative prompts. Given the prompt to write a metaphor about time, the outputs collapsed into two dominant clusters: variations on "time is a river," and variations on "time is a weaver."

Twenty-five models. Different companies. Different architectures. Two ideas.

The researchers call this the "artificial hivemind," and identify two layers: intra-model repetition (a single system repeating itself) and inter-model homogeneity (different systems from different developers converging on the same output). The causes are not fully established due to overlapping training data, synthetic data contamination and similar alignment practices are all candidates, but the empirical pattern is clear.

Why this matters commercially: if you and your three closest competitors are all using AI assistance for positioning, campaign concepts, product naming and strategy, you are not just using similar tools. You are drinking from a well that has a strong central tendency. The convergence is not a rounding error. It is a structural property of how these systems work: they are trained to predict what typically comes next, which means they reliably surface the statistically common association, not the improbable one.

Improbable associations are what creativity is made of.

Is homogenisation inevitable? No, and this is the useful part

Here is where the research gets genuinely actionable rather than merely alarming.

Follow-up experimental work has shown that the homogenisation effect appears to stem from how generative AI was used, not from an inherent property of using it. In the studies that produced the strongest homogenisation, every participant was effectively assigned the same AI collaborator, prompted the same way, optimising for the same thing: a better individual answer. Nobody was optimising for diversity.

When researchers deliberately diversified the AI's contribution through distinct personas, different cultural or disciplinary perspectives, varied cognitive styles, multi-model ensembles, randomisation strategies and collaborative role architectures to the homogenisation effect was substantially mitigated.

Think about what that means. In pre-AI creative teams, we intuitively understood that a room full of identical thinkers produces identical thinking. We hired for cognitive diversity. Then we quietly handed every person in the organisation the same collaborator and expected the collective output to stay diverse.

It did not. That is not a technology failure. That is a design failure.

What this does to creative work as a profession

The labour market picture is not "creative jobs disappear." It is a bifurcation, and it is already visible.

A 2026 design-industry survey found that around 91% of surveyed designers now use AI weekly and roughly half have shipped AI-generated code. Yet a separate 2026 business survey found that while 88% of businesses used AI design tools, only 18% said it reduced their need for designers.

Those two findings only look contradictory if you assume creative work is one thing. It is not. It is splitting:

Under severe pressure, commodity creative work:

• Template and variant production

• Basic layout, resizing, localisation

• Stock and library assets

• Volume copywriting and SEO filler

• First-pass edits and proofing

Appreciating in value, judgement-led creative work:

• Creative direction and taste

• Brand strategy and positioning

• Original research and reporting

• Concept development that departs from convention

• Editorial judgement about what not to make

• Anything requiring accountability for the outcome

The pricing follows the split. Commodity creative rates are collapsing toward the cost of compute. Strategic and premium creative roles are holding or rising. The middle, competent execution without distinctive judgement, is the dangerous place to be, which is the same structural argument I make about specialists generally in The Deep Generalist Advantage.

The human-made premium

There is a countervailing force worth naming, because it is becoming a real commercial position rather than a sentimental one.

As AI-generated content saturates search results, feeds and inboxes, a segment of buyers is actively paying more for work that is verifiably human. Music licensing catalogues are creating "human-composed" premium tiers. Photographers and illustrators are being asked for provenance documentation in ways they were not three years ago. Independent writers and strategists are finding that positioning around a specific point of view, a distinct voice and documented lived experience commands higher rates than positioning around volume.

This is not nostalgia. It is scarcity economics. When the cost of producing competent output falls to near zero, the value migrates to whatever cannot be produced that way: judgement, taste, accountability, and the credibility that comes from having actually done the thing.

Interestingly, the world's largest search engine has taken a position that maps onto this precisely. Google's published guidance is that its ranking systems focus on the quality of content rather than how it was produced, rewarding demonstrated experience, expertise, authoritativeness and trustworthiness, and it explicitly encourages creators to evaluate their work in terms of who created it, how it was created and why. You can read their guidance on creating helpful, reliable, people-first content directly.

