The Competence Penalty

A large technology company spent a year rolling out an AI coding assistant across 28,698 engineers. It was pre-installed, integrated into normal workflows, and promoted internally.

Twelve months later, adoption was still sitting at 41%.

What were the reasons? Perhaps the coding assistant wasn’t good enough yet, or worse, training hadn’t worked and people simply didn’t know how to use it well? Maybe incentives weren’t strong enough?

Those explanations seem pretty convenient because they turn adoption into a capability problem. They also miss something important in the dynamics.

Researchers tested what happened when other engineers knew AI had been involved.

In a preregistered experiment, 1,026 engineers reviewed identical snippets of Python code. Some reviewers were told AI had helped produce the code and others weren’t. When reviewers believed AI was involved, they rated the engineer’s competence 9% lower. The code didn’t change. What changed was what the reviewer thought of the person who wrote it.

The effect wasn’t limited to women. Male AI users took about a 6% competence hit. Female AI users took about 13%, with the heavier penalty coming particularly from male reviewers who hadn’t adopted AI themselves.

That difference matters, but the broader pattern isn’t limited to one group. Visible AI use changed what coworkers thought of the person whose name was on the work. Once that enters the decision, 41% adoption starts looking less mysterious. Someone deciding whether to use a coding assistant is also wondering, “if I use this and people find out, what will they think of me?”

What mandates don’t fix

Companies have responded to slow adoption by making AI use harder to avoid. Microsoft told employees that using AI was “no longer optional.” Shopify’s CEO told teams not to request more headcount until they could show that AI couldn’t do the work.

If adoption was too low, just raise the cost of not adopting it. Seems like solid logic, right?

Mandates change one side of the calculation. Now there may be a cost to not using AI, but the professional cost of being seen using it is still sitting there untouched. You can require AI use all day and still leave everyone wondering what happens when a reviewer sees it in their work.

Research from KPMG and the University of Melbourne covered 48,340 people across 47 countries. Among employees, 57% reported hiding their AI use and presenting AI-generated work as their own.

A company can tell its team members that they should be using AI and, without missing a beat, turn around and kneecap the effort by teaching them the opposite lesson in a single code review. One skeptical “walk me through this” can carry a lot of weight. The question itself may be completely reasonable. A reviewer should understand a risky change. But context and tone can make the question mean, “do you actually understand any of this?”

One person can use AI on a nasty refactor, catch the bad calls, challenge what comes back, and walk a reviewer through exactly why the final change is safe. Another person can hand everything to a coding agent and not come up for air until the tests are passing.

Those are not the same kind of work.

People notice when nobody seems to care about the difference. After a while, careful AI use gets harder to see because the ones doing it well have a reason to keep quiet.

So bring it up with the team

The reflex might be to go write another policy. Resist that temptation. Bring the dynamic right out into the open with your team members.

“I’m noticing there is some professional hesitation around whether it’s safe to be seen using AI. Let’s talk about it.”

Start right off with emotional safety and see what happens next.

If people think admitting AI involvement puts their competence on trial, another training session isn’t going to fix it. And turning the adoption dial harder until the dashboard looks healthier definitely isn’t going to fix it.

Before you schedule another training session, ask what people learned the last few times AI showed up in review. You might find out they understand the system just fine.