Keeping my voice in AI writing

I think I can often tell which model wrote something, because ChatGPT, Gemini and Claude all have habits I've come to know from using them, even when the topic changes.

The phrasing. The same tidy close. The way the model builds a point, then explains that point again.

Claude has been the most obvious to me. Opus 5.5 feels less so, but that's still just how it reads to me.

The writing can be clear and still sound like the model. If I'm publishing something under my name, I want it to sound like me.

The watermark is in the choices

Text watermarking works through the words a model picks, influencing how it picks its next word or word piece and leaving a pattern that a detector can check across enough text. There's nothing added to the font.

Google calls its tool SynthID. It shifts the chances of some words being picked, so the choices leave a watermark. Google says it uses this for text made in the Gemini app and web experience. I wouldn't see it by reading the words.

Anthropic uses a watermark for supported Claude models too. It affects word choices and adds no hidden characters. A model's style is a separate thing, as Anthropic also explains. The habits I hear in Claude's writing aren't proof of a watermark.

OpenAI said on October 5 that it plans to roll out text watermarking to eligible ChatGPT and Codex users in the EU. It also offers the option on selected API models worldwide (off by default). These tools don't all mark every piece of text they produce.

I don't know whether Grok uses a text watermark, because xAI describes watermarks for images and video but doesn't say what happens to text. That doesn't support calling Grok the only model without one.

Keeping my words

The writing habits are what I want more control over. A draft can be well written and still sound nothing like something I'd say, even if the thought came from me.

That's part of why I built Murmur. It takes samples of my own words and turns them into a voice card. Sentence rhythm. Word choice. Punctuation. The card gives the model a guide based on how I actually write.

The model gets that card when it drafts. Asking it to sound natural leaves a lot open, because its idea of natural might be quite far from mine. My own words give it something to work from.

A call has someone else's words in it too. I want to keep their voice apart from mine. The way they speak can give me something to compare my own words against.

Meeting AI Assist helps with that. It keeps my microphone audio apart from the other person's audio, which lets me keep track of who said what. Their phrasing shouldn't quietly become part of what I call my voice.

Then I give Grok the content with my voice card. Same card. I still need to read the draft and decide whether the words fit what I meant.

A voice card doesn't tell me whether the text has a watermark, but it gives me a way to guide the draft and check how it sounds. I want help putting the thought into words, and I want those words to feel like mine.