---
id: "t-2026-08-18-a-watermark-on-ai-text-only-hurts-if"
type: "thought"
title: "A watermark on AI text only hurts if AI help is something to hide"
date: "2026-08-18"
week: "2026-W34"
url: "https://nosov.ai/writing/a-watermark-on-ai-text-only-hurts-if/"
linkedin: "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7495537376199557120"
author: "Ivan Nosov"
---

# A watermark on AI text only hurts if AI help is something to hide

A watermark on AI text only hurts if being AI-touched is something to hide.

Anthropic turned on invisible text watermarking for Claude this month, a way to solve for EU AI Act requirement in force since August 2. The feed filled with cancellation threats. Nothing about the model changed.

Now you can just be open about AI everywhere.

The draft, the edit, the summary. The same tools sit in every workflow, and have for some time. Watermarked text mostly means Claude processed it. Unmarked text probably means another model did. Neither says who owned the call, or whether the judgment behind it was any good.

Panic about the mark only appears if the assistance was supposed to stay off the record: keep the speed of the model, keep the credit of a blank page. Those two claims cannot both be true.

The technical objections arrived next.

The mark only acts where the model has free choice, a comment or a synonym, so code that has to compile is largely untouched.

Google has used the same class of watermark on Gemini for two years, tested on about 20 million responses, and reported no measurable drop in quality. The signal sits in word choice, not in the reasoning that produced the answer.

If the analysis is right and the judgment is yours, the adjective Claude picked is a detail.

Demanding that AI use stay a secret is the part that does not add up.

For research and writing for this post a mix of Grok 4.6, Gemini 3.1 Pro, Kimi K3 Max was used through Cursor. So probably no watermarks, but the AI is clearly here.