TL;DR

Three studies this year say the same thing from three directions: women worry they’ll be seen as cheating when they use AI at work, that worry is justified, and the rational response—using AI quietly or not at all—is exactly what will cost them.

Lean In found that women are 32% more likely than men to fear the “cheating” label. A UK study gave 1,000 people the same AI-assisted résumé and changed only the first name: “Emily” was 22% more likely to have her trustworthiness questioned and twice as likely to have her competence questioned.

Harvard, Berkeley, and Stanford researchers found women adopt AI at roughly 25% lower rates, and the gap doesn’t close with access or training.

In a profession that is two-thirds women, this isn’t a diversity footnote. It’s the mechanism by which the job loss Alex Karp predicted actually happens.

Key Insights

Lean In’s March 2026 survey found that 29% of women worry they’ll be perceived as cheating when they use AI at work, compared with 22% of men.
Zehra Chatoo’s résumé study gave 1,000 UK adults the same AI-assisted CV with the same disclosure. The only variable was the name. Emily’s trustworthiness was questioned 22% more often than James’s, her competence twice as often, and James scored a 97% approval rating.
Harvard Business School, Berkeley, and Stanford synthesized 18 studies involving 140,000 people and found that women adopt generative AI at about a 25% lower rate than men—across every country and education level, in identical jobs, and even after equal access and training.
Women anticipate larger penalties for using AI and are more likely to view it as cheating. The fear isn’t irrational. It’s accurate.
Public relations is 65.6% women. The people adopting AI fastest on comms teams are the ones most penalized for saying so, while the people setting AI policy are the ones least penalized. Adoption isn’t low. Admission is low.
The fix isn’t for women to use AI less. It’s for leaders to make disclosure a team norm rather than a confession, evaluate output rather than process, and watch who gets asked, “Did you really do this?”

When Men Use AI, They’re Resourceful. When Women Use AI, They’re Cheating.

A few months ago, I wrote about Alex Karp telling the Davos crowd that AI would reduce the economic and political power of highly educated, often female professionals.

My argument then was: believe him. And build demonstrable strategic value so fast that nobody can pretend you’re replaceable.

I still believe that. But I missed something.

I assumed the threat was the technology. That AI would get good enough at the text-heavy, judgment-heavy work women dominate, and the jobs would just … go.

It turns out there’s a faster route to the same outcome, and it doesn’t require AI to replace anyone.

It just requires women to hide their AI fluency long enough for it to stop counting as a skill.

That’s what Tara McDonagh and I discussed on her podcast this week. She’s been reading the same research as me, and we’re both fired up about it.

She has also published her thoughts on LinkedIn, which has been shadowbanned by the social network.

Not for being wrong. Not for being political. Not for anything crude or inappropriate.

For being “controversial.” She used the words “women” and “gender bias” in a post about AI, and the algorithm decided that was a conversation nobody needed to see.

It’s exhausting. For every step we take forward, it feels like we take 10 times as many backward.

The Research, in the Order It Should Scare You

There are three studies Tara and I discussed, and I’m going to talk you through them in the order I think is most important. The first is about fear. The second proves the fear is justified. The third shows what happens when smart people respond rationally to a justified fear.

Spoiler alert: they retreat.

Women Worry They’ll Be Seen as Cheating

In March, Lean In surveyed 1,015 American adults about AI at work.

Twenty-nine percent of women said they worry they’ll be perceived as cheating when they use AI. Twenty-two percent of men said the same. That’s a 32% relative gap, and it shows up everywhere else in the data, too.

Men are 27% more likely to have been praised for their use of AI. Women are 38% more likely to have ethical reservations and 29% more likely to question the accuracy of AI-generated outputs.

Now, if you’re a leader reading that, the tempting response is, “That’s a confidence problem. We’ll do a lunch-and-learn.”

Hold that thought.

And They’re Right to Worry

Earlier this year, Zehra Chatoo, a former Meta strategist who now runs Code For Good Now, ran an experiment that should be required reading for every hiring manager on earth.

