AI Text Now Has an Invisible Watermark, Until You Rewrite It
Model and Machine
118 views
On October 5, OpenAI explained how it will watermark the text its AI models write, and it published the first numbers on how well that mark survives editing. There are no hidden characters and nothing is added to the text. The watermark lives in which words the model picks, and in OpenAI's test, swapping a quarter of the words for synonyms cut detection from about 92% to 17%.
You'll see how a language model picks each word, how a secret key turns that random choice into a pattern only the key holder can check, how OpenAI's textGrain and the SynthID-Text method behind Claude's watermark work, why short passages and math are hard to detect, what editing does to the signal, and what a positive or negative result can and can't tell you.
About the evidence: the detection rates (about 80% at 200 tokens and 95% at 400 tokens at a 1% false-positive rate, and 92% → 66% → 17% after 10% and 25% synonym swaps on 400-token passages) and the benchmark comparison are OpenAI's own tests. Anthropic's statements about editing, cost and what a match proves are Anthropic's own. The group-and-key diagrams are simplified illustrations of OpenAI's technical report. "200 tokens is roughly 150 words" is our rough conversion. We did not run either detector; access is limited to approved organizations. Sources checked October 6, 2026.
Chapters
00:00 The watermark you can't see
01:01 How a model picks each word
03:17 Who ships it, and where
04:28 Why length matters
05:11 What editing does
05:51 What a match proves
06:50 A signal, not proof
Sources
OpenAI, "Our approach to EU text provenance rules", October 5, 2026: https://openai.com/index/eu-text-prov...
OpenAI, "textGrain: Entropy-Calibrated Watermarking for Language Model Text" (technical report), October 5, 2026: https://cdn.openai.com/pdf/e9508624-d...
Anthropic, "How Claude's text watermark works", August 14, 2026 (updated September 1): https://www.anthropic.com/news/claude...
Dathathri et al., "Scalable watermarking for identifying large language model outputs", Nature 634, 818–823 (2024): https://doi.org/10.1038/s41586-024-08...
European Commission, Code of Practice on AI-generated content: https://digital-strategy.ec.europa.eu...
AI Text Now Has an Invisible Watermark, Until You Rewrite It
Model and Machine
118 views
On October 5, OpenAI explained how it will watermark the text its AI models write, and it published the first numbers on how well that mark survives editing. There are no hidden characters and nothing is added to the text. The watermark lives in which words the model picks, and in OpenAI's test, swapping a quarter of the words for synonyms cut detection from about 92% to 17%.
You'll see how a language model picks each word, how a secret key turns that random choice into a pattern only the key holder can check, how OpenAI's textGrain and the SynthID-Text method behind Claude's watermark work, why short passages and math are hard to detect, what editing does to the signal, and what a positive or negative result can and can't tell you.
About the evidence: the detection rates (about 80% at 200 tokens and 95% at 400 tokens at a 1% false-positive rate, and 92% → 66% → 17% after 10% and 25% synonym swaps on 400-token passages) and the benchmark comparison are OpenAI's own tests. Anthropic's statements about editing, cost and what a match proves are Anthropic's own. The group-and-key diagrams are simplified illustrations of OpenAI's technical report. "200 tokens is roughly 150 words" is our rough conversion. We did not run either detector; access is limited to approved organizations. Sources checked October 6, 2026.
Chapters
00:00 The watermark you can't see
01:01 How a model picks each word
03:17 Who ships it, and where
04:28 Why length matters
05:11 What editing does
05:51 What a match proves
06:50 A signal, not proof
Sources
OpenAI, "Our approach to EU text provenance rules", October 5, 2026: https://openai.com/index/eu-text-prov...
OpenAI, "textGrain: Entropy-Calibrated Watermarking for Language Model Text" (technical report), October 5, 2026: https://cdn.openai.com/pdf/e9508624-d...
Anthropic, "How Claude's text watermark works", August 14, 2026 (updated September 1): https://www.anthropic.com/news/claude...
Dathathri et al., "Scalable watermarking for identifying large language model outputs", Nature 634, 818–823 (2024): https://doi.org/10.1038/s41586-024-08...
European Commission, Code of Practice on AI-generated content: https://digital-strategy.ec.europa.eu...