AI Content Labeling Laws in 2026: EU, US, China Requirements Compared

AI Content Labeling Laws in 2026: EU, US, China Requirements Compared

If your company publishes AI-generated text, synthetic images, cloned voices, or deepfake video in 2026, the legal question is no longer only “who owns the copyright?” It is also “did we tell people what this is?”

That shift matters because labeling duties do not map neatly onto copyright ownership. A marketing image can be unregistrable as a mostly machine-generated work under U.S. Copyright Office policy and still trigger disclosure duties under platform rules, consumer-protection law, election law, or AI-specific regulation. A cloned voice may raise right-of-publicity and privacy claims even when no copyrighted sound recording was copied. A model provider may disclose training-data summaries under one law while the deployer must label output under another.

This guide compares the three disclosure regimes that matter most for global AI copyright compliance in 2026: the European Union’s AI Act, the United States’ patchwork of federal agency enforcement and state laws, and China’s rules for deep synthesis and generative AI services. It is written for product, legal, trust-and-safety, and marketing teams that need an operational policy, not a generic explainer.

Related reading: if your immediate problem is ownership rather than labeling, start with our guide to copyrighting AI-assisted content. If you are building a broader governance program, pair this with the AI copyright compliance checklist and the AI output clearance workflow for marketing teams.

The short answer

In 2026, the safest default is to label AI-generated or materially AI-altered content whenever a reasonable viewer, listener, reader, customer, voter, or business partner could be misled about whether a human created it, whether a real person said or did something, or whether the work came from a licensed human source.

That does not mean every spellcheck, crop, background removal, or internal brainstorm needs a public warning. It does mean companies should stop treating disclosure as a PR afterthought. Disclosure is becoming a compliance control.

A practical baseline policy looks like this:

1. Label synthetic or materially altered images, video, and audio at the point of publication.

2. Use stronger labeling for deepfakes, voice clones, public figures, election content, minors, medical/legal/financial advice, and advertising.

3. Keep internal records showing the tool used, prompt or instruction summary, human edits, rights clearance, and final approval.

4. Contractually require vendors to disclose AI-generated assets and preserve provenance metadata where feasible.

5. Do not rely on invisible metadata alone; use visible or audible notices when deception risk is material.

The reason is simple: copyright law, consumer-protection law, and AI regulation are converging around provenance. The law may not always say “copyright disclosure,” but in practice the same evidence proves both compliance and authorship.

Why labeling has become a copyright issue

For years, copyright teams focused on two questions: whether AI training infringes input works and whether AI outputs are protectable. Those questions remain live. The U.S. Copyright Office’s January 29, 2025 report on copyrightability, for example, reaffirmed that copyright protects human authorship and not purely machine-generated expression. Courts are also still testing training-data theories in cases such as The New York Times Co. v. Microsoft Corp. and OpenAI, Inc., filed December 27, 2023 in the Southern District of New York, and Andersen v. Stability AI Ltd., filed January 13, 2023 in the Northern District of California.

But labeling has become the bridge between those legal questions and day-to-day publishing. A company that labels output well usually has a workflow that can answer: who prompted it, who edited it, what source material was used, what license covers it, and what human contribution exists? Those are the same facts needed for registration, fair-use analysis, indemnity claims, and takedown disputes.

The opposite is also true. If a company cannot say which public-facing assets were generated by which model, it probably cannot defend authorship, originality, or clearance with confidence. That is why labeling belongs in the same operating system as the training-data audit trail and vendor copyright indemnity checklist.

European Union: the AI Act turns transparency into a product requirement

The EU AI Act is the most important AI-specific disclosure law in force. The final regulation was approved by the European Parliament on March 13, 2024 and adopted by the Council on May 21, 2024. It entered into force on August 1, 2024, with obligations phased in over time. For copyright teams, two areas matter most: transparency duties for AI-generated or manipulated content and copyright-related duties for general-purpose AI model providers.

Article 50 is the core output-transparency provision. It requires certain AI systems to inform users that they are interacting with AI, unless this is obvious from the context. It also requires providers of AI systems that generate synthetic audio, image, video, or text content to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, where technically feasible. Deployers that publish deepfake content generally must disclose that the content has been artificially generated or manipulated, subject to limited exceptions such as law enforcement and some artistic, satirical, or creative contexts.

Article 53 adds a separate obligation for providers of general-purpose AI models. They must draw up and make publicly available a sufficiently detailed summary of the content used for training, according to a template from the AI Office, and put in place a policy to comply with EU copyright law. This is not the same as output labeling, but it matters to copyright clearance. If your company licenses, fine-tunes, or deploys a general-purpose model, Article 53-style documentation may become part of procurement diligence.

For deeper EU-specific analysis, see our breakdown of the EU AI Act copyright transparency requirements.

What EU compliance should look like in practice

For most companies, EU compliance should include four layers.

First, classify the content. Is it text, image, video, audio, code, or mixed media? Was AI used only to assist a human, or did it generate expression that appears in the final asset? Does it depict a real person, imitate a real voice, or present a factual event?

Second, classify the audience risk. A synthetic product mockup used inside a design review is low risk. A synthetic testimonial, realistic news image, political ad, educational medical video, or cloned executive voice is high risk.

Third, decide the disclosure format. Machine-readable provenance metadata is useful, especially where industry standards such as C2PA are supported. But if a reasonable audience could be misled, visible or audible disclosure should appear near the content itself. “AI-generated illustration,” “Synthetic voice recreation,” or “This video includes AI-generated scenes” is usually better than vague language like “enhanced with technology.”

Fourth, preserve evidence. Keep the final prompt or instruction summary, the model or vendor used, the human editor, approval date, license review, and disclosure text. This record will be valuable if a platform removes the asset, a claimant sends a takedown notice, or a regulator asks how the content was produced.

United States: no single AI labeling law, but many ways to get in trouble

The United States does not have one federal AI Act equivalent. Instead, AI labeling obligations come from a patchwork: Federal Trade Commission authority, Copyright Office registration policy, state deepfake and election laws, privacy and biometric statutes, platform rules, advertising law, and sector-specific regulators.

That patchwork can be more dangerous than a single statute because teams often assume “no federal AI labeling law” means “no labeling duty.” That is wrong.

The FTC has repeatedly warned companies not to mislead consumers about AI. Its authority under Section 5 of the FTC Act covers unfair or deceptive acts or practices. If an AI-generated endorsement, review, testimonial, product demo, investment claim, or health claim gives consumers a false impression, disclosure may be necessary even without an AI-specific statute. The label is not magic; it must be clear enough to prevent deception.

The Copyright Office adds a different disclosure problem. In March 2023, it issued registration guidance requiring applicants to disclose and disclaim AI-generated material that is more than de minimis. On February 21, 2023, the Office partially cancelled the registration for Zarya of the Dawn, concluding that the human-authored text and selection/arrangement could be protected, but the Midjourney-generated images were not. That decision is not an advertising-labeling rule, but it teaches the same lesson: document what the human did and what the machine produced.

Courts have reinforced that copyright protection still turns on human authorship. In Thaler v. Perlmutter, decided August 18, 2023, the U.S. District Court for the District of Columbia upheld the Copyright Office’s refusal to register an image generated autonomously by the “Creativity Machine.” The D.C. Circuit affirmed on March 18, 2025, emphasizing the statutory requirement of a human author. For a business, the practical takeaway is that AI-use records are not optional if you expect to claim copyright in a final work.

State laws create additional traps. Many states have enacted or proposed rules on election deepfakes, synthetic sexual imