ClearFactAll topics

Does generative AI replace rather than augment human creativity?

7 peer-reviewed studies for, 12 against

Key takeaways

Stronger evidence: Against

Verdict

The evidence indicates that generative AI is, on the whole, augmenting rather than replacing human creativity at the individual and task level, while at the same time homogenizing the collective output of AI-assisted creative work — a real harm to creative diversity but not a replacement of human creative capacity. On the "against" side, Noy and Zhang (Science, 2023) randomly assigned 444 professionals to writing tasks with or without ChatGPT and found 40 percent faster completion and 18 percent higher quality, with the largest gains for lower-ability writers and human direction preserved throughout. Brynjolfsson, Li, and Raymond (Quarterly Journal of Economics, 2025) studied 5,179 customer-support agents and reported a 15 percent productivity gain concentrated in less experienced workers, consistent with AI transferring the implicit problem-solving of top performers rather than replacing creative contribution. Huang (Journal of Experimental Psychology: General, 2025) reported a curvilinear effect of AI involvement on creativity — moderate collaboration was best, and high-substitution use of AI underperformed both extremes. On the "for" side, Doshi and Hauser (Science Advances, 2024) showed GPT-4 assistance made individual stories rated more creative but reduced the diversity of the resulting corpus, narrowing the creative space at the population level. Meincke, Nave, and Terwiesch (Nature Human Behaviour, 2025) found ChatGPT-assisted brainstorming produced more and higher-quality ideas but with substantially lower semantic diversity than human-only groups. Xu et al. (PNAS, 2025) showed LLM-generated stories collapse onto a narrow set of plot skeletons, roughly an order of magnitude less diverse than human writing. The synthesis: the "for" evidence is real and important, but the harm it documents is a homogenization of collective output, not a substitution of machine creativity for human creativity. Humans remain the editors, directors, and selectors of AI-assisted creative work, and the experimental evidence shows higher creative quality when humans retain substantive control (Huang's Goldilocks effect). The "against" side therefore prevails for the question as asked — AI is augmenting, not replacing, human creativity — with the essential caveat that AI-augmented work tends to be more similar to other AI-augmented work, and that this collective-diversity loss is the most credible long-term threat to the variety that defines human creative culture.

Strongest study on each side

Ranked by citation count.

Strongest study for

Strongest study against

Experimental evidence on the productivity effects of generative artificial intelligence (Science, 2023)

Shakked Noy and Whitney Zhang · 1,327 citations

Noy and Zhang randomly assigned 444 college-educated professionals to perform realistic mid-level writing tasks with or without ChatGPT access. Workers using ChatGPT completed tasks 40 percent faster and produced output rated 18 percent higher in quality, with the largest gains accruing to the lowest-ability writers in the sample. The work was still human-directed — workers shaped prompts, edited drafts, and decided what to keep — supporting an augmentation rather than replacement model in which AI raises the floor of creative production without eliminating human authorship.

Open the interactive page →

The evidence

For (7)

Generative AI enhances individual creativity but reduces the collective diversity of novel content (Science Advances, 2024)

Anil R. Doshi and Oliver P. Hauser · 537 citations

Doshi and Hauser ran an online experiment in which writers produced short stories with or without GPT-4 idea suggestions. Stories from writers using AI were rated as more creative and better written on average, particularly for less-creative writers — but the same AI assistance made the corpus of stories markedly more similar to each other. The authors warn that the population-level outcome is a narrowing of the creative space: individuals benefit, but the diversity of novel content shrinks, which is a form of AI substitution for the heterogeneity that makes human creativity valuable.

Echoes in AI: Quantifying lack of plot diversity in LLM outputs (PNAS, 2025)

Weijia Xu, Nebojsa Jojic et al.

Xu, Jojic, Rao, Brockett, and Dolan systematically quantified plot diversity across stories generated by leading LLMs given matched prompts. Across models and sampling temperatures, LLM-generated stories collapsed onto a narrow set of recurring narrative skeletons — roughly an order of magnitude less diverse than a comparable human writing sample. If AI-generated content increasingly populates the cultural record, the result is a measurable reduction in the variety of stories humans encounter, an empirical substitution of model echoes for the heterogeneity of human imagination.

ChatGPT decreases idea diversity in brainstorming (Nature Human Behaviour, 2025)

Lennart Meincke, Gideon Nave et al.

Meincke, Nave, and Terwiesch ran experiments in which individuals brainstormed product ideas alone, with another human, or with ChatGPT. Sessions involving ChatGPT generated more ideas per minute and higher average idea quality, but the semantic diversity of ideas was substantially lower than in human-only groups. The result indicates that GPT-assisted ideation substitutes a narrow distribution of model-shaped ideas for the wider distribution humans produce on their own, displacing rather than expanding creative range.

Against (12)

Experimental evidence on the productivity effects of generative artificial intelligence (Science, 2023)

Shakked Noy and Whitney Zhang · 1,327 citations

Noy and Zhang randomly assigned 444 college-educated professionals to perform realistic mid-level writing tasks with or without ChatGPT access. Workers using ChatGPT completed tasks 40 percent faster and produced output rated 18 percent higher in quality, with the largest gains accruing to the lowest-ability writers in the sample. The work was still human-directed — workers shaped prompts, edited drafts, and decided what to keep — supporting an augmentation rather than replacement model in which AI raises the floor of creative production without eliminating human authorship.

