Most of what gets called an “AI disclosure statement” on this site so far is a publisher or journal editorial requirement — a condition of manuscript submission set by COPE, ICMJE, or an individual publisher, not a government mandate. This guide covers the other, less-documented half: what actual law currently requires around disclosing AI-generated content, and how that differs from what a journal’s submission policy asks for. As of mid-2026 the two most concrete legal disclosure regimes that can touch research and its outputs are the EU AI Act’s Article 50 transparency obligations and a fast-moving, still-unsettled patchwork of US state AI-transparency laws. Neither is a research-specific law — both are general-purpose transparency statutes that happen to reach content a researcher, publisher, or institution might produce or publish. Verified against EUR-Lex, the European Commission’s AI Act Service Desk, and current US legal-tracker reporting as of 2026-07-18. Because AI regulation is moving fast and jurisdiction by jurisdiction, treat every date and requirement below as a snapshot rather than a permanent rule, confirm against the linked primary source before relying on it for a compliance decision, and don’t read this guide as legal advice.
Three distinct layers of “AI disclosure” — don’t conflate them
A researcher, editorial office, or compliance officer can be looking at three genuinely different kinds of “AI disclosure” obligation at once, and confusing them is the most common practical error:
| Layer | Who sets it | Is it legally binding? | Example |
|---|---|---|---|
| Government/legal transparency law | National or state legislatures | Yes — statutory, enforceable by a regulator | EU AI Act Article 50; California SB 942; Colorado SB 189 |
| Research-funder policy | NIH, NSF, ERC, and similar bodies | A condition of the funding agreement, not a general statute — enforced through grant administration, not a regulator | NIH’s NOT-OD-25-132 originality standard; NSF’s encouraged-not-required disclosure notice |
| Publisher/journal editorial policy | COPE, ICMJE, individual publishers | A contractual condition of submission, not law | A manuscript’s generative-AI disclosure statement |
This guide is about the first row only. For the funder layer, see CASRAI’s NIH vs NSF vs ERC comparison. For the publisher layer, see Journal and Publisher Policies on Generative AI in Manuscripts and the Generative-AI disclosure statement dictionary entry.
The EU AI Act’s Article 50: the actual legal transparency requirements
Article 50 of Regulation (EU) 2024/1689 (the AI Act) is the provision that actually creates a legal AI-content-disclosure duty, as distinct from the research-exemption question CASRAI’s EU AI Act obligations and exemptions guide covers in depth. Article 50 sets out four separate obligations, not one general “disclose AI use” rule:
- Article 50(1) — a provider of an AI system intended to interact directly with people must ensure those people are informed they’re interacting with AI, unless that’s already obvious to a reasonably well-informed person.
- Article 50(2) — a provider of a system generating synthetic audio, image, video, or text content must mark that output in a machine-readable format, detectable as artificially generated or manipulated.
- Article 50(3) — a deployer of an emotion-recognition or biometric-categorization system must inform the people exposed to it.
- Article 50(4) — a deployer of a system producing a deepfake, or AI-generated/manipulated text published to inform the public on matters of public interest, must disclose that the content is artificially generated. Two exemptions matter here: content that is evidently part of an artistic, creative, satirical, or fictional work (a lighter, non-intrusive disclosure suffices), and — more relevant to scholarly text — content that has undergone a process of human editorial review, where a natural or legal person holds editorial responsibility for its publication.
Article 50’s general compliance date, per Article 113 of the Regulation, is 2 August 2026. As CASRAI’s EU AI Act guide details, a separate legislative package (the “Digital Omnibus,” part of the EU’s “Omnibus VII” simplification effort) reached political agreement in May-June 2026 to defer the Article 50(2) synthetic-content-marking obligation specifically to 2 December 2026 — that deferral had cleared Parliament and Council but had not been independently confirmed as published in the Official Journal as of that guide’s most recent verification. Article 50(1), (3), and (4) were not reported as part of that specific deferral. Check the AI Act Service Desk’s Article 50 page directly before relying on either date.
Does Article 50 add a new disclosure duty on top of a journal’s own policy?
This is the question a research administration or editorial office actually needs answered, and it doesn’t have a Commission-issued answer yet. Working through the text: Article 50’s obligations attach to providers (the entity that builds and puts an AI system into service) and deployers (the entity using that system under its own authority), not directly to “an author whose manuscript contains AI-assisted text.” A researcher using a general-purpose chatbot to help draft a paragraph is, in the Act’s terms, a deployer of that tool for that use — and Article 50(4)’s human-editorial-review exemption is squarely aimed at exactly this situation: text that has gone through human editorial review, with a person or organization holding editorial responsibility for publishing it, is not subject to the deepfake/public-interest-text disclosure duty in the same way unreviewed synthetic content is. A peer-reviewed journal article, by definition, has gone through exactly that kind of human editorial review.
