David Monnerat

Product + AI | Systems Thinker | Enterprise Reality

Tag: writing

  • What If I’m Discovered

    What If I’m Discovered

    Watermarking isn’t the problem. Not knowing what to do with what it finds is.

    I’m writing a memoir. It’s about my son, Mitchell, and the years since his epilepsy diagnosis, the seizures, the hospital rooms, the slow work of building a life around a condition that doesn’t go away. The stories are mine. The experiences are mine. Most of the words are mine too. I use AI to help structure it, to push back when a chapter drags, to catch the places where I’ve buried the actual moment under too much explanation.

    When Anthropic announced it would start watermarking AI-generated text, my first thought wasn’t about the policy. It was: what if a passage gets flagged, and someone assumes the whole book was written by a machine.

    The Fear Isn’t About Cheating

    I know what I did. I lived every scene in that manuscript. Nobody handed me the chapter where my son’s friend waits outside the house while I drive him to a dance I never got to go to myself. That’s mine, down to the tie he wore. What AI helped with is structure and editing, the same job an editor does, the same job Grammarly does every time it flags a passive sentence. Nobody accuses a memoirist of fraud for having an editor.

    But a watermark doesn’t know any of that. It can’t tell the difference between “AI wrote this scene” and “AI helped me cut forty words from a paragraph that was already mine.” It measures something closer to word-choice statistics than authorship. The detection isn’t even reliable yet, the watermark can be worked around, by AI rewriting over it, by a person editing it out without meaning to. But most rules aren’t built for perfect enforcement. The intention is the point, not the catch rate. And intention doesn’t need to be accurate to make someone nervous. The fear isn’t that I did something wrong. It’s that a technical signal, imperfect or not, might override the actual truth of how the book got written, and I won’t get a say in which one people believe.

    The Same Fear Shows Up At Work, Wearing Different Clothes

    I use AI constantly at work too, drafting messages, working through how to explain something technical to people who don’t have my context. I don’t use it blindly. Left alone, it doesn’t sound like me. It reaches for more words than the point needs. I spend real time editing it down, adjusting how I prompt it the next time, the same way I’d revise my own first draft.

    Here’s the problem. If someone reads a message from me and thinks, this must be AI, the message can die right there. Not because the reasoning was wrong. Not because the point wasn’t true. Because the moment AI gets suspected, some readers stop evaluating the content and start evaluating the process, and they don’t like what they imagine the process to be.

    That’s a different fear than the memoir one. With the memoir, I’m afraid of being falsely accused of not doing the work. At work, I’m afraid of doing the work and still losing the reader, because the tool I used to communicate becomes the reason they stop listening.

    The Real Cost Isn’t Wrongdoing, It’s Losing the Reader

    Both fears come from the same place. It’s not about whether I did something wrong. It’s about losing control of how the work gets read the moment someone detects, or thinks they detect, AI in it. That loss of control has nothing to do with whether my reasoning held up or my story was true. It’s a judgment about method that overrides a judgment about substance, and once it fires, there’s no getting the reader back to the actual page.

    I don’t think the answer is announcing every tool I touch. Grammarly is AI. Autosuggest is AI. Nobody’s asking for a disclosure footnote every time spellcheck catches a typo. But somewhere between that and a watermark quietly flagging a memoir chapter, there’s a line, and right now nobody’s drawn it. We don’t have a shared sense of what counts as “helped me think” versus “did my thinking for me,” so readers default to the worst assumption whenever the question comes up at all.

    That’s the actual problem. Not the tool. The fact that we built the technology to detect AI assistance faster than we built any shared understanding of what to do with what it finds. A norm based on degree, how much of the thinking was mine, how much was structure, how much was polish, would serve people better than one based on a binary yes or no. I don’t know exactly where that line sits. I’m not sure anyone does yet. But I’d rather sit with that question honestly than pretend a watermark has already answered it for me.