Every AI conversation I have with C-suite leaders lands on the same two slides: productivity lift and adoption curves. Almost nobody has a slide on what AI is doing to the diversity of ideas their organization actually produces. That gap is the same one Cathy Cobey and I wrote about in The AI Precipice: innovation and safety only work when they move together — and creativity is where that balance is hardest to see, because the damage never shows up as an incident report. It shows up as sameness.
Here’s the short version. AI makes individual work better, especially for people who started weaker. At the same time, it makes a group’s work more alike. Novelty per person goes up; novelty in the room goes down. MIT calls this the hidden cost of AI-assisted creativity. I call it the trap most boards haven’t noticed they’re funding.
Five things worth five minutes of your next leadership meeting:
- AI boosts individual creativity — and quietly shrinks collective diversity.
A Science Advances study found AI-assisted work rose 8 to 11 percent in novelty, especially for weaker performers — but the resulting work became 10.7 percent more similar to everyone else’s. MIT Sloan and Wharton studies since have found the same pattern.
- It also costs cognitive muscle.
MIT Media Lab found AI-assisted writers had the weakest brain connectivity of any group tested — and stayed weaker even after they stopped using AI. Call it cognitive debt: capability quietly not built because the tool built it instead.
- Sequence is the fix — not less AI.
When AI generates the first idea, sameness follows. When humans go first, and AI evaluates or refines afterward, most of the collective cost disappears without losing the productivity gain.
- The bias runs deeper than most boards realize.
UNESCO found leading AI models default to one gender’s and one culture’s view of the world. Global research shows the same flattening effect well beyond the US and in other languages.
- Diverse teams are a built-in defense.
Research from PNAS, BCG, and McKinsey ties diverse team composition to more novel ideas and stronger financial performance. Protecting that composition protects the organization from exactly the sameness AI can introduce.
The rest of this briefing is the evidence and the playbook. If this is all you read, you already have what your board or senior leadership team needs to ask about.
The Paradox, in Plain Terms
Researchers at Wharton gave about 300 writers a short-fiction task, some with an AI idea generator, some without. AI-assisted stories were rated 8.1 percent more novel and 9 percent more useful — with the biggest gains going to below-median writers. Read no further, and this looks like exactly the democratizing tool AI has promised.
Then look across the stories instead of within any one of them: story-to-story similarity rose 10.7 percent. The same tool that lifted the weakest writers made the room’s output measurably more alike. MIT Sloan’s newer research explains why: it depends on where in the process AI enters. Used for ideation, it narrows the field everyone starts from. Used for evaluation and refinement, after humans have generated options, the collective cost mostly disappears. Eighty-three percent of executives MIT surveyed rank innovation as a top-three priority — almost none designed their AI rollout around this distinction.
Why does the model default to sameness? It’s built to predict the most statistically likely next output — so its most confident suggestion is, by construction, the most common one. Ask a thousand people for a first idea and they scatter. Ask one AI system and it doesn’t. That’s not a flaw to patch. It’s how the tool works. It also means the rare, unconventional idea — the kind a Northwestern study of 17.9 million papers found drives nearly double the impact of purely conventional work — is exactly what a mean-seeking model is least likely to produce.
CINDY’S PERSPECTIVE
Can algorithms trap human creativity? Only if an organization lets the algorithm go first every time. The fix isn’t ‘use AI less’ — it’s ‘use AI later.’ Humans lead ideation, drawing on tacit and contextual knowledge only they have; AI supports what comes next. That’s the whole argument.
Whose Default, and Why Diverse Teams Win
This isn’t confined to English or to American workplaces, and it isn’t only about ideas in the abstract — it reaches gender and culture directly. UNESCO tested the leading language models in December 2024 and found a consistent pattern: women shown in domestic roles, men in high-status careers, with further bias against LGBTQ people and certain ethnic groups. Separately, a 2026 Nature feature found AI tools measurably shifted Indian writers’ prose to sound more American, and UNESCO’s own international expert panel has called out ‘algorithmic homogenization’ as a threat to cultural pluralism, warning that AI is advancing faster than the governance built to manage it. If an AI’s ‘neutral’ default is quietly male, quietly Western, and quietly the statistical center of its training data, every team that starts there is starting from a narrower point than it realizes.
The strongest counterweight isn’t a technology fix — it’s an organizational one. Mixed-gender research teams produce more novel, higher-impact work than same-gender teams of equal size (PNAS, 2022). Companies with above-average leadership diversity earn 45 percent of revenue from innovation, against 26 percent for their peers (BCG). Top-quartile diverse executive teams are 25 to 36 percent more likely to outperform on profitability (McKinsey). The mechanism is simple: differently trained people notice different signals. An AI-mediated workflow that quietly narrows everyone’s starting point erodes that advantage — no matter how diverse the humans around the table remain.
