Harvard Research: A 22% Gender Gap In AI Use That Is Barely Budging

Men are using generative AI at a markedly higher rate than women, and the latest research out of Harvard Business School suggests the gap is not closing on its own.

In Global Evidence on Gender Gaps and Generative AI Over Time, Katelyn Cranney, Solène Delecourt, and Rembrand Koning synthesize 76 sources across more than 100 countries and find a divide that is, in their words, nearly universal. Among the roughly 319,000 respondents in studies reporting both men and women, adoption runs 47.8% for men against 39.3% for women – a relative gap of 22%. The gap has narrowed over time but stalled at around 16% since early 2025, and web-traffic data on the ten most-visited AI tools shows women spending less time in the tools, with the divide widening for frontier products.

The authors argue the gap shrinks as familiarity grows, but that institutional, organizational, and social frictions will keep it from closing on its own. For any business school trying to graduate an AI-fluent class, the finding lands close to home: the people who arrive already comfortable with these tools are not arriving in equal numbers.

THE BIGGER PATTERN

The gender-gap study is one of many pieces of new research out of HBS, and it sits inside a larger preoccupation running through the faculty’s latest work – the steady migration of AI out of the IT department and into territory that used to belong to clinicians, compensation committees, and auditors. The faculty are no longer asking whether AI matters. They are measuring what it does to the people and institutions it touches.

Nowhere is that clearer than in mental health, where two studies cut in opposite directions. In Benchmarking the Safety of General Purpose Large Language Models for Suicide Risk Detection and Response, Julian De Freitas and a large clinical team run ten leading models through a simulated multi-turn evaluation. Newer models score better than their predecessors at three of four providers, with Claude Sonnet 4.6 and GPT-5.2 topping the table and GPT-4o and Grok 4 at the bottom. The catch is sobering: models reliably flag potential risk but fail to ask direct follow-up questions about suicidal thoughts 61% of the time and fail to steer users toward human care 33% of the time.

Set against that caution is a more optimistic result. In a forthcoming NEJM AI paper, AI for Proactive Mental Health, Julie Cachia, De Freitas, and colleagues report the first multi-institutional randomized controlled trial of a generative-AI well-being app called Flourish. Across 486 undergraduates at three U.S. institutions, students prompted to use the app twice a week report greater positive affect, resilience, and social well-being, and are buffered against declines in mindfulness. The pairing is the story: the same technology that stumbles in acute crisis may still deliver real, measurable benefit as everyday prevention.

AI ARRIVES ON THE BALANCE SHEET

The arms race also shows up in the accounting. Suraj Srinivasan’s case Meta Platforms: Accounting for the AI Arms Race drops students into early 2026, when Meta discloses a $27 billion off-balance-sheet data-center project structured, the case suggests, to flatter the balance sheet, alongside depreciation choices that shave billions off expense. Meta’s auditor flags the treatment as a critical audit matter, four U.S. senators call for an investigation, and the private-credit partner holding 80% of the off-balance-sheet entity faces a liquidity crisis. Taught from a credit analyst’s perspective, it is a timely lesson in how AI’s capital intensity is reshaping financial reporting.

Alex Chan pushes the legal frontier in Optimal Medical Liability for AI, asking who is at fault when AI acts as the doctor rather than a tool the doctor uses. His answer hinges on the medical record itself: when it cleanly separates AI error from patient nonadherence, standard liability works; when it is coarse, the first-best may be unreachable.

PAY, FAILURE & THE LIMITS OF PREDICTION

Not everything is about AI. In Making Equity Incentives Actionable, Sam Karasik, Ethan Rouen, and Ashley Whillans run a field experiment at a building-materials supplier and find that shareholder letters framing stock rights around concrete behaviors, rather than abstract values, cut turnover by 13.2 percentage points. The retention edge is sharpest during operational disruption – in this case, Hurricane Helene – when behavioral clarity matters most.

Sen Chai and Anil Doshi, in The Promise–Risk Balance, dissect Virgin Galactic’s 2014 test-flight crash to show how a firm recalibrates engineering choices and public messaging in tandem after catastrophic failure, and why de-risking the technology can quietly dampen the very promise that sustains stakeholders.

And for a closing note that ties back to the beat, Lauren Cohen’s Working Knowledge piece, If AI Knows Your Next Trade, What Happens to Money Managers? warns that fund managers running predictable strategies are the ones AI can most easily mimic. The managers who survive, he argues, are the ones doing something the machine cannot yet copy.

DON’T MISS NEW HARVARD BUSINESS SCHOOL RESEARCH TAKES AIM AT SOME OF AI’S BIGGEST ASSUMPTIONS

© Copyright 2026 Poets & Quants. All rights reserved. This article may not be republished, rewritten or otherwise distributed without written permission. To reprint or license this article or any content from Poets & Quants, please submit your request HERE.