Our approach to EU text provenance rules
Published October 5, 2026
Not law. This is a company's own public position on AI regulation. It is not law, and it carries no legal force.
What it argues for
OpenAI’s web article, dated October 5, 2026, explains its response to the EU AI Act’s machine-readable marking requirement for generated text. Its central argument is that provenance should promote transparency without treating detection as conclusive evidence of authorship, accuracy or responsibility. The rollout separates consumer products from developer access: eligible ChatGPT and Codex text outputs will receive invisible watermarks across EU plans over the coming weeks, while API customers worldwide can opt in for select models, with watermarking "off by default in the API". Detector access initially goes only to approved researchers and expert organizations; OpenAI says it is "not making it publicly available at launch". Its textGrain technology inserts an "invisible statistical signal" into word choices, and OpenAI plans an open-source release. Evaluations show why the company limits access: at a 1% target false-positive rate, detection reached about 80% for 200-token psychology passages and about 95% for 400-token passages, with substantially poorer results for mathematics. In a separate editing evaluation, replacing 25% of words with synonyms reduced detection from about 92% to 17%. OpenAI reports no meaningful benchmark performance differences from watermarking its Astra model. It stresses that "A watermark does not measure human contribution" and that "The absence of a detected watermark does not prove human authorship". Watermarks also do not identify users or establish ownership, lawfulness or truth. The article places text marking within a layered provenance strategy using Content Credentials, image and audio watermarks, public verification tools and abuse enforcement, and commits to revisiting deployment and access as evidence and standards develop.
Stated positions (12)
- Text watermarking will initially be a regional default rather than a worldwide one: "We are not making text watermarking a global default at launch." Eligible ChatGPT and Codex users across EU plans are the announced rollout population.
- API customers decide whether to enable watermarking for supported models: "Text watermarking will remain off by default in the API."
- Text detection will initially be restricted to specialist applicants: "Access will initially be limited to approved researchers and expert organizations that can help us evaluate and improve the technology."
- The technology marks word choices rather than attaching a visible label: "textGrain, adds an invisible statistical signal to the model’s word choices."
- Strong evaluation results do not establish dependable real-world detection: "strong performance under ideal conditions does not guarantee reliable detection in everyday use."
- Editing can sharply reduce detection: "Replacing 25% of words reduced it to 17%."
- OpenAI reports no meaningful performance cost on its Astra benchmarks: "we do not see meaningful performance differences with and without watermarking."
- A watermark cannot quantify human involvement or settle legal responsibility: "A watermark does not measure human contribution." Separately, "A watermark does not establish ownership or responsibility."
- Detection must not be treated as an identity or truth check: "A watermark does not identify the user." Also, "A watermark does not verify accuracy."
- Failure to detect a watermark is not evidence of exclusively human authorship: "The absence of a detected watermark does not prove human authorship."
- Provenance should combine complementary methods rather than rely on one signal: "No single provenance technique is enough on its own, so we take a layered approach that combines open standards, durable watermarking, and verification tools."
- The technology and access policy remain subject to revision: "We also plan to make the technology available in open source so that others can build on it", and OpenAI will "expand detector access when we believe results can be interpreted responsibly."
About this document
A web article published by OpenAI, presented under Safety and dated October 5, 2026. It uses headed sections to explain the rollout, detection performance, output quality, interpretive limits and broader provenance strategy, with descriptions of two charts and a benchmark comparison table.
How this sits against AI law
Each stance compared with what EU and US instruments actually require. Where no instrument addresses a theme, that gap is shown rather than hidden.
EU rollout and optional API marking
OpenAI will watermark eligible ChatGPT and Codex text outputs across EU plans, while leaving API watermarking off by default and allowing customers worldwide to opt in for select models.
Article 50(2) requires providers to ensure that covered synthetic outputs are machine-readable marked and detectable as artificially generated or manipulated. For EU API outputs within that scope, customer opt-in alone is weaker than the provider duty; the provision allows specified exceptions and technical qualifications, but not a general customer-choice exemption.
The California AI Transparency Act’s output-marking duties concern covered image, video and audio content under Business and Professions Code Section 22757.3. It does not impose a corresponding generated-text watermarking duty.
Restricted access to the text detector
Initial detector access will be granted case by case to approved researchers and expert organizations, rather than opened to the public, because false positives and missed watermarks create interpretation risks.
