Right now, the most rewarded thing you can do with AI at work is use it heavily and never mention it.
That is not cynicism. It is roughly what the evidence says. And it should bother us more than it does.
Concealment works and candor costs
Andras Molnar and Jiaqi Zhu at the University of Michigan put more than 1,300 American participants in front of personal messages. When people read AI written text with no label attached, they formed impressions just as warm as the people who had been told a human wrote it.
They could not tell.
The detail that should stop you is who could not tell. Heavy AI users, light users and people who never touch the tools all assumed a person was behind the words. Daily familiarity bought no advantage whatsoever.
Then the researchers added the label. Shown identical text and told a machine produced it, readers called the sender lazy and insincere. The same words credited to a person got called genuine and thoughtful.
Nothing about the writing changed. Only the admission did.
A University of Florida study of 1,100 professionals found the workplace version. When a supervisor's message leaned heavily on AI, between 40 and 52 percent of employees judged that supervisor sincere. When it did not, 83 percent did. Ratings of professionalism fell from 95 percent to the low seventies.
Stack those findings and the incentive barely needs spelling out. Use is close to free. Admission is expensive. Detection is close to nonexistent.
Every honest person in your organization is currently paying a tax for their honesty. Nobody designed that. It is simply what the numbers produce.
A norm that holds only while nobody checks
Here is why this is not a comfortable place to settle.
The penalty in this research attaches to disclosure, not to use. That is a strange thing for a penalty to attach to, and it holds only while two conditions hold with it. Detection has to stay unreliable, and readers have to keep extending unlabeled text the benefit of the doubt.
Neither is a law of nature. Both are features of one particular moment.
The moment readers begin assuming machine authorship by default, the penalty stops attaching to the admission and starts attaching to the suspicion. And suspicion cannot sort the honest from the concealing. That is the entire problem with it. It lands on everyone, including the person who wrote every word unaided and now has no way to prove it.
So here is the prediction, and note what it does not require. It does not require detection to improve. The Michigan result suggests that ceiling is low and may stay low. It requires only that assumption changes, which needs no technology at all. One widely reported incident will do it. So will simple volume, the ordinary process of people receiving enough polished, frictionless, faintly hollow messages that they stop giving the benefit of the doubt.
Everyone who spends this period optimizing for concealment is building credibility on a condition with an expiry date.
Trust already moved from the text to the record
There is a clue about where this goes in how buyers have already adjusted.
TrustRadius surveyed 1,862 technology buyers and 444 vendors in January. Among buyers using AI in their purchase process, 94 percent said they fact check what it tells them at least some of the time. Seventy four percent lean on reviews. The things that actually move a decision are demos, trials, prior experience, and other people's accounts.
Notice what every item on that list has in common. Not one of them is a piece of writing you are asked to take on faith. They are all expensive to fake, and all checkable by somebody other than the person making the claim.
Buyers did not decide to distrust text. They repriced it, quietly, and moved their trust to whatever survives verification.
That points at the real answer, which is not better prose and not a cleverer way to hide the tell.
Stop defending authorship and start producing a record. Almost nobody outside your company sincerely cares whether a model drafted a paragraph. What they care about is whether the claims were checked, whether the sources can be named, and whether a specific person is willing to be wrong in public about it.
So disclose at the level of the process rather than the sentence. "A model helped write this" is noise, and it is precisely the disclosure the research says gets punished. "Here is where this number came from, and here is who stands behind it" is the one that does work. It also happens to survive the shift that is coming, because it never rested on who did the typing.
The honest summary of this moment is that it rewards quiet use and punishes candor, and most organizations are responding exactly as you would expect. That is a rational answer to this month's incentives and a poor answer to next year's.
The credibility worth building is the kind that does not depend on anyone believing you wrote it.