She gave 1,000 UK adults the same résumé. Same skills. Same experience. Same disclosure that AI helped build it.

One group saw the name Emily Clarke at the top. The other saw James Clarke.

That’s it. That’s the whole experiment.

James got a 97% approval rating. Emily was rated strongly by 76%. Reviewers were 22% more likely to question Emily’s trustworthiness, and doubts about her competence were twice as high. Gen Z men called Emily’s résumé “weak” 3.5 times more often than James’s, which, if you were hoping the next generation would fix this, is not great news.

But the numbers aren’t what got me. The comments are.

On James, “He just needed a bit of help putting it together.”

On Emily, “She can’t even write a CV herself—not sure she has the skills.”

Same résumé. Same disclosure. He’s resourceful. She’s a fraud.

Chatoo’s summary of the finding is the line Tara built our discussion around, and it’s the line I can’t stop thinking about: “When men use AI, we question their effort. When women use AI, we question their integrity.”

So, no. It’s not a confidence problem. The women in the Lean In survey aren’t imagining the penalty. They’re absolutely correct.

So They Use It Less

Which brings us to…

Harvard Business School, Berkeley, and Stanford pulled together 18 studies covering more than 140,000 people and found that women adopt generative AI at about a 25% lower rate than men.

Their updated global synthesis puts current adoption at 47.8% for men and 39.3% for women, and says the gap has remained roughly 16% since early 2025.

Here’s what makes it damning rather than just disappointing.

The gap holds in both rich and poor countries. It holds at every education level. It holds inside the same job: at one tech company. And when Kenyan entrepreneurs were given identical access to ChatGPT and training, women were still 13% less likely to try it.

Equal access didn’t fix it. Training didn’t fix it. Because access and training aren’t the problem.

The researchers say women “anticipate larger penalties” for using AI and are more likely to view it as cheating. Koning’s read is that women worry others will think that “even though they got the answer right, they ‘cheated’ by using ChatGPT.”

Put the three studies together, and you understand that women truly fear the penalty. The penalty is real. So they use AI less, or use it quietly. And then, when AI fluency becomes the skill everyone’s hiring for, the people who were smart enough to see the trap are the ones who look like they never learned the tool.

Why This Is a Communications Problem, Not a Women’s Issue

I can hear some of you already. “Gini, this is important, but it’s an HR thing.”

Nope.

Our industry is 65.6% women. Two-thirds of our profession. Add in content, social media, and internal comms, and the number goes up, not down.

So do the math on what the research means for a comms team specifically.

The people on your team most likely to be using AI every day for drafts, research, media lists, and analysis are, by headcount, mostly women. Those are the people most penalized for admitting it.

Meanwhile, the people setting your organization’s AI policy—the ones deciding what’s allowed, what gets disclosed, and what counts as “real” work—are disproportionately the ones who get called resourceful for using it.

What you end up with is a team full of quiet AI users, no shared practices, no disclosure norms, and a leader who looks at the survey results and concludes adoption is low.

Adoption isn’t low. Admission is low.

In a measurement framework we are building for a client, we track AI use among their 11,000 employees. We survey employees monthly to get a read on how they’re using it.

At the start, fewer than 10% of the workforce reported using their internal AI, and even fewer used their CoPilot instance. The majority of those users are men.

We just looked at the August survey, and while the numbers have increased in both internal AI and CoPilot use, the percentage increase among women is minuscule compared to that of men.

And then last week, I was in a meeting where the chief comms officer showed me data on actual AI use. You know, the report that shows how many people are using it. Turns out almost everyone uses it daily, but only male employees will admit it in the survey.

And this is where it connects back to Karp. He didn’t say AI would replace women. He said it would reduce their economic and political power.

You don’t need a robot to do that. You just need a workplace where a woman’s most valuable new skill is the one she’s been taught to hide.

The Just-World Trap

Tara made a point that I am going to shamelessly steal because it explains why this research gets read and then ignored.

When people are shown evidence of bias, they have two easy exits. They can feel powerless and change the subject. Or they can decide it doesn’t apply to them personally, even with proof in hand.

