AI Makes You More Creative, but Less Original: The Research

Does AI Make You More Creative? Yes, but Less Original

A case study of the research, the documented failures and their costs, the fixes that have actually been tested, and the rules now in force. Last updated 12 October 2026.

The short answer

Yes, AI usually makes individual work look more creative, especially for people who start out less creative. But the work of different AI users converges: ideas, stories and essays become more alike than they would be without AI. Researchers call this a social dilemma. Each person is better off; the pool of ideas narrows.

Three things most coverage leaves out:

The effect is not universal. One peer-reviewed experiment found that a base language model did not reduce diversity while a feedback-tuned model did.
Switching AI tools does not fix it. A 2026 PNAS Nexus study found 22 different chatbots converged with each other.
Some fixes have been tested and work. Structured prompting and deliberately diverse AI inputs preserved diversity in controlled experiments.

Key numbers

Finding	Figure	Evidence type
Novelty gain for writers given up to five AI story ideas	+8.1%	Peer-reviewed (Science Advances, 2024)
Usefulness gain, same group	+9%	Peer-reviewed
Novelty gain for less creative writers	+10.7%	Peer-reviewed
Brainstorm comparisons where ChatGPT narrowed idea diversity	37 of 45	Peer-reviewed (Nature Human Behaviour, 2025)
Studies pooled in the 2026 homogenization meta-analysis	19 studies, 61 effect sizes	Preprint (OSF)
Time saved on professional writing with ChatGPT	−40%	Peer-reviewed (Science, 2023)
Quality gain on those tasks	+18%	Peer-reviewed

Part 1. The evidence, graded by strength

Tier 1: Controlled experiments and a meta-analysis

Doshi & Hauser (Science Advances, 2024). UCL and Exeter researchers had 293 writers produce short stories under three conditions: no AI, one GPT-4 idea, or up to five. A separate panel of 600 evaluators, blind to the conditions, rated the stories.

With five AI ideas, stories scored 8.1% higher on novelty and 9% higher on usefulness.
Writers who scored low on a baseline creativity test gained most (+10.7% novelty, +11.5% usefulness); the most creative writers gained least.
AI-assisted stories were more similar to each other than human-only stories.

Paper via PMC · UCL summary

Meincke, Nave & Terwiesch (Nature Human Behaviour, 2025). Across five experiments, ChatGPT-assisted brainstorming produced narrower idea sets, with the diversity drop significant in 37 of 45 comparisons. In a toy-design task using a brick and a fan, 94% of ChatGPT-assisted ideas shared overlapping concepts, and nine participants independently chose the same name, “Build-a-Breeze Castle.” Penn Today reported that only 6% of the AI-assisted ideas counted as unique, against 100% in the human-only group. (I could not open the paper itself; figures come from the Wharton Mack Institute and Penn Today summaries.)

Mack Institute · Penn Today · Axios

Wenger & Kenett (PNAS Nexus, 2026, vol. 5, pgag042). Duke and Technion researchers gave 22 commercial chatbots and more than 100 human participants the same creativity tests. A single chatbot often matched or beat the average human. But the chatbots’ answers clustered far more tightly with each other than human answers did, even across companies. Turning up randomness did not help; it produced nonsense.

PsyPost report · Preprint, arXiv

de Rooij & Biskjaer (2026 preprint, OSF). A systematic review and meta-analysis of 18 articles (19 experiments, 61 effect sizes) published 2022 to early 2026 found a statistically significant homogenization effect from generative AI use. It was strongest in idea-generation tasks (specific problems with constraints) and weaker in open-ended tasks. The authors also found a small but significant effect in real-world, before-and-after comparisons, and preliminary evidence that the effect persists after people stop using the tool. Both the persistence and real-world findings rest on small groups of studies and are tentative. This is a preprint, not yet peer-reviewed as far as I could verify.

