The AI Bully Problem

KPMG and the University of Melbourne ran the numbers in 2025 across 48,340 people in 47 countries. 57% of employees hide their AI use and present the output as their own. Slack's 2024 Workforce Index pegs the share who would feel uncomfortable telling their manager at 48%. Microsoft and LinkedIn put the bring-your-own-AI rate at 78%, with 52% hesitating to disclose it on the work that matters to them.

Ethan Mollick has a name for them. He calls them secret cyborgs.

The stigma is the cause. The hiding is the symptom. And every workplace I have walked into in the last two years has it.

I have lived this

I have experienced AI stigma directly. I have mediated it for others. I have coached leaders through how to head it off. The how comes later in this piece.

The pattern is consistent. A peer or a manager learns that someone used AI to draft, write, code, or decide. The reaction lands somewhere on a spectrum from passive ("oh, you used the bot for that?") to performative ("must be nice to skip the work") to actively punitive (write-ups, lost assignments, terminations). The Duke Fuqua study published in PNAS in 2025 confirms this empirically. When colleagues knew you used AI, they rated you as lazier, less competent, and less trustworthy. The penalty held across age, gender, and occupation. In the field experiment, workers gave up 3.4% accuracy to avoid being seen using AI when evaluators were watching.

Read that twice. Workers traded measurable quality for the appearance of independent judgment. That is the cost of letting the stigma sit.

The slop is real, and it is a user error

Now the honest part. A lot of AI output is slop.

Simon Willison gave us the word in May 2024. Merriam-Webster put it in the dictionary as the 2025 word of the year. NewsGuard now tracks 3,006 AI-generated content farms, up from 49 in April 2023. More than half of long-form LinkedIn posts are likely AI-written. Daniel Stenberg, the maintainer of curl, shut down his bug bounty program in January 2026 because he had spent six years receiving zero valid AI-generated security reports. Zero. The Workslop study from BetterUp Labs and Stanford Social Media Lab, published in HBR last September, put a price on it: 1 hour 51 minutes of rework per incident, $186 per employee per month, $9 million a year for a 10,000-person company. And the senders of workslop got rated 54% less creative, 42% less trustworthy, and 37% less intelligent by the people who received it.

So yes, the bad output is real, and the social penalty for being caught producing slop is rational. The problem is the conclusion most people draw from this. They conclude that AI is the problem. It is not. The user is. And the user is downstream of leadership choices about training, expectation, and what gets tolerated. Hold that thought.

The BCG and Harvard jagged-frontier study put 758 consultants through 18 tasks. Inside the frontier where AI is competent, AI users produced 40% higher quality output, 25% faster. Outside the frontier where AI is wrong, AI users were 19 percentage points less likely to land correct answers. Same tool. Same population. Different tasks, different outcomes. The skill that separates the lift from the loss is knowing which situation you are in.

Worth pausing on the dates. BCG ran that experiment on GPT-4 in 2023. Nobody serious uses GPT-4 in 2026. Multiple generations of frontier models have shipped since, alongside dramatic improvements in process and technology, while the cost of inference (the part where AI computes and sends you a reply) has fallen many times over. Almost every study I have cited above was run on tools that are already obsolete. That does not soften the findings. It sharpens them. The lift inside the frontier is larger now. The failure modes outside it are stranger. The gap between trained and untrained users is wider, not narrower. The studies on the shelf describe a world the tools have already left behind.

Microsoft's 2024 numbers say the training side cleanly. 75% of knowledge workers use AI at work. Only 39% have ever received any AI training from their company. Only 25% of companies planned to offer it that year. And the gap held forward. KPMG's 2025 global study, the same 48,340-person survey I cited at the top, found 47% of employees had received any AI training. The tools have moved. The training has not.

Using AI well is a learned skill. It is not magically intuitive. Context-first prompting, iteration, validation, taste, knowing when to pull the work back into your own hands. None of that comes free. And the people who are mocking colleagues for AI use are mocking people who are doing untrained work in a tool nobody taught them to operate.

The stigma is the second-order effect of a training failure. The fix is not less AI. The fix is more skill.

This is not the first time

Every meaningful workplace technology has produced the same stigma cycle.

It is not hard to trace the pattern back five hundred years. When the printing press arrived in the late 1400s, the abbot Johannes Trithemius wrote a treatise called In Praise of Scribes, arguing that handwritten manuscripts were spiritually superior to printed books. He had it printed on a press, because by then that was the only way to circulate a book widely enough to be read. The contradiction did not slow him down. Scholars and clerics fought to keep printed books out of legitimate scholarship for decades. Mark Twain was teased for using a typewriter in the 1870s, and literary writers held out for the dignity of the pen for another generation. Schools in the United States banned ballpoint pens into the 1960s on quality grounds, after they had already become standard in business.