Psychologists call the second one the just-world hypothesis. It’s the belief that people basically get what they deserve, so if I’m doing fine, the system must be fine.

In a comms team, it sounds like this, “Nobody’s ever questioned my AI use.”

Maybe. Or maybe you’ve never disclosed it. Or maybe you’re a man. Or maybe the question got asked about you in a room you weren’t in, the way it got asked about Emily.

You don’t have to have been Emily to care that Emily exists. And you especially don’t get to run a team, set a policy, or hire people while pretending she doesn’t.

If You Lead a Team

So what do you do about it?

If you lead a team, there are four very specific things you can do:

Make disclosure a norm, not a confession. The Emily problem exists because disclosure is optional and individual. When one person says, “I used AI on this,” and nobody else does, that person is volunteering for judgment. When everyone does it, as a routine step in the workflow, disclosure stops carrying information about the person and starts carrying information about the process. This is what it should have been all along.
Evaluate the output, not the origin story. If the brief is good, the brief is good. The Chatoo study shows the penalty lands on competence—reviewers decided Emily couldn’t do the work, not that she didn’t. So build your review process around whether the work holds up, and stop asking how it got made unless you ask everyone, every time.
Watch who gets asked, “Did you really do this?” You will not fix a bias you can’t see. So, for a month, pay attention to who on your team gets the follow-up question about their process and who gets a “nice work” and a move-on. If it’s not evenly distributed, you have your answer, and it’s not a training problem.
Praise AI use out loud, and check the distribution. Lean In found men are 27% more likely to have been praised for using AI. That’s not because they used it better. It’s because “resourceful” is the default read for one group and not the other. Praise is cheap, and it’s one of the fastest ways to reset a team’s sense of what’s rewarded.

If You’re the One Using It

If you’re the one using AI, there are four things you can do, too!

Use it out loud, and describe it as a workflow. Not “I used AI on this” (which invites the Emily read) but “I ran the first draft through Claude, threw out the intro, rewrote the argument, and fact-checked the three stats.” The résumé study tells you that the penalty lands on competence. So pair every mention of AI with the judgment you applied.
Get very good at the part AI can’t do. The skills Karp said AI would devalue—ethical reasoning, audience judgment, knowing what not to say—are the ones every serious AI researcher says we’re short on. Be the person in the room who knows why the AI draft is wrong. That’s not a hedge against AI. That’s the job.
Educate opportunistically. Don’t lecture. Wait for the moment when AI use comes up naturally, in a review, in a hiring conversation, in a planning meeting, and say, “I read something interesting the other day.” Then tell them about Emily and James. People check their own perceptions when the information arrives as relevant.
Share the research. LinkedIn buried Tara’s post. It cannot bury a link in an email, a Slack channel, or a team meeting. Forward the research.

The Part Most Teams Underestimate

I’ve spent the better part of this year writing about how AI systems decide whether to trust a brand. They look at what you claim about yourself and then check whether anyone else backs it up. Claims versus corroboration.

The PESO Model® works because owned media makes the claim, and earned, shared, and paid media corroborate it.

It turns out people evaluate people the same way.

A woman who says “I used AI to build this” is making a claim about her own competence. Whether it gets believed depends entirely on whether there’s corroboration around her: a visible track record, a leader who vouches for her judgment, or a team norm that says this is how good work gets made here. Without that, the claim gets discounted, and she gets read as Emily.

What most teams underestimate is that corroboration doesn’t build itself. Organizations that come to us thinking they have an AI tools problem or an AI training problem almost always turn out to have a trust problem, which is a communications problem, which is a leadership problem.

The tools are the easy part.

Two-thirds of this industry can’t afford to hide its most valuable new skill. And the industry can’t afford to let it.

If You’re Auditing AI Use on Your Team

If your team’s AI-use survey says “occasional” and your gut or the actual data says “daily,” the gap between those two numbers is the thing to investigate—and it usually isn’t a tools problem.

That is something we would love, love, love to help you solve. Shoot us a note, and we’ll happily spend 30 minutes brainstorming with you.