CINDY’S PERSPECTIVE
Diversity of team composition was never a compliance line item. It’s a signal-detection system a single model standing in for many minds cannot replicate — and it’s worth protecting from the algorithm as deliberately as boards once protected it from groupthink.
Who Wins This Well, and the Playbook
Not everyone pays the same cost. HBR’s research found the deciding factor was metacognition — the habit of questioning a model’s answer instead of accepting it. Employees with that habit used AI to sharpen their thinking; employees without it anchored to whatever the model returned first. The good news: metacognition is trainable, and cheaper than almost anything else on the AI budget line.
The commercial stakes are already visible. Dentsu’s 2026 CMO survey found 79 percent worried algorithmic optimization is making brands indistinguishable from competitors, and 87 percent believe winning now requires more distinctly human qualities, not fewer. In The AI Precipice, Cathy Cobey and I laid out a six-step blueprint for balancing innovation with safety. The same logic applies here, one layer deeper: balance isn’t only about preventing catastrophic misuse of a model — it’s protecting the human judgment and originality that make an organization worth safeguarding in the first place.
CINDY’S PERSPECTIVE
Innovation and humanness aren’t opposite ends of a dial. The research says they move together when AI is sequenced correctly, and apart when it isn’t. That makes this a design choice, not a values debate.
Five moves, each drawn directly from the research above, worth putting on this quarter’s agenda:
- Sequence AI after human ideation, every time.
Make solo human ideation the default first step. Bring AI in only at evaluation, stress-testing, or refinement — the stage where it protects diversity instead of narrowing it.
- Train metacognition as deliberately as you train prompting.
Most companies invest heavily in prompt training and almost nothing in teaching people to question what a model hands back. Reverse that ratio.
- Measure diversity of output, not just speed — and track cognitive debt like any other liability.
If a team can’t see convergence happening, it can’t manage it. Ask learning leaders to report on skill practice the way finance reports technical debt.
- Protect diverse team composition, and diversify the AI itself.
Diverse teams are a demonstrated performance advantage — defend it as deliberately as you’d defend against groupthink. Rotating models, prompts, and AI personas also measurably widens the pool of first ideas.
- Ask every AI vendor what its bias testing actually covered.
Most enterprise bias audits stop at hiring and credit. Whichever model drafts strategy, marketing, or product ideation deserves the same scrutiny — UNESCO’s testing found real gender and cultural bias in exactly these everyday outputs.
Conclusion
The uncomfortable version: the AI tools making your best people’s work better today may also be making your organization’s collective thinking quietly more alike, and you’re least likely to notice it in the earliest, most human part of the creative process. The reassuring version: the fix is known. Sequence matters more than adoption speed, metacognition is trainable, and diversity of output is measurable the moment anyone decides to measure it. Boards and senior leaders don’t need to choose between innovation and humanness; however, they need to stop treating creativity as a productivity metric and start treating it as a portfolio worth protecting.
CINDY’S PERSPECTIVE
I know this is an uncomfortable topic for a board that’s spent two years being told, correctly, to move faster on AI. I’m not walking that back. Speed and sameness are not the same risk, and only one of them shows up on the dashboard everyone’s already tracking.
The organizations that lead the next decade won’t be the ones that used AI the most. They’ll be the ones that stayed curious about what their own people, thinking first and unaided, would have come up with — and built a culture disciplined enough to protect that moment on purpose. That’s not nostalgia. Once everyone has access to the same models, it’s the most practical advantage left.
Lady Whistledown’s Society Papers on the Suspiciously Similar Sonnets
Dearest Gentle Reader,
This author has lately observed a most curious sameness descending upon the ton. No fewer than eleven young ladies arrived at Wednesday’s ball in gowns of the identical dusky rose, each insisting the shade had been chosen for her alone. London’s dressmakers, it transpires, have all acquired the same marvelous engine for suggesting trims and silhouettes — and the engine, being terribly agreeable, keeps suggesting the very same dusky rose to everyone who asks.
Each gown, alone, has never looked finer. The ballroom, taken as a whole, resembled a rather expensive rehearsal for the same scene, worn eleven times. Only the debutantes who sketched a first idea themselves, engine consulted afterward for trimmings, were not standing in matching rose when the dancing began.
This author is further informed the very same engine is consulted in Paris and in Delhi, and renders, wherever asked, a suspiciously London — and suspiciously gentlemanly — opinion of good taste. One begins to wonder whose ballroom it believes itself to be dressing.
Society does love an efficient engine. It is rather less prepared for the ballroom it produces.
Until our next gathering, dear reader,
Lady Whistledown