Read plainly, that suggests the Act’s Article 50(4) duty is unlikely to impose an additional disclosure requirement on top of what a journal’s own AI-disclosure policy already asks for, for ordinary AI-assisted manuscript text that goes through normal peer review and editing. That reading is CASRAI’s own analysis of the statutory text, not confirmed Commission guidance specific to scholarly publishing — the Commission has separately signalled (see the EU AI Act guide’s coverage of the Article 2(6)/2(8) research-exemption boundary) that further sector-specific implementation guidance is still coming, and scholarly publishing has not been named as a priority area the way pre-clinical/medical-device research has. An institution or publisher with an active compliance question here should get its own legal reading rather than rely on this guide’s inference.
US state AI-transparency laws: a fragmented, fast-moving patchwork
There is no single comprehensive US federal law requiring disclosure of AI-generated content as of this writing; AI-transparency legislation in the US has moved almost entirely at the state level, and unevenly. Two of the more developed examples, current as of mid-2026:
- California SB 942 (the California AI Transparency Act) requires providers of generative-AI systems with more than one million monthly users to offer a free AI-detection tool and to support both “manifest” (visible) and “latent” (embedded, machine-readable) disclosure of AI-generated image, video, and audio content. Its original 1 January 2026 effective date was pushed to 2 August 2026 by a follow-on bill, AB 853.
- Colorado replaced its original Colorado AI Act with SB 189, passed by the state legislature in May 2026, which drops the earlier law’s duty-of-care/risk-assessment framework in favor of a narrower, disclosure-based approach, taking effect 1 January 2027.
Both laws are general-purpose consumer/platform-transparency statutes aimed at large-scale generative-AI providers — neither targets research, academic publishing, or manuscripts specifically, and neither is likely to apply directly to a researcher or a journal in the way a journal’s own AI-disclosure policy does. They matter to a research organization mainly at one remove: if the institution or a vendor it uses (an AI writing tool, an image-generation tool embedded in a communications platform) meets the scale threshold to be a “provider” under one of these laws, that vendor — not the individual researcher — carries the compliance burden. This is a genuinely unsettled, jurisdiction-by-jurisdiction area with more state bills proposed than enacted; this guide is not an exhaustive 50-state survey, and a compliance question tied to a specific state should go to institutional counsel or a maintained legal tracker rather than be resolved from this page alone.
What this means practically for a research administration office
- Keep the three layers in the table above separate in any policy document: satisfying a journal’s disclosure statement requirement does not automatically satisfy an EU AI Act or state-law obligation, and vice versa — they’re enforced by different parties for different reasons.
- The EU AI Act’s Article 50 duties fall on providers and deployers of AI systems, not on authors as such — the institution’s exposure is largest where it operates or deploys a generative-AI system with public-facing output (a communications office publishing AI-assisted content, an AI tool embedded in a public-facing service), not in the ordinary case of a researcher using a chatbot to help draft a reviewed manuscript.
- US state AI-transparency laws generally reach large-scale AI providers, not individual researchers or most universities as such — the more relevant question for most institutions is whether a vendor’s tool is itself covered, not whether the institution needs its own disclosure statement.
- None of this changes what a journal’s own AI-disclosure policy requires at submission — see CASRAI’s publisher-policy landscape guide for that separate, already-operative requirement.
- Track the EU AI Act’s Article 50 dates and any US state developments through the primary sources linked throughout this guide rather than a cached summary — this is one of the fastest-moving areas of research compliance right now.
For the broader compliance picture, see CASRAI’s Research Integrity & Compliance cluster hub and the Scholarly Publishing cluster hub, plus related Dictionary entries on AI-generated content, watermarking (AI output), and AI tool disclosure.
Frequently asked questions
Does the EU AI Act require me to disclose AI use in a published research paper?
Not as a rule separate from what applies to any AI-assisted text. Article 50(4)’s deepfake/public-interest-text disclosure duty exempts content that has undergone human editorial review with a person or organization holding editorial responsibility for its publication — which ordinary peer-reviewed manuscript text satisfies. This is CASRAI’s reading of the statutory text, not confirmed Commission guidance specific to scholarly publishing.
What’s the difference between a journal’s AI disclosure policy and a legal AI disclosure requirement?
A journal’s policy (e.g., a required generative-AI disclosure statement) is a contractual condition of submission set by the publisher, enforced by the editorial office. A legal requirement, like the EU AI Act’s Article 50, is a statutory obligation enforced by a government regulator and applies regardless of whether a paper is ever submitted anywhere.
Do US state AI-transparency laws like California’s SB 942 apply to academic publications?
Not directly. These laws generally regulate providers of large-scale generative-AI systems (over a stated monthly-user threshold), not individual authors or most research institutions publishing a paper. They matter to a research organization mainly if it, or a vendor it uses, meets the threshold to be a covered “provider” in its own right.
Is there a US federal law requiring disclosure of AI-generated content?
Not a comprehensive one as of this writing. US AI-transparency regulation has developed almost entirely at the state level, unevenly and still changing — California and Colorado are among the more developed examples, but this is not an exhaustive survey and state positions are shifting quickly.
Does NIH’s or NSF’s AI policy count as a legal AI-disclosure requirement?
No — funder policies are conditions of the funding agreement, enforced through grant administration (cost disallowance, award suspension, misconduct referral), not general statutes enforced by a regulator. See CASRAI’s NIH vs NSF vs ERC comparison for how those specific funder policies work.
