Sometime in the next year, in an office no stranger than yours, somebody will paste a colleague’s report into a watermark checker ten minutes before the meeting. The verdict will land in the room before the author does: AI involvement likely. She wrote that report herself — the analysis, the judgement calls, the careful paragraph explaining why the May numbers dipped — and then ran it through Claude to tighten the wording, the way millions of us now do. The paste-in strips all of that away and leaves one word on the table: detected.

The machinery for this moment arrived in August. Anthropic announced that Claude’s output now carries an invisible watermark (a statistical pattern in its word choices, added to comply with the EU’s AI Act), with a detection API to follow, and the other major model providers have signed up to the same code of practice. John Gruber, who called the whole scheme a perversion of writing, spent a long Daring Fireball post imagining precisely the scene above: office workers copying each other’s emails into detectors, then wondering what the match would even get them.

I’d answer Gruber’s question less breezily than he asks it, because offices have always known what to do with ammunition, and the people it hits won’t be the bad actors the policy imagines. AI’s great gift to the ordinary knowledge worker is superpowers — an average performer suddenly producing more, and better, at the same job. A workplace where AI use can be checked, and where a positive check carries a smell, poisons exactly that. The stigma won’t reduce AI use; it will drive it underground, and the people who pay will be the honest ones.

The asymmetry favours the accuser

Running the check costs thirty seconds and no courage at all, while answering it takes a paragraph of nuance that nobody in a tense meeting wants to hear. The watermark means processed by, not written by — Anthropic’s own FAQ is explicit that it can’t distinguish “Claude wrote this” from “Claude heavily edited this”. So a report drafted by a human and polished by the machine can test positive, while a fully machine-written report run through a paraphrasing tool can test clean, since anyone motivated to hide their use can strip the mark with tools that already exist in the open. The practical result is a mark carried mostly by people who had nothing to hide, while the committed evaders skate. How reliable the detection tooling will prove is unclear to me — Gruber doubts the scheme survives contact with the real world at all — but office politics doesn’t need a reliable verdict, only a fast one, and in a hostile room a positive result on an honestly made report reads one way: caught.

Consider what the verdict would say about Allison Stanger, the Middlebury College professor who published a piece in the Wall Street Journal last week titled “AI Makes Me a Better Writer”. Her rules are a model of open use: never outsource the important parts, because “writing is a means to learning who you are and what you value”; use the machine as an adversarial reader rather than a writer; hand it the mechanical work without apology. “AI didn’t write this article,” she says, “but it helped me write it.” Paste her columns into a detector and they could still register, because the mechanical work is made of the machine’s word choices. The most virtuous AI-assisted writing process I’ve seen described in print is one a colleague’s thirty-second check could flag just the same.

The penalty was measurable before the detector shipped

We don’t need to speculate about whether people will hold the mark against each other, because researchers measured the penalty before any watermark existed. A 2025 study in PNAS ran four experiments with more than 4,000 participants and found that people who use AI at work attract harsher judgements of their competence and motivation than colleagues doing the same work without it — and that the users anticipate this, and become less willing to disclose. Atlassian ran the blunt version this year: 961 knowledge workers evaluated an identical piece of work from the same hypothetical colleague, the only difference being a short note saying AI had helped. The disclosers were rated ten times lazier and were markedly less likely to be recommended for high-visibility projects.

Workers have done this arithmetic for themselves. PagerDuty found this summer that two-thirds of office professionals have used AI tools they believed weren’t permitted, and with workplace AI use now past half the US workforce on Gallup’s count, the safe assumption is that some of your own team is among them and saying nothing. The silence is rational, too, since showing your work invites a judgement and hiding it has, so far, invited none.

Managers hold the other half of this problem, because downgrading a watermarked report is the cheapest available way to look rigorous — you never have to engage with the content at all. Every manager who does it once teaches a whole team what disclosure costs.

What the hiding costs your business

If you run a business, the cost of all this lands on you before it lands on anyone’s feelings. Once stigma sets in, your adoption dashboard ends up measuring who’s willing to be seen using AI rather than who’s using it. Your keenest adopters go invisible, the training budget gets aimed at the half of the room comfortable being watched, and the adoption numbers you report upward become a measure of social courage. I’ve written before that your AI strategy is probably theatre — and your employees know it; stigma is the same play running in the other direction, with employees performing non-use for an audience of managers.

The first cost is the one the whole business case rests on: the average performer doing more — the superpowers — across every report, proposal and analysis. A status penalty on visible use shrinks that gain at the moment you most want it growing. The second cost may be worse. Peers running checks on peers corrodes something small firms can’t buy back: a business of fifteen people runs on trust the way big firms run on process, and having spent seventeen years running an agency, I’d have traded almost any process failure for the damage a culture of mutual suspicion does. I’ve argued that agent sprawl is the new shadow IT and treated hidden AI as a governance problem, which it is; but underneath the governance problem sits a status problem. People hide what they expect to be judged for, and no acceptable-use policy fixes a judgement problem.

What ended it for the calculator

We’ve run this experiment before. Calculators were banned from exams as cheating, spellcheck marked you out as semi-literate, and Grammarly was something you used but didn’t mention. Nobody ever pasted a colleague’s spreadsheet into a calculator checker; each of those stigmas faded when the people whose judgement mattered used the tools openly and kept judging the work on whether it was good.

That’s also what the Atlassian research found when it looked past the headline penalty: the laziness stigma nearly disappears in companies that actively celebrate AI use. The stigma tracks what leadership rewards, which means an owner or manager can end it deliberately:

  1. Use AI openly yourself, and say so. Nothing normalises faster than the boss’s visible example. Stanger putting her name over a national-paper column describing exactly how AI helps her write is the move in miniature: disclosure converts the watermark from a gotcha into a footnote.
  2. Write down that a detection result is never a verdict. Make it policy: the mark records processing, and says nothing about authorship or quality, so no evaluation of anyone’s work may cite one.
  3. Praise sanctioned use in public. The Atlassian finding cuts both ways, because the same social machinery that produces the penalty can produce the norm.
  4. Judge output on quality, with a named human standing behind it. Openness keeps the bar where it always was — attached to the work and to the person who owns it. Accountability survives; only the ambush goes.

Gruber ends his piece advising us to stop caring who — or what — wrote a thing, and simply judge whether it’s good. He’s right about the destination. But “just judge the writing” is advice for readers, and inside an organisation it only becomes safe to live by when a leader goes first.

The watermark is already in the text, and the detectors are coming. What remains open is what your workplace decides a positive result means. Decide — out loud, as policy, before the first paste-in happens — that it means nothing without a human judgement attached, and the only live question about any piece of work remains whether it’s good, and who stands behind it.