Your legal team’s AI-assisted output is starting to all sound the same—and research explains why.
The finding that should change how legal members of boards read “our people are more productive” is this: AI can raise individual creativity scores while reducing collective idea diversity. The lawyer looks better. The set of ideas the board can choose from gets smaller.
In metallurgy terms, each grain can look finer under the microscope while the bulk material loses toughness—the ability to take a hit from an unexpected direction.
What the Studies Actually Show
MIT Sloan’s synthesis of four studies (short stories, circular-economy solutions, humour, and collaborative storytelling) is blunt: AI consistently limits the diversity of ideas across groups, even when average quality rises. You get higher mean quality and fewer breakthrough outliers (MIT Sloan).
The underlying short-story experiment by Doshi and Hauser is the one boards can hold in one sentence: access to generative AI ideas made individual stories more creative by standard ratings, and made the stories more similar to each other (Science Advances).
A 2026 Psychology Today synthesis of this literature puts the operational question in the form legal directors should steal: when everyone has the same creative partner, are we all more creative, or are we all thinking the same? (Psychology Today)
That is not a humanities puzzle. It is a description of a modern legal department.
Translate It Into Legal Work
Creativity in law is not poetry. It is the production of positions that are both novel enough to fit these facts and useful enough to survive a regulator, a counterparty, or a court. Diversity is the portfolio property: how many genuinely different theories, structures, and challenge questions exist in the set.
Here is the paradox on a legal committee:
| What you measure | What AI often does |
|---|---|
| Speed of a first memo | Up |
| Apparent quality of the average memo | Up |
| Junior lawyer’s confidence and fluency | Up |
| Distinct legal theories on the table | Down |
| Chance of an outlier defence or structure | Down |
| Board’s ability to hear a real dissent | Down |
An individual counsel who uses AI will often look more complete, more cited, and more “board-ready.” The group of counsel using the same models will produce papers that occupy a narrower band of legal possibility. The board sees a high average. It does not see the missing tail.
This is dangerous in exactly the work legal board members oversee:
- Litigation strategy, where the winning theory is frequently not the first plausible one.
- Regulatory engagement, where the industry’s modal talking point is what the supervisor has already heard.
- Transactional structuring, where “market standard” is the statistical mean—and the value is often in the non-standard allocation that fits this deal.
- Investigations, where the uncomfortable hypothesis is the one the model is least likely to volunteer.
You do not need the model to hallucinate a case. You need it only to be helpful. Helpfulness pulls toward the centre.
The Goldilocks Effect: Moderate Use Beats Heavy Use
A 2025 set of experiments in the Journal of Experimental Psychology: General finds a curvilinear (“Goldilocks”) relationship: moderate human–AI collaboration outperforms both low and high AI use for creative performance. The mechanism is knowledge diversity at the brainstorming stage—not “more prompting is always better.” Heavy collaboration compressed diversity again (PubMed; DOI).
For a legal function, “high use” is not a moral failing. It is the default under time pressure: prompt, paste, circulate. The Goldilocks finding is a control design, not a wellness slogan.
Low use: you miss cheap issue-spotting and you stay stuck on the first human idea (classic fixation).
Moderate use: you use the model to expand the set of hypotheses, then you do the lawyer’s job—rank, kill, and own a thesis.
High use: the model’s first adequate cluster becomes the committee’s universe. You have outsourced the search space.
Legal members of boards should treat “we use AI a lot” as an incomplete sentence. The missing clause is at which stage, with what cap on reliance, and with what requirement to table a non-AI position.
Why Individual Metrics Fool the Board
Boards are used to individual KPIs: cycle time, matter cost, outside-counsel spend, paper quality ratings. Those metrics will often improve with AI. The paradox is that group-level legal quality is a diversity statistic, and almost nobody reports it.
If you only measure the average memo, you will celebrate homogenization. The metallurgy analogue is measuring yield strength on a coupon that was never tested for fracture toughness. The part looks strong until it meets a load from a new angle—a novel fact pattern, a hostile counterparty, a supervisor who does not accept the industry script.
Ask for evidence of the set, not the specimen:
- How many distinct legal characterisations of the same facts were written down?
- Which one did we reject, and why?
- Was any rejected position generated without a model in the loop?
- Did outside counsel independently reach a different view, or did they start from our AI-assisted draft?
If the answer is “we all landed in the same place quickly,” that is a productivity story. It is not automatically a judgment story.
What Legal Board Members Should Do With This
You do not need to ban tools. You need to stop reading individual fluency as collective coverage.
- Demand a diversity artefact on material matters: a one-page map of alternative legal theories or structures, including at least one the model did not suggest.
- Cap “first adequate answer” behaviour in the operating model: time-boxed challenge, a named devil’s advocate, or a second firm that is instructed not to start from the internal draft.
- Watch the Goldilocks zone in policy: AI for expanding the issue list; humans for the thesis and the filing.
- Change what you praise. If you only praise speed and polish, you will get the admissions-essay problem Thomson Reuters describes: the template was already rewarded; AI just makes it cheaper (Thomson Reuters Institute).
If you want a practical way to measure idea diversity in legal reporting—not just cycle time—I work with boards and GCs on AI governance that preserves challenge culture. Contact me.
Relevant Sources
- The Hidden Cost of AI-Assisted Creativity — MIT Sloan Management Review — https://sloanreview.mit.edu/article/the-hidden-cost-of-ai-assisted-creativity/
- Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content — Doshi & Hauser, Science Advances (2024) — https://www.science.org/doi/10.1126/sciadv.adn5290
- Will AI Kill Creativity—or Set It Free? — Psychology Today (Aug 2026) — https://www.psychologytoday.com/us/blog/creative-insights/202608/will-ai-kill-creativity-or-set-it-free
- Unlocking Creativity with AI: The Goldilocks Effect of Human–AI Collaboration — Huang, Journal of Experimental Psychology: General (2025) — https://pubmed.ncbi.nlm.nih.gov/41051839/
- Organizations Are Misdiagnosing What’s Killing Their Innovation — Thomson Reuters Institute — https://www.thomsonreuters.com/en/institute/articles/feature-misdiagnosing-whats-killing-innovation