Engineers carried slide rules until 1972 because slide rules were the symbol of the profession, the way a stethoscope is for a doctor. Then the HP-35 landed at $395, and within a decade the same teachers who had policed against calculators were teaching with them. NACOME had to formally argue in 1975 that students should be allowed to use them. Spreadsheets were considered a threat to the integrity of the accounting profession in the 1980s. Executives in the 1990s wouldn't type their own emails because typing was secretarial work. The same pattern hit the internet at work (cyberslacking, 1999), Wikipedia at school (banned, mid-2000s), and GPS in aviation, where instructors actively criticized students for using GPS instead of dead reckoning. The line endured. "Real pilots didn't need GPS."

Real pilots didn't need GPS. Real engineers didn't need calculators. Real accountants didn't need Excel. Real executives didn't type their own correspondence. Real students didn't cite Wikipedia. Real programmers didn't use Copilot. Real scholars copied by hand.

The sentence pattern is older than this argument by centuries. It reverses every time, and the reversal is faster in places where leadership took an explicit stance, slower in places where leadership stayed neutral.

What leadership has to do

This is where I take the line a lot of people don't like.

Mocking a colleague for using AI to do better work is bullying. It is not banter. It is not professional skepticism. It is not "just keeping standards." It is repeated mistreatment of a colleague for adopting a tool the company benefits from. SHRM's progressive discipline framework already has the right ladder for it. Verbal warning. Written warning. Suspension. Termination. The structure exists. Most companies just refuse to apply it to this category of behavior.

There is also the question of the holdout. The honest answer is that refusing progress has rarely been a viable path inside a working business. Luddite patterns persist but they do not prevail. Religious communities that slow technological adoption in favor of family and shared life are the principled exception, and that choice deserves respect, not derision. Outside that frame, an employee who refuses to learn the dominant tool of the job is opting out of the job.

The school analogy holds, with one important nuance. The research on school discipline does not endorse blanket zero-tolerance, and neither am I. Zero-tolerance shortcuts produce worse outcomes than progressive discipline. So the right read is not "fire on first offense." The right read is the ladder. Make the expectation explicit. Have the private conversation. Document the formal warning. Suspend if it persists. Terminate if it persists past that. The ladder works precisely because it is a ladder, not a guillotine.

The reason this matters is mechanical. The bystander research is unambiguous. When leadership tolerates the mocking, mocking becomes the culture. When leadership intervenes consistently and visibly, it stops. And the Duke PNAS study has the cleanest finding of all on this question. The penalty for using AI vanishes when managers use AI themselves. Leadership modeling is more powerful than leadership messaging. Silence from leadership is, in practice, endorsement of the stigma.

There is also a do-not-do-this list. Tobi Lütke's 2025 Shopify memo is the cleanest hard-line precedent on record. "Reflexive AI usage is now a baseline expectation. Everyone means everyone, including the CEO." That part is right. Klarna's overreach (claiming AI could do all the jobs, then quietly rehiring) and Duolingo's tool-usage performance metric (which the CEO walked back a year later) show the failure modes. Mandate the use. Do not score people on token counts. Do not pretend AI replaces judgment. There is now a whole genre of managers tracking employees by AI token count, including Meta's internal "Claudeonomics" leaderboard, because tokens have nothing to do with productivity. They are the byproduct of producing something. The cleanest version of the warning is this. "Use AI" is not a metric. "This work product would have taken 10 hours and now takes 2" is.

Progress is not optional

Some people will not like any of this. I am aware. I have heard the objections, often from very smart people, often delivered as concern for craft, integrity, or fairness.

The line I keep coming back to is this. The engineers who refused calculators are not engineers anymore. The accountants who refused spreadsheets are not accountants anymore. The executives who refused to type their own correspondence type their own correspondence now. The pattern is reliable enough to bet on, and the cost of betting against it is paid by everyone who tolerates the bullying instead of stopping it.

That said, no one is truly forced. Trithemius was not forced to print In Praise of Scribes on a press. He chose to. Had he chosen not to, that would have been perfectly fine. We just would never have heard of him. Opting out is a real option, and sometimes it is the right one. The cost is invisibility. The pattern is not coercion. It is the math of what gets carried forward.

The hard line is not cruelty. It is the actual mechanism that protects the people who want to do the work.

One question

What conversation, on your team, are you avoiding because you are worried about the reaction, when the reaction is exactly what is keeping the stigma alive?

Reply with one. I read every response.

Kent