Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs™. Each week I dig into AI, knowledge management, and future-of-work research so you don't have to.
EDITOR'S NOTE
Last week ended on the line I could not stop thinking about, that the people who find the most have the least reason to tell anyone. This week is why, and it is less about character than I expected.
They are not being cagey, they are being correct. Saying you used AI carries a cost you can measure, and workers predict the size of it accurately. But the worst outcome is neither disclosing nor staying quiet. It is being found out by somebody else.
Plenty of people are making that bet. Across 48,340 employees in 47 countries, more than half told the University of Melbourne and KPMG they hide their AI use and present the output as their own, while only two in five said their workplace has any policy on it. What surprised me is that the penalty driving all this is barely attached to the tool.
———————————————————————————
THIS WEEK'S READING:
If You're Very Clever, No One Knows You've Used It: CHIWORK 2026
Researchers interviewed 19 UK knowledge workers about how they learned to use AI at work, from solicitors to lecturers to HR officers, and found a second skill nobody had assigned.
Why It Matters: One participant, a lecturer, learned to spot AI in his students' papers by its vocabulary, then turned that detector on his own writing. "There are indicator words, like 'leverage', 'delves', 'underscores'. So when I write, I make sure that those words are not in." The literacy on display is not prompting, it is cue removal, and nobody taught him to do it.
Notice who is doing the scrubbing. A lecturer's standing rests on being the person in the room who knows, so looking assisted costs him plenty. A different participant says flatly, "I don't hide it. I'm not ashamed of using it," because nothing about his position is threatened by the admission. Same tool, same output, and what differs is only what the job quietly claims about whose brain did the work.
Nineteen people, though. Are they right to hide it?
Four experiments with more than 4,000 evaluators found that people expected to be seen as lazier and less competent for using AI, and when observers rated them, that is what happened.
Why It Matters: Workplace fears are usually inflated, so it is worth pausing on one that is calibrated. The paper reaches back to Plato, who worried whether people leaning on a new invention, writing, would ever develop real wisdom. The lazy read is that he was wrong and we are fine. He was half right. Individual memory did degrade, nobody recites epics anymore, and what improved instead was the collective record, its accuracy and the fact that you can go check it. The trade was real, it just was not the one he was measuring.
The better example comes three sentences later, where the authors note that patients assume physicians who use diagnostic aids are less capable. That is the same penalty decades early, where the stakes are somebody's life and the tool is unambiguously good. And it is a misread. The aid is not the doctor doing less, it is the doctor checking himself and considering an angle he might not have reached alone.
So say nothing. What does that actually cost?
The Transparency Dilemma: How AI Disclosure Erodes Trust: Organizational Behavior and Human Decision Processes
Thirteen experiments covering 4,093 participants, including supervisors, investors and legal analysts, produced a clean ranking. On a trust scale where higher is better, saying nothing scored 4.02, disclosing scored 3.15, and being exposed by a third party scored 2.49.
Why It Matters: That ordering is the whole game. Silence beats disclosure, which is why so many people are silent, and why calling them dishonest misses what they are doing. But exposure is the worst outcome available and it is the one you do not control, so every hidden use is an open position on nobody ever noticing.
The uncomfortable part is that the researchers tested six disclosure formats, including a human-review note and an explicit transparency statement, and all six scored worse than saying nothing. Before concluding that nothing works, notice what these studies have in common. They are strangers judging people they will never meet again. Nobody has tested disclosure inside a relationship where you already have a track record, and that is where almost all of it happens.
One study moved the number, by changing who the value was for.
Vendor research, so weigh it accordingly, but the experiment is clean. Across 961 knowledge workers, disclosing AI use cost 24 percentage points on being recommended for a high-visibility project with the output held identical. Then they changed a single clause.
Why It Matters: Workers who framed the use as supporting the team and the client, rather than as saving themselves time, scored 11 points higher on effort, 45 percent to 56, and 8 points higher on recommendation, 43 percent to 51. On its own that reads as a communications trick. I do not think it is. If the only honest sentence available to you is that it saved you time, the value really did stop at you, and the penalty is tracking something true.
Which is why the force-multiplier version holds up where the wording exercise does not. When my product scopes get more accurate, engineering gets a workload projection they can plan against and the company stops misallocating people. That is not a story I tell about myself, it shows up in JIRA. Same as the support agents in last week's issue, legible because somebody was already counting chats per hour. If nothing counts your work, your gain stays invisible no matter how you phrase it.
———————————————————————————
Put that together and the pattern is not that people distrust AI. The penalty scales with how much your profession's standing rests on looking unassisted. The lecturer scrubs his vocabulary, the physician gets read as weaker for checking a reference, and the data analyst could not care less. Which is why "just be transparent" is useless as universal advice.
My rule is that you disclose when the machine's judgment is load-bearing, the interpretation and the nuance and the numbers, and you do not when it is only your own judgment moving faster. Reaching for the right tool and then checking what it hands you is a hammer instead of whatever else was lying around, and nobody files a disclosure for owning a hammer.
But all of this sits downstream of how the work gets priced. Under hourly billing, every efficiency you disclose is revenue you hand back, so concealment there is not psychology, it is arithmetic. Bill per deliverable and the incentive inverts, because the compression is yours to keep, and it shows up as a client who can suddenly afford a scope that was out of reach. None of the research above models that, and I suspect it is the largest lever in the problem.
So, mine. This newsletter is built with AI doing the retrieval and helping me articulate, and with my own research and industry experience deciding what makes it in. I will also say the obvious thing, which is that this is the cheapest disclosure in the world. I run an AI research lab, so being on brand costs me nothing. A solicitor or an underwriter has the opposite problem, and I will not pretend my position generalizes to theirs.
Last week I asked what you tried and who you told. If the answer was nobody, the research says you were probably right, and that is the part worth fixing. It raises the next question too, because if judgment is what makes any of this legitimate, somebody has to be doing the checking. In that same global survey, two in three said they rely on AI output without evaluating it. That is next week.
Know someone quietly sitting on a workflow they have never told anyone about? Forward this.
Have a take on any of this? Reply directly, I read every response.