I'm Fine Being Flagged

Tue, 25 Aug 2026 07:00:00 -0400

Anthropic recently announced that Claude models are going to start watermarking the text they generate, with older models getting the feature added over the coming months. The stated reason is compliance with the EU AI Act, which requires AI-generated content to carry some kind of disclosure marker, and it’s a poorly written law, in my opinion, vague enough that I’m not entirely sure how it’s supposed to apply to something like my own workflow.

Reading the announcement, the watermark attaches to words Claude chooses, not to text a person wrote and Claude only lightly edited, so a heavily rewritten draft might not carry much of a signal by the time it’s published. I don’t know exactly where my own drafts land on that spectrum, and I’d rather not guess. What I do know is that it’s likely, or at least plausible, that some of what I publish here is going to get flagged as Claude-processed at some point. I want to be upfront about my process before that happens, rather than after.

So here’s how I actually write these posts. I come up with the idea. It’s mine. Then I talk through it out loud and use voice transcription to get everything I want to say down in words. From there, I’ll ask Claude for any follow-up questions, the kind that surface gaps or push me to be more specific, and I answer those the same way, out loud. Claude takes everything I’ve spoken across that transcript and organizes it into a draft. Then I go through it and decide, line by line, whether it actually sounds like me. If it’s expressing what I meant in the way I’d like to say it, I’m OK with leaving it alone. If it doesn’t, which is often, I fix it. I’ve also built up reference documents on my own writing style and tone, with examples pulled from years of my own writing, specifically so that draft has a better shot at sounding like me in the first place. That’s been my process for months now, refined over a lot of iteration, but the shape of it, talk first, write second, has been true for even longer than that.

I use Claude for other things too. Site development. Processing links for my newsletter. It’s woven into a lot of what I do, not just the writing. I don’t think that should be hidden, so it’s going in the disclosure along with everything else.

The distinction I actually care about is not “AI touched this, therefore it’s suspect.” I think we can all agree to loathe a “meat proxy,” someone who forwards AI-generated output without reading it, understanding it, or validating it before passing it along. That’s one end of an extreme. My actual issue is narrower: it’s when something took someone very little effort to produce, including the thinking behind it, and I’m still expected to spend real time and attention engaging with it. A model arranged some of the sentences in this piece. That’s true, and I’m not going to pretend otherwise just because I read every word before it went up. What I’d point to instead is the whole shape of the process: the idea is mine, the raw material comes from me talking it through out loud, and I stay involved enough in the editing that I can tell you, honestly, when I left something alone versus when I rewrote it. That’s a different thing than handing a topic to a model and pasting back whatever it returns.

I also won’t pretend I need Claude to produce work worth reading. What it actually gives me is sustainability. I’m not paid for anything on this site. I write in whatever hours I can carve out around a demanding job and the rest of my life, and time is the scarce resource, not ideas. This workflow lets me get from a spoken idea to a polished piece inside the time I actually have, instead of the time writing it alone would take. This site has now outlasted every other project I’ve tried to keep running online, and I think that’s largely because this process is sustainable in a way my earlier attempts weren’t.

Now, about Anthropic’s stated reasoning. To be fair to them, the announcement is explicit: they signed the EU’s Code of Practice on Transparency of AI-Generated Content and are watermarking to comply with the AI Act. That’s the actual, on-the-record reason, and I have no evidence it’s false. But I can’t shake the suspicion that it’s also a convenient reason, one that sounds responsible and externally imposed while covering something they’d likely want anyway. Being able to identify their own models’ output once it’s loose on the internet would help them keep it out of future training data, where synthetic text can quietly degrade the next generation of models, a problem researchers call model collapse. I can’t prove that’s part of the calculus here. It’s a guess. But a law that happens to hand you a tool you’d want to build anyway is the kind of coincidence worth pausing to consider.

None of which changes my answer. I’m fine with my posts getting flagged as Claude-processed, whenever and however much that ends up applying. If that’s a dealbreaker for a reader, I respect that and I’d rather they know upfront than feel misled after the fact. What I’d push back on (with apologies to Hank) is the reflex to treat “AI was involved” as shorthand for “not worth your time.” I originate every idea on this site, I supply the raw material by talking it through myself, I put a lot of work into making my content as good as it can be, and I hold final judgment over (and responsibility for) every sentence that goes up. That’s a different thing than the picture “AI-generated” usually conjures, and I think it’s worth more consideration than a watermark can convey either way. I’ve written up the specifics of how I use these tools on a separate AI disclosure and policy, if you want the fuller picture.

Transparency on tomgromak.com

I'm Fine Being Flagged