The AI Watermark Is Hidden in the Words

Name: The AI Watermark Is Hidden in the Words
Uploaded: 2026-08-15T10:00:20.000Z
Duration: 1 min 51 s
Channel: Authority Hacker Podcast
Description: AI watermarks can be embedded in language models' word choices, making simple removal techniques less effective.

Authority Hacker Podcast15 August 2026Watch on YouTube

Part of series

Ep. 9 · Onzichtbare AI-Watermerken

Alles over de opkomst van verplichte watermerken in AI-gegenereerde content en de implicaties voor gebruikers.

View the series

What you'll learn

AI text watermarks are not hidden code characters but statistical patterns in a model's word choices.
Across enough words, subtle shifts in chosen synonyms create a detectable signature.
Unlike the code characters ChatGPT once added by accident, this watermark cannot be stripped with a text editor or an OCR pass.
The system is already live in Claude, and text generated over the past week already carries the signature.

Frequently asked questions

How does the AI watermark in Claude's text work?

At moments where several words fit, the model gives some choices a slightly higher weighting. Across enough words this creates a statistical pattern in word choice that acts as a recognizable signature.

Why was ChatGPT's earlier watermark easy to remove?

Back in April ChatGPT accidentally added hidden unique code characters to its output. They were easy to spot and delete in a text editor, after which the software removed them.

Can the new watermark be bypassed by copying or scanning text?

No. The speakers say text keeps its key even after copying, photographing or an OCR pass, because the key is embedded in the text itself.