The evidence says individuals get better with generative AI and the set of their work gets narrower. The loss sits in the configuration, not in the person, and only the configuration can fix it.
In brief
- Across the controlled studies, people working with generative AI produce more ideas, rated higher, faster. The same studies find their ideas more alike. A 2026 meta-analysis of 19 studies confirms the pattern.
- The narrowing travels through one conduit: a shared model returns its central tendency to each worker as finished-looking text, and every review inside the firm judges the individual, not the set. The first to see the set is the client.
- Half of senior executives surveyed by BCG in 2026 already see the erosion; one firm in ten has a strategy. The levers are organizational: measure variety where it is lost, place the tool by task, keep more than one conduit, keep authorship with the person, and protect unassisted practice.
The worry is real, and it is pointed at the wrong place
A widely held view in boardrooms is that generative AI is eroding human creativity, and that the defense is to use it less. The view captures something real: people who lean on the tool report thinking less, and the studies that measure engagement find it. But it locates the loss in the wrong place. The data shows that individuals with AI produce more creative work, rated higher. What they lose is their difference from one another. I would argue that the creativity problem of the AI-assisted workplace is a problem of configuration, and that it is answered by how work is organized rather than by how sparingly the tool is used.
What the evidence shows
Individuals gain. In a field experiment with 758 BCG consultants, those with GPT-4 produced work rated more than 40% higher in quality on tasks inside the tool’s range, and the bottom half of performers gained most (Dell’Acqua et al., 2023). Among 293 writers, AI story ideas raised judged novelty by 8.1%, and the least creative writers gained 10.7%, enough to put them level with the most creative (Doshi and Hauser, 2024).
The set narrows. In the same BCG study, the ideas of consultants using AI varied markedly less than those of consultants without it, and 100 simulated AI-only sessions were the most homogeneous group of all. The AI-assisted stories were more alike and about 5% closer to the AI’s suggestion. A comparative study found that ChatGPT users produced more ideas across more categories than users of a non-AI prompt deck, yet at the group level their ideas were less distinct, with no difference at the individual level: the tool was suggesting similar things to different people (Anderson, Shah and Kreminski, 2024). The 2026 meta-analysis by de Rooij and Biskjaer pools 19 studies and finds a small but consistent homogenization effect (d = .334), strongest in constrained ideation tasks, with signs that it persists after the tool is put away. Their own words: subtle in isolated instances, but consequential through widespread engagement with similar AI systems.
Judgment thins where confidence in the tool is high. Among 319 knowledge workers describing 936 real tasks, higher confidence in the AI predicted less critical thinking; higher confidence in oneself predicted more. The barriers named were practical: no time under a quota, verification treated as someone else’s job, too little domain knowledge to judge the output (Lee et al., 2025).
Executives are starting to see it. In BCG’s June 2026 survey of 70 C-suite leaders, half already observe de-skilling, almost 90% cite overreliance on AI outputs without challenge, and 49% cite less diversity of thinking. One company in ten has a response (Goel, Martin and Kaffe, 2026).
What improves for the individual, what narrows for the organization
Every study that measures both finds the gain and the narrowing together.
| Study | Setting | What improved for the individual | What narrowed for the group |
|---|---|---|---|
| Dell’Acqua et al., 2023 | 758 BCG consultants, field experiment | Quality rated +40%; 12.2% more tasks; 25.1% faster | Ideas measurably less varied; modal worker retained ~87% of the AI’s text |
| Doshi and Hauser, 2024 | 293 writers, 600 evaluators | Novelty +8.1%; least creative writers +10.7% | Stories more alike; ~5% closer to the AI’s suggestion |
| Anderson, Shah and Kreminski, 2024 | ChatGPT vs a non-AI prompt deck | ~15% more ideas across 27% more categories | Less distinct at group level; no change at individual level; self-credit 48% vs 64% |
| Lee et al., 2025 | 319 knowledge workers, 936 tasks | Less effort on recall, synthesis, evaluation | Confidence in the AI predicts less critical thinking |
| de Rooij and Biskjaer, 2026 | Meta-analysis, 19 studies, 61 effects | Individual creativity improves (as the included studies show) | Homogenization d = .334; strongest in constrained ideation; may persist |
| Goel, Martin and Kaffe (BCG), 2026 | 70 C-suite leaders surveyed | Productivity, as the reason for adoption | 49% less diversity of thinking; 43% fewer debates; 53% slower junior development |
Sources: as listed below. Table: Mentor Cana.
How the narrowing travels
The chain from individual gain to collective loss has six links, each visible above.
- One model becomes the shared source for many people inside the firm and across its competitors.
- Each person receives something close to the model’s central tendency for that prompt; prompting style made no measurable difference.
- The output arrives finished-looking, so the distance from suggestion to submission is short and cheap to close. The modal BCG consultant kept about 87% of the AI’s text.
- Confidence in the tool lowers the perceived effort of checking, and time pressure rewards accepting the saving.
- Each person’s work improves; the set converges.
- Reviews and dashboards judge the individual, not the set, so the convergence stays invisible until a client sees it.
In two decades of research and advisory roles, the product was a distinct point of view. A set of notes that all say the same better thing is, to the reader, one note. That is how a client experiences the loss, and why the first to notice is usually outside the firm.
A fair rival explanation is that teams converged long before AI, through templates, benchmark decks and group discussion. Partly true. The tool did not invent homogenization; it industrialized it, by routing more people through one channel at lower cost and higher polish.
A different reading: who shapes whom along the conduit
Read another way, as a question of who shapes whom and through what, the same chain shows where the levers are.