Article 50(2) requires machine-readable marking and detectability, but does not expressly require every provider to publish an unrestricted public detector. Specialist-only access is therefore compatible with that provision, although access restrictions do not remove the underlying marking and detectability duties.
Business and Professions Code Section 22757.2 requires covered providers to offer a free public detection tool for their image, video and audio outputs. It does not establish an equivalent access rule for a text watermark detector.
Technical limits and detection reliability
OpenAI treats watermarking as an imperfect signal: shorter passages and constrained subjects are harder to detect, and editing can substantially weaken the watermark. It will study reliability, editing and translation rather than treat ideal-condition performance as sufficient.
Article 50(2) requires effective, interoperable, robust and reliable technical solutions as far as technically feasible, taking account of content-specific limitations, implementation costs and the generally acknowledged state of the art. OpenAI’s emphasis on testing and technical limitations matches that qualified standard; its evaluation results alone do not establish legal compliance.
Sections 22757.2 and 22757.3 address detection and provenance for covered image, video and audio outputs. They do not set a reliability standard for generated-text watermarks or their resistance to rewriting.
Human contribution and authorship
A watermark can indicate that an OpenAI system generated or processed part of a passage, but cannot measure human judgment, editing or creativity. Failure to detect a watermark does not prove human authorship.
Article 50(2) governs marking and detectability, while Article 50(4) separately addresses disclosure of certain public-interest text and includes an exception involving human review or editorial control and editorial responsibility. Neither provision makes a watermark result a measure of human contribution or establishes a presumption of human authorship when detection fails.
The California Act’s detection and provenance provisions do not establish how much human authorship a text contains or make an absent text watermark proof of human authorship. Its principal output-transparency duties concern image, video and audio.
No identity, ownership or truth inference
The detector will not identify users or reveal prompts or conversations. OpenAI also rejects using a watermark to determine ownership, lawful use, disclosure duties, responsibility or factual accuracy.
Article 50(2) requires a signal of artificial generation or manipulation, not identification of the user, prompt or conversation. It does not make that signal a determination of ownership, lawfulness, responsibility or truth; Article 50(4) sets separate disclosure duties rather than deriving them from a detector result.
The California Act regulates detection tools and provenance disclosures for covered media. Sections 22757.2 and 22757.3 do not give text watermark results legal effect as determinations of ownership, responsibility or factual accuracy.
Layered provenance across media
OpenAI combines Content Credentials, invisible watermarks and verification tools, with public image and audio verification remaining available. It pairs provenance signals with policies, abuse detection, reporting, investigations and enforcement.
Article 50(2) sets an output-marking and detectability obligation across synthetic audio, image, video and text, but does not prescribe OpenAI’s combination of Content Credentials, watermarks, public verification services and abuse-enforcement processes. The proposed operational package goes beyond the specific provenance mechanisms that provision mandates.
Business and Professions Code Sections 22757.2 and 22757.3 combine public detection access with provenance disclosures for covered image, video and audio outputs. OpenAI’s public image and audio tools and embedded provenance match that layered structure, though this does not establish compliance with every statutory requirement.
Open-source technology and continuing revision
OpenAI plans to release textGrain in open source and update its technical report. It will revisit rollout and detector access as technology, evidence, standards and regulatory requirements evolve.
Article 50(2) ties technical solutions to feasibility and the state of the art, but does not require providers to release their watermarking technology as open source. That planned release goes beyond the provision’s marking and detectability duty.
The California Act requires covered detection and provenance functionality, but Sections 22757.2 and 22757.3 do not require an open-source release of watermarking technology. OpenAI’s planned publication adds a form of technical access beyond those duties.
Article 50(2) of the EU AI Act places the machine-readable marking and detectability duty on providers; it does not make compliance depend on customers opting in. OpenAI’s EU consumer rollout follows that approach, but its API default is weaker for EU outputs within the provision’s scope unless another compliant marking method applies. By contrast, Article 50(2) does not expressly require a publicly accessible detector, so restricted detector access is not itself a conflict with that provision. California’s AI Transparency Act requires a public detection tool for covered image, video and audio outputs, not an equivalent tool for generated text.
Source
https://openai.com/index/eu-text-provenance/- Date on the page:
- "October 5, 2026" printed in the article header, category Safety.
- Source checked:
- opened and confirmed on 2026-10-30