The Ethics of Creative AI (The Language of Creative AI, 2022)

Catherine Flick and Kyle Worrall · 25 citations

Flick & Worrall present a virtue ethics framework examining ethical issues in creative AI (copyright, artistic replacement, bias, and human flourishing), arguing that responsible creative AI should enhance rather than replace human creativity and align with societal values.

Generative AI at Work (Quarterly Journal of Economics, 2025)

Erik Brynjolfsson, Danielle Li et al.

Brynjolfsson, Li, and Raymond analyzed staggered rollout of a generative-AI conversational assistant to 5,179 customer support agents and found a 15 percent average increase in issues resolved per hour, with the largest gains for less-experienced and lower-skilled workers and essentially no measurable effect on top performers. The pattern is consistent with AI codifying the tacit creative problem-solving of high performers and transferring it to less experienced workers — augmentation that broadens access to skilled creative work rather than substituting machine output for human contribution.

Unlocking Creativity With Artificial Intelligence: Field and Experimental Evidence on the Goldilocks (Curvilinear) Effect of Human–AI Collaboration (Journal of Experimental Psychology: General, 2025)

Hsuan-Che (Brad) Huang

Huang combined field data from creative professionals with controlled experiments and reported a curvilinear (inverted-U) relationship between AI involvement and creative-output quality: moderate human–AI collaboration produced the highest-rated outputs, while both no-AI and high-AI conditions underperformed. The finding directly contradicts a replacement story: maximum substitution of AI for human input degrades creativity, and the optimum is a partnership in which the human retains substantive creative control.

The debate

Lisa: The replacement question has a clear answer in the experimental literature. Noy and Zhang in Science randomized 444 professionals to writing tasks with or without ChatGPT — 40 percent faster, 18 percent higher quality, biggest gains for the weakest writers. The work was still human-directed. That is augmentation, not replacement.

Daniel: One experiment on writing tasks does not settle the cultural question. Doshi and Hauser in Science Advances ran a similar study with short stories and got the same individual effect — but they also measured corpus-level diversity, and it collapsed. Individual writers were more creative; the collective creative output was less diverse. That is replacement at the population scale even if no single writer feels replaced.

Lisa: Reduced diversity is a real concern, but it is a different claim from replacement. Brynjolfsson, Li, and Raymond in the Quarterly Journal of Economics studied over 5,000 support agents and found a 15 percent productivity gain concentrated in newer workers — AI is transferring the tacit knowledge of high performers to less experienced ones. That broadens access to creative problem-solving rather than substituting for it.

Daniel: Support agents are not the cleanest test of creativity. Meincke, Nave, and Terwiesch in Nature Human Behaviour studied brainstorming directly. ChatGPT teams generated more ideas per minute and higher-rated ideas, but the semantic diversity of those ideas was substantially lower than in human-only teams. We are getting more outputs that look more alike. That is a meaningful loss.

Lisa: Loss of diversity is bad, but it is not replacement. Humans still set the brief, choose which ideas to develop, and execute the work. Huang in the Journal of Experimental Psychology: General found a Goldilocks pattern — moderate human–AI collaboration produced the best creative outputs, and high-AI substitution conditions actually underperformed. The empirical answer to "should AI replace the human" is no — quality falls when it does.

Daniel: That is the right empirical answer at the task level. But the cultural concern is downstream. Xu et al. in PNAS quantified plot diversity in LLM stories — an order of magnitude lower than comparable human writing. If AI content increasingly populates the cultural record — newsrooms, marketing copy, screenplays, textbooks — the variety humans encounter shrinks. Replacement at the cultural layer is the version that matters.

Lisa: That is a sharp version of the concern, and the homogenization papers establish it. But it is replacement of the variety in produced outputs, not replacement of the human capacity to be creative. The remedy is to design AI workflows that preserve the human creative choice, which is exactly what the Goldilocks finding recommends. The answer to the literal question is no — AI is not replacing human creativity in the sense of eliminating it; it is augmenting individual creators while homogenizing what they collectively produce.

Daniel: I will accept that framing. The honest verdict: humans are not being replaced as creators, but their collective output is being narrowed by tools that converge on similar suggestions. The augmentation story is correct at the individual level, the substitution story is correct at the diversity level, and the policy implication is to use the Goldilocks zone — moderate AI, human-directed — rather than maximizing AI involvement.

Lisa: Agreed. The replacement framing oversimplifies; the augmentation framing is more accurate for individuals; and the homogenization concern is the part that needs design and policy attention, not panic.

Frequently asked questions

Is AI more creative than humans?

One study found large language models often score higher than humans on standard divergent-thinking tests. A large benchmark against 100,000 people found LLMs can exceed average human performance but fall short of highly creative humans.

Does using AI make creative work more similar?

Several studies suggest it does. AI-assisted stories and brainstormed ideas were rated higher on average but were less diverse than human-only work, and LLM-generated stories collapsed onto a narrow set of plot structures.

Who benefits most from AI creative and writing tools?

Studies of writers and customer-support agents found the largest gains among lower-skilled or less experienced workers. Top performers showed little measurable change in one workplace study.

How much should people rely on AI for creative tasks?

One study reported an inverted-U pattern: moderate collaboration with AI produced the highest-rated creative output, while both no AI and heavy AI use did worse. This suggests results are best when humans keep substantive creative control.

Related questions