The model vendor shapes the tool through training and default settings, one way. The tool shapes the worker through finished-looking output; the worker shapes the tool back only through prompts, which made no difference to how alike the group’s ideas were. That is the strongest and most lopsided connection in the picture. Worker and review process shape each other in roughly equal measure: work earns a score, and the score, the quota and the time budget decide whether checking feels worth doing. The executive team shapes the review process strongly through what it measures, and the vendor hardly at all. The firm delivers its set of work to the client, and nothing on record comes back about variety.
The tool shapes the worker; the firm’s review shapes nothing back
Influence runs from the vendor through the tool to each worker and on to the client; the firm’s own review looks only at individuals and sends nothing back up the chain. Line weight shows how much one party changes the other; line style shows whether the connection is on record, reported, or my reading.
Source: Mentor Cana analysis of the studies in Exhibit 1.
Two things follow. The conduit that carries the narrowing is controlled outside the firm; the firm can only decide what sits around it. And the one loop the firm fully owns, from executive team to review process to worker, measures the one thing that is not being lost. Wan and Kalman’s 2026 experiment shows what changing the deployment does: writers given plots from ten deliberately different AI personas, instead of one default voice, wrote stories as varied as a human-only group’s. The trade-off, they conclude, comes from uniform deployment, not from the technology.
What to do about it
Each move follows from the evidence above; where it is my reading rather than a study, I say so.
Executive team
- Measure variety where it is lost. Add a portfolio-level measure of idea diversity beside the individual quality score: semantic similarity across submissions where you have the tooling, or one reviewer reading the whole set where you do not. Report both to the people who decide. The firms in BCG’s survey see the erosion and have no instrument for it.
- Put judgment in the job, not on top of it. Time pressure and “verification is someone else’s job” are the barriers workers name most (Lee et al.). Budget verification and stewardship as part of the role, assess how outcomes were reached and not only what they scored, as BCG recommends, and treat near-verbatim retention of AI text as a review flag rather than a productivity win.
- Protect unassisted practice, starting with juniors. The 2023 field-experiment authors warned that firms which stop delegating in-range work to junior staff create training deficits; three years later 53% of executives report exactly that. Schedule AI-off work on real tasks where the stakes are low enough to learn, and keep senior-junior apprenticeship as a structured practice.
Transformation leaders
- Place the tool by task. For constrained ideation (name the ten options, propose the campaign, segment the market), where the meta-analysis finds the strongest effect, run an independent human-only round first and bring the tool in afterward to expand, stress-test and refine. For drafting and polishing, use it freely. One caution, and it is mine: same-session ordering made no difference in the comparative study, so treat the sequence as a hypothesis to test, with real separation between rounds.
- Keep more than one conduit. Wan and Kalman show that deliberately varied AI voices preserve diversity; the 2023 field-experiment authors suggested several models or more human-only involvement. Build the plurality in: multiple models, persona-varied prompts set by design rather than left to each user, non-AI creativity aids, outsiders from other functions. No single system should be the passage point for every idea.
- Keep authorship with the person. Require the submitter to state what was kept, changed and rejected, and why. Ownership drops with AI in every study that measured it (48% versus 64% self-credit), and ownership is where judgment gets exercised.
What could change
If buyers begin to discount AI-typical output, as the 2023 field-experiment authors anticipated, variety becomes priced and the firm that preserved it holds an asset competitors cannot buy back quickly. If vendors ship diversity-optimized generation, part of the burden shifts to the tool. Until either happens, it sits with how work is organized.
The significance of generative AI for workplace creativity lies less in what the tool can do than in the configuration through which it is used: one conduit, finished-looking output, incentives for speed, reviews of individuals rather than of the set. Change that configuration and the gains remain while the narrowing is contained. The relationship most worth managing is not the one between a person and the tool. It is the one among all the people who share it.
Where in your workflow does the center start to pull, and would you be able to see it?
Sources
- Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013, September 22, 2023. https://ssrn.com/abstract=4573321 (foundational field experiment; the only source for the consultant figures)
- Doshi, A. R., and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10(28), eadn5290, July 12, 2024. https://doi.org/10.1126/sciadv.adn5290 (foundational experiment the 2026 meta-analysis and Wan and Kalman build on)
- Anderson, B. R., Shah, J. H., and Kreminski, M. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation. Proc. Creativity and Cognition (C&C ’24), June 2024. https://doi.org/10.1145/3635636.3656204 (foundational; the only study separating group-level from individual-level effects)
- Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., and Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proc. CHI ’25, April 2025. https://doi.org/10.1145/3706598.3713778
- de Rooij, A., and Biskjaer, M. M. (2026). Does generative AI make us think alike? A systematic review and meta-analysis of homogenization effects in human-AI co-creation. Behaviour & Information Technology, e-pub August 31, 2026 (preprint April 2026). https://research.tilburguniversity.edu/en/publications/does-generative-ai-make-us-think-alike-a-systematic-review-and-me-2/ ; companion with the effect size: Generative AI Makes Creative Output More Homogeneous, Proc. ECCE 2026, September 14, 2026. https://doi.org/10.1145/3822301.3822304
- Wan, Y., and Kalman, Y. M. (2026). Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation. Computers in Human Behavior: Artificial Humans, Vol. 8, May 2026. https://cris.openu.ac.il/en/publications/diverse-ai-personas-can-mitigate-the-homogenization-effect-in-hum/ ; preprint arXiv:2504.13868, March 2025. https://arxiv.org/abs/2504.13868
- Goel, S., Martin, D., and Kaffe, C. (2026). When Everyone Uses AI, Companies Risk Losing Critical Skills. Boston Consulting Group, June 2026. https://www.bcg.com/publications/2026/when-everyone-uses-ai-companies-risk-critical-skills
Leave a Reply