---
title: "A text watermark is provenance evidence, not proof of authorship"
description: "OpenAI's EU text watermark can survive copying but degrades under editing. Enterprises should treat detection as one provenance signal, never as an authorship, identity or accuracy verdict."
url: "https://devencelab.com/insights/2026/10/05/a-text-watermark-is-provenance-evidence-not-proof-of"
date: "2026-10-05"
section: "Insights"
tag: "AI Regulation"
author: "Devence Lab"
reading_time: "2 min read"
site: "Devence Lab"
license: "Readable and quotable with attribution to the canonical URL."
---

# A text watermark is provenance evidence, not proof of authorship

OpenAI's EU text watermark can survive copying but degrades under editing. Enterprises should treat detection as one provenance signal, never as an authorship, identity or accuracy verdict.

OpenAI said on 5 October that eligible ChatGPT and Codex text in the European Union will gain an invisible watermark over the coming weeks. Its textGrain detector identified about 95% of 400-token psychology passages at a target 1% false-positive rate in company testing. That figure sounds like an authorship detector. It is not.

The operational value is narrower: textGrain adds a statistical signal to model word choices so copied text can retain provenance information after metadata disappears. OpenAI explicitly says the signal does not identify a user, establish ownership, measure human contribution or verify whether the text is accurate.

## Editing turns detection into probabilistic evidence

OpenAI's own evaluation shows why policy must account for uncertainty. For 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. Shorter passages and constrained domains such as mathematics were also harder to detect.

> Devence Lab assessment: provenance signals belong in an evidence chain, not at the end of a disciplinary decision.

That makes a negative result especially weak. An undetected passage may have been edited, translated, generated by an unsupported model, created before watermarking or produced by another provider. A positive result is also bounded: it indicates a statistical signal associated with generation or processing, not who submitted the text or how much human judgement shaped it.

## Enterprise policy needs separate questions

Security, compliance and education teams should separate four decisions that are easy to collapse. Provenance asks whether a technical signal is present. Identity asks who performed an action. Authorship asks who contributed the substantive work. Accuracy asks whether the content is correct. textGrain directly addresses only the first.

OpenAI is reflecting that limitation in access. API customers can opt in globally for select models starting 5 October, while eligible ChatGPT and Codex users in the EU will receive watermarking in a phased rollout. Detector access initially goes to approved researchers and expert organisations rather than becoming a public accusation tool.

## Design controls for corroboration, not certainty

Organisations adopting AI-content detection should record detector version, threshold, text length and transformation history with every result. High-impact decisions should require corroborating evidence such as document history, account activity, declared workflow and review records. A detector score alone should not trigger fraud, misconduct or access decisions.

The useful control is therefore not 'detect AI and decide'. It is 'collect provenance signals and investigate proportionately'. The EU rollout makes machine-readable provenance more available, but OpenAI's published limitations make the governance boundary equally clear: detection evidence is informative precisely because it is not a verdict.

Produced through Devence Lab's automated research and publishing workflow, without prepublication human review. Devence Lab retains editorial responsibility. Sources are linked; report factual errors to contact@devencelab.com.

## Sources

- [Our approach to EU text provenance rules](https://openai.com/index/eu-text-provenance/) - OpenAI
- [OpenAI will start watermarking ChatGPT’s text in the EU](https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/) - TechCrunch