PsyPost report · Preprint

Tier 2: Smaller or narrower studies pointing the same way

Padmakumar & He (ICLR 2024). In an experiment where people wrote argumentative essays, writing with InstructGPT (a feedback-tuned model) significantly reduced diversity and increased similarity between authors. Writing with the base GPT-3 did not. The reduction was mainly because the model contributed less diverse text; the users’ own contributions were unaffected. ICLR abstract · arXiv
Anderson, Shah & Kreminski (Creativity & Cognition 2024). In a 36-participant study, ChatGPT users produced more detailed ideas but less semantically distinct ideas across users than users of a non-AI tool, and felt less responsible for their ideas. arXiv
Agarwal, Naaman & Vashistha (CHI 2025). In a 118-participant study (India and the US) using GPT-4o autocomplete, efficiency gains were larger for Americans, and AI suggestions pushed Indian participants toward Western writing styles. So homogenization is not culturally neutral. arXiv
ECIS 2024 (n = 102). AI collaboration produced more, more detailed ideas in more categories, but less original ones than working alone. AIS eLibrary

What this evidence cannot tell you

Most studies use short online tasks. Whether a novelist or design team iterating for months converges the same way is not established.
Models change quickly. The studies used GPT-3, GPT-4 and ChatGPT-era systems, and the 2026 PNAS Nexus study tested 22 commercial models, so the pattern is not confined to old models. No study I found shows it has disappeared.
“Less original” here means less varied across users, not that any one AI-assisted piece is worthless.

Why “more creative” and “less original” don’t contradict each other

This is my interpretation. A creativity rating judges one piece on its own merits. Originality in a market or field is about rarity relative to everyone else’s output. A model leans toward the most probable good idea. That raises the floor for an individual and flattens the distribution across a population. Padmakumar & He’s finding that the effect came from the model’s contributions, not the users’, fits this reading, as does the ChatGPT clustering in the brick-and-fan task.

Part 2. Fixes that have actually been tested

This is the section most competing pages lack. Evidence quality varies, so each fix is graded.

Approach	Evidence	Verdict
Structured prompting. Chain-of-thought prompts for GPT-4 produced the most diverse idea pool of the methods tested, close to human groups (Meincke, Mollick & Terwiesch, SSRN 2024)	Experiment on AI idea pools, not human-AI co-creation	Promising
Diverse AI personas. Ten distinct personas generated 300 story plots; writers given them wrote stories as diverse as the human-only baseline (Wan & Kalman, Computers in Human Behavior: Artificial Humans, May 2026)	Peer-reviewed, one task type	Promising; the authors suggest the trade-off comes from uniform deployment, not inherent limits
Use ChatGPT alongside your own ideas and ideas from other people (Terwiesch, per Axios and Penn Today)	Expert recommendation tied to the study, not separately tested	Reasonable, untested as a package
Switch to a different AI model	Contradicted: Wenger & Kenett found cross-model homogeneity	Do not rely on it
Raise the model’s randomness	Contradicted: Wenger & Kenett found it produced gibberish	Do not rely on it

Sources: SSRN, Prompting Diverse Ideas · Wan & Kalman abstract · Penn Today

Note that Wharton’s Meincke has suggested using multiple AI models for variety. The 2026 cross-model evidence suggests that, on its own, this is weak.

Part 3. Case files: eight documented AI failures and what they cost

Each case is rated by how well the facts are established. Costs appear only where a court, regulator, audit, filing or reputable report supports them. The cases cannot honestly be summed: they differ in currency, type of AI and kind of harm.

#	Case	Year	AI type	Documented outcome	Fact status
1	Moffatt v. Air Canada	2024	Chatbot	C$650.88 damages; C$812.02 with interest and fees	Ruling
2	Mata v. Avianca	2023	ChatGPT legal brief	US$5,000 sanction; case dismissed	Ruling
3	Deloitte Australia / DEWR	2025	GPT-4o in a government report	Partial refund of a ~A$440,000 contract	Reported; amount undisclosed
4	Willy’s Chocolate Experience	2024	AI-generated marketing	Cancelled day one; refunds; police called	Widely reported; cost unquantified
5	Sports Illustrated / Arena Group	2023	Alleged AI articles, fake bylines	Content removed; AdVon partnership ended; shares fell ~27% next day	Allegation, disputed
6	Coca-Cola AI holiday ad	2024 onward	AI-generated ad	Online backlash; ad tested strongly with audiences	Contested, see below
7	MD Anderson / IBM Watson	2012–2016	Clinical decision support	US$62M spent; never clinically used	Audit
8	Zillow Offers	2021	Pricing algorithm (ML)	US$304.4M inventory write-down; ~25% workforce cut	SEC filing

Moffatt v. Air Canada (2024)

The airline’s chatbot told a grieving customer he could claim a bereavement fare retroactively; the airline’s own webpage said otherwise. The British Columbia Civil Resolution Tribunal rejected the argument that the chatbot was a separate entity responsible for its own actions and found negligent misrepresentation. The sum was tiny; the principle was not: a company answers for what its bot says. (Tribunal decisions are not binding precedent outside that tribunal.)