The lesson generalises well beyond search: provenance and judgement are becoming the differentiators, not production capability.

Five practices that protect creative range

If you run a team, a brand, an innovation function or a business where original thinking is the product, here is what the research actually implies you should do.

1. Diverge before you prompt.

Individual ideation first, unassisted, with a hard time box. Capture the raw, weird, half-formed ideas. Only then bring AI in as an expander and stress-tester of ideas that already exist, not as the origin of them. The homogenisation effect is strongest when AI is present at the point of origination.

This is the one practice with direct neurological evidence behind it. In an MIT Media Lab study titled Your Brain on ChatGPT: Accumulation of Cognitive Debt, researchers used EEG to compare people writing unaided, with a search engine, and with an AI assistant. Brain connectivity scaled down as external support increased, the unaided group showed the strongest and widest-ranging neural networks, the AI-assisted group the weakest coupling. The AI-assisted group also showed poorer recall of their own work and a reduced sense of ownership over it, and human graders described many of their essays as generic. The researchers additionally found consistent homogeneity in named entities, phrasing and topic selection within groups to the same convergence pattern, observed at the level of brain activity.

The finding that matters most for how you run a team is the one most coverage skipped. In a follow-up session, participants who wrote unaided first and then brought AI in, the researchers call this the Brain-to-LLM sequence to show higher neural engagement, better recall and more strategic use of AI suggestions than those who started with the tool.

The order of operations is not a stylistic preference. It appears to change what happens in the brain.

Two honest caveats: this was 54 participants, published as a preprint, on a single task type in an educational setting, and the authors themselves state the findings may not generalise across tasks. It is suggestive rather than settled. But it points in exactly the same direction as the behavioural research, which is the strongest thing you can say about any single study.

2. Deliberately diversify your AI collaborators.

Assign different personas, disciplinary lenses and cognitive styles to different team members. Use more than one model family. Randomise. If everyone in the room is prompting the same system the same way, you have functionally cloned one team member fifteen times.

3. Measure variance, not just quality.

Most creative review processes evaluate the best idea in the room. Start evaluating the spread of ideas in the room. If your concept set is tighter than it was two years ago, that is a warning sign, not an efficiency gain.

4. Treat taste as a trainable organisational asset.

Taste is the ability to recognise which of many competent options is actually the right one. It is developed through exposure, critique and consequence, not through tooling. Organisations that stop investing in it will end up with teams who can generate infinite options and cannot choose between them.

5. Protect the sources AI cannot reach.

Original research. Customer conversations. Field observation. Failed experiments. Institutional memory. Lived experience. These are the inputs that produce genuinely improbable associations, and they are precisely the inputs no model has access to. In an AI-saturated market, your proprietary experience is your only durable creative moat.

What leaders should stop doing

Three things I would end tomorrow:

• Stop measuring creative productivity by volume. Volume is now free. Measuring it rewards exactly the behaviour that destroys differentiation.

• Stop treating "faster ideation" as the goal. Speed to consensus is the enemy of originality. Most organisations were already too fast to converge before AI accelerated it.

• Stop hiring for output and start hiring for judgement. The portfolio question that matters in 2026 is not "what did you make?" but "what did you decide not to make, and why?"

The deeper point

Here is what I have come to believe after years of working on innovation with organisations across sectors, and it is the argument at the centre of The Bold Ones.

Creativity was never primarily about generation. Generation was just the visible part. The actual work has always been the harder, less photogenic stuff: noticing what everyone else has stopped noticing, tolerating the discomfort of an idea before it makes sense, and having the courage to back the option that the room finds uncomfortable.

AI is extraordinarily good at generation. It is not good at any of the rest of it, not because of a temporary technical limitation, but because those capacities require having something at stake.