Decision, CanLII · Law360 Canada · Dentons via Mondaq

Mata v. Avianca (2023)

Lawyers filed a brief citing court opinions that did not exist, generated by ChatGPT, and kept defending them after the court questioned them. Judge P. Kevin Castel found subjective bad faith under Rule 11 and imposed a $5,000 sanction jointly on the lawyers and their firm.

Opinion and order, 678 F. Supp. 3d 443 · Bloomberg Law

Deloitte Australia (2025)

A report for Australia’s Department of Employment and Workplace Relations, valued at about A$440,000, contained non-existent references and a fabricated court quote. A corrected version disclosed use of Azure OpenAI GPT-4o, and Deloitte agreed to repay the final instalment of its contract; the amount was not disclosed. Deloitte and the department said the findings and recommendations were unaffected.

City AM · TechRadar · AI Incident Database

Willy’s Chocolate Experience (2024): creative AI that over-promised

A Glasgow event was marketed with AI-generated imagery of an elaborate wonderland; attendees found a sparsely decorated warehouse. It was cancelled partway through its opening day, 24 February, and the organizers promised full refunds. Tickets were reported at £35. Some customers later said they were still struggling to get refunds. No verified total cost exists, and the organizers did not clearly confirm which assets were AI-made.

The Register · NBC News · STV News

Sports Illustrated / Arena Group (2023): the disputed case

In November 2023 Futurism reported that Sports Illustrated had published articles under bylines and headshots of writers who appeared not to exist, with AI-generated profiles. The Arena Group said the pieces were product reviews licensed from a third party, AdVon Commerce, which assured it that humans wrote them. Arena ended the partnership and removed the content. Reports said the fake bylines were used, in AdVon’s words, to protect author privacy, a practice Arena said it did not condone. The Globe and Mail reported Arena’s shares fell almost 27% on the following Tuesday; one-day stock moves have many causes, so I do not attribute the whole drop to this story.

Scripps News via KTNV · Mediaite · Globe and Mail via PressReader

Coca-Cola’s AI holiday ad: backlash is not the same as failure

Coca-Cola’s 2024 AI remake of its 1995 “Holidays Are Coming” spot drew loud online criticism for looking soulless. But Adweek reported the research firm System1 gave it a score of 5.9 in both the US and UK, an “overwhelmingly positive” response, before sentiment turned online. eWeek later reported that System1’s metrics suggested audiences were more forgiving in private than online comments implied, and that Coca-Cola continued using AI the following year. Verdict: a reputational controversy among creative professionals, with no verified commercial harm. This is the best reminder that online outrage and measured audience response can diverge.

Adweek · eWeek

MD Anderson and IBM Watson: US$62 million, never in clinical use

A University of Texas System audit found MD Anderson spent about $62 million on its Oncology Expert Advisor project, roughly $40 million to IBM and over $21 million to PwC, much of it without board approval. As of September 2016 the system was not in clinical use. The auditors did not rule on the system’s scientific merit, and IBM said it reached 90% accuracy on lung cancer decision support. The documented failure is largely governance, procurement and integration.

Network World · The Register · Engadget

Zillow Offers (2021)

Zillow wound down its home-buying business after a $304.4 million inventory write-down for homes bought above its estimates of future selling prices, flagged further charges, and cut roughly 25% of its workforce. Its filings cite pricing unpredictability, capacity constraints and a volatile market, so the algorithm was not the only cause. This is predictive machine learning, not generative AI.

Zillow Form 8-K (SEC) · Zillow Q3 2021 results

What the verified cases show

Across the eight cases, the pattern is that cost tracked how much was delegated without checking, not how advanced the AI was. A chatbot answering a refund question cost about C$800. A project run without competitive procurement or integration cost $62 million. Coca-Cola’s case shows the counter-example: public criticism did not translate into measurable audience rejection. This is my synthesis from eight cases I selected, so treat it as a hypothesis, not a finding.

Part 4. Was it worth it?

It depends on the task and on who checks the output.

Evidence for yes. In a preregistered experiment with 453 college-educated professionals (MIT, Science), ChatGPT cut task time by 40% and raised rated quality by 18%, and it narrowed the gap between weaker and stronger writers. That matches the story study, where the least creative writers gained most.

Caveats. One co-author noted that real-world speed gains are smaller than in the lab because users must prompt and fact-check. The Agarwal study found gains were unequal: non-Western users had to spend more effort adapting suggestions and saw less net benefit.

The cost side. Documented direct costs range from about C$800 to $62 million, and reputational costs (Sports Illustrated, Willy’s) are real but unpriced.

What no study I found establishes. A universal return on investment. Productivity experiments measure short tasks; failure cases are chosen because they went wrong. Neither gives a net result for an organization.

My reading: AI pays off for first drafts and idea expansion that a human then verifies, and especially for weaker starting performers. It costs most when unverified output reaches a customer, court, client or market, and when a whole team or industry converges on the same outputs.

Noy & Zhang, Science 2023 · MIT News · IANS report with author comments

Part 5. The rules as of 12 October 2026

European Union

The legal text is now enacted, not proposed. Regulation (EU) 2026/1744 (the Digital Omnibus on AI), dated 8 July 2026, amends the AI Act (Regulation (EU) 2024/1689). EUR-Lex records it entering into force on 27 July 2026; commentary adds it was published in the Official Journal on 24 July.

Delayed: high-risk obligations for stand-alone (Annex III) systems moved to 2 December 2027, and for product-embedded (Annex I) systems to 2 August 2028.
Not delayed: the Article 50 transparency rules apply from 2 August 2026. Chatbots must disclose that users are interacting with AI, synthetic media and deepfakes must be labelled, and AI-assisted text published on matters of public interest needs a label or documented human editorial control, according to compliance analyses.
Grace period: providers of generative systems placed on the market before 2 August 2026 must comply with Article 50(2) machine-readable marking by 2 December 2026, a transition rule stated in the amended text as quoted in consolidated timelines.
Penalties: compliance analyses put most operator obligations, including Article 50, in the tier of up to €15 million or 3% of worldwide turnover.
New prohibition: the Omnibus added bans on AI systems built to generate non-consensual sexual imagery or child sexual abuse material, applying from 2 December 2026.

Verification note: I confirmed the regulation’s title, date and entry-into-force date on EUR-Lex. The page’s full text did not load for me, so the Article 50 details above rest on consistent law-firm, compliance and EU-association summaries. Read the Official Journal text before acting on exact obligations. Earlier proposals floated different dates (for example February 2027 for watermarking), but the adopted text appears to use December 2026.

EUR-Lex, Regulation 2026/1744 · Law and Technology: OJ publication · Legalithm · Stephenson Harwood takeaways · FEDIL

United States: who owns AI-assisted creative work

Thaler v. Perlmutter. The Supreme Court denied certiorari on 2 March 2026 (No. 25-449), leaving in place the D.C. Circuit’s ruling that copyright requires human authorship. A work generated entirely by AI with no human creative contribution cannot be registered. SCOTUSblog docket · Mayer Brown
Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (29 January 2025). The Office concluded that, with currently available technology, prompts alone do not give enough human control to make a user the author. Protection is still available where a person’s own expression is perceptible in the output or where a person creatively selects, arranges or modifies AI material, assessed case by case. IPWatchdog · Squire Patton Boggs via NatLawReview

Why this matters for originality: the more a work leans on convergent AI output, the less distinctly human expression there is to protect.

Courts as de facto regulators

Moffatt (the deploying company owns its chatbot’s statements) and Mata (professionals remain responsible for what they file) work like rules in practice, though neither is a statute.

Not covered here

I did not verify US state AI laws, UK or Asia-Pacific rules, or the status of copyright suits over model training. This is information, not legal advice.

Part 6. A practical protocol, with evidence labels

Generate your own ideas first. Evidence: expert recommendation from the Wharton team; consistent with Padmakumar & He’s finding that user-written text stayed unaffected. Not tested as a standalone intervention.
If you use AI for ideas, vary the inputs deliberately. Evidence: tested. Chain-of-thought prompting and multi-persona inputs kept idea pools and stories closer to human diversity.
Don’t assume a second model solves it. Evidence: the 2026 PNAS Nexus study found cross-model convergence.
Weigh the task type. Evidence: the 2026 meta-analysis found the strongest convergence in constrained idea-generation tasks. Open-ended work converged less.
Verify every fact, quote and citation. Evidence: every fabricated-content failure above (Mata, Deloitte, Moffatt-style misstatements) was the kind of error a verification step catches.
Disclose where required, and keep a record of your human contribution if you may need to register copyright.

FAQ

Does AI make people more creative? In controlled experiments on short tasks, AI-assisted work was rated more creative, with the biggest gains for people who started lower on creativity.

Does AI make work less original? At the group level, yes. A 2026 meta-analysis of 19 experiments found a significant homogenization effect, and individual studies in stories, essays and brainstorming agree.

Is it just ChatGPT? No. A 2026 study of 22 commercial chatbots found their outputs converged with one another. One earlier study found a base GPT-3 did not reduce diversity while InstructGPT did.

Can the sameness be fixed? Partly. Structured prompting and diverse AI personas preserved diversity in experiments. Switching models or raising randomness did not help.

Can I copyright AI-generated content in the US? Purely AI-generated material cannot be registered. AI-assisted work can be protected to the extent of its human authorship.

Do the EU’s AI transparency rules apply now? Per EUR-Lex-confirmed amendments and compliance summaries, Article 50 applies from 2 August 2026, with a marking grace period for existing generative systems until 2 December 2026.

Who is liable when a chatbot gives wrong information? In Moffatt v. Air Canada, the company that deployed it. That was a Canadian tribunal decision, not binding elsewhere.

Was Coca-Cola’s AI ad a failure? Not on the evidence I found: it drew strong online criticism but scored well in System1’s audience testing.

How this article was checked

Primary or near-primary sources: court opinions, an SEC filing, EUR-Lex metadata, journal and arXiv abstracts, a Supreme Court docket, and institutional press releases.
Secondary sources are labelled where a figure rests on them (for example the Nature Human Behaviour figures, which I took from Wharton and Penn Today summaries, and the 2026 meta-analysis, which I read through PsyPost’s detailed report of an OSF preprint).
Excluded: claims I could not verify, including a newsroom AI pause mentioned in only one article and a circulating “12,000 LLM ideas” statistic that traced only to a columnist’s summary.
Interpretation is flagged. Where I draw a conclusion beyond the sources (the originality explanation, the cost pattern, the “worth it” reading), the text says so.

Limitations

Experimental findings come from short tasks and online participants.
Some regulatory details rest on summaries because the EUR-Lex full text did not load.
Cost figures are those reported in the cited filings, rulings and press; undisclosed amounts are marked as such.
The meta-analysis is a preprint, and some of its sub-findings rest on few studies.

Source index

Research

Doshi & Hauser, Science Advances 10(28), 2024. DOI 10.1126/sciadv.adn5290
Meincke, Nave & Terwiesch, Nature Human Behaviour, 2025. Mack Institute
Wenger & Kenett, “Large language models are homogeneously creative,” PNAS Nexus 5, pgag042, 2026. Preprint
de Rooij & Biskjaer, meta-analysis preprint, 2026. OSF
Padmakumar & He, ICLR 2024. arXiv
Anderson, Shah & Kreminski, C&C 2024. DOI 10.1145/3635636.3656204
Agarwal, Naaman & Vashistha, CHI 2025. DOI 10.1145/3706598.3713564
Meincke, Mollick & Terwiesch, “Prompting Diverse Ideas,” SSRN 2024. SSRN
Wan & Kalman, Computers in Human Behavior: Artificial Humans 8, 2026. Abstract
Noy & Zhang, Science, 2023. DOI 10.1126/science.adh2586

Law and filings

Moffatt v. Air Canada, 2024 BCCRT 149. CanLII
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023)
Zillow Group Form 8-K, 2 November 2021. SEC
Regulation (EU) 2026/1744. EUR-Lex; base text Regulation (EU) 2024/1689
Thaler v. Perlmutter, cert. denied 2 March 2026. SCOTUSblog
U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability, 29 January 2025