So what happens to human creativity in an AI world? It gets more valuable and more concentrated. The half of creative work that was really production becomes a commodity. The half that was really judgement becomes the whole job.

The uncomfortable implication is that a lot of people whose creative identity was built on production will need to rebuild it on judgement, and organisations that do not deliberately develop that capacity will find themselves with faster teams producing increasingly interchangeable work.

The bold move here is not to reject the tools. It is to use them relentlessly for everything they are good at, while ruthlessly protecting the human capacities they cannot touch. That is the balance I explore on stage in Innovation in a World of AI, and across my writing on innovation and disruption.

Being different has never been a comfortable strategy. It has just never been this valuable.

Frequently Asked Questions

Does AI reduce human creativity?

Not at the individual level, controlled studies consistently find that people produce work rated as more creative when assisted by AI. The reduction happens collectively. A 2026 meta-analysis of 19 studies found a small but statistically significant homogenisation effect, meaning AI-assisted creative outputs become more similar to one another even as each individual output improves.

What is creative homogenisation?

Creative homogenisation is the convergence of creative outputs toward similar ideas, structures and phrasing when many people use the same generative AI systems. It occurs because these models are trained to predict statistically common associations, which reliably surfaces conventional solutions rather than improbable ones. It has been observed in creative writing, idea generation, advertising, visual art, essays and academic publishing.

Will AI replace creative jobs?

It is splitting them rather than eliminating them. Commodity creative work such as templates, variants, resizing, localisation, volume copy and stock assets is under severe pricing pressure. Judgement-led work such as creative direction, brand strategy, original research and editorial decision-making, is holding or rising in value. One 2026 survey found 88% of businesses used AI design tools but only 18% said it reduced their need for designers.

How can teams avoid producing the same ideas as everyone else?

Diverge before prompting: generate ideas unassisted first, then use AI to expand and stress-test them. Deliberately diversify AI collaborators through distinct personas, disciplinary lenses and multiple model families. Measure the variance of your idea set rather than only its best member. Research shows homogenisation is strongly mitigated when diversity is designed into the process rather than assumed.

What is the "human-made premium"?

The human-made premium is the additional value buyers assign to work when they know a person created it. It shows higher willingness to pay, stronger trust signals and explicit "human-composed" or "human-made" premium tiers in music licensing, photography, illustration and independent writing. It reflects scarcity economics: as competent production becomes nearly free, judgement, voice and provenance become the scarce inputs.

Does Google penalise AI-generated content?

Google's published position is that its ranking systems focus on the quality of content rather than how it was produced, rewarding experience, expertise, authoritativeness and trustworthiness. Its guidance encourages creators to evaluate content in terms of who made it, how it was made and why. Content produced primarily to manipulate rankings, regardless of method, is what its systems are designed to discount.

What is "cognitive debt" and should leaders worry about it?

Cognitive debt is the term MIT Media Lab researchers used for the reduced engagement, weaker recall and diminished sense of ownership observed when people rely on AI assistance for demanding cognitive work. Their EEG study found brain connectivity scaled down as external support increased, and that participants who used AI showed poorer memory of their own output. Critically, participants who worked unaided first and introduced AI afterwards showed higher engagement and better recall than those who started with the tool. The study was small and task-specific, so treat it as suggestive rather than conclusive, but the sequencing implication is practical and free to adopt.

What creative skills matter most in 2026?

Taste and judgement (choosing correctly among many competent options), original research and first-hand observation, editorial restraint (deciding what not to make), cross-disciplinary pattern recognition, and the ability to direct AI systems toward unconventional rather than conventional outputs. Production speed has largely stopped being a differentiator.

How should leaders measure creativity now that output volume is free?

Shift from volume metrics to differentiation metrics: the semantic spread of the concept set, how often the chosen direction departs from category convention, the proportion of ideas sourced from proprietary inputs such as customer research or field observation, and the commercial performance of distinctive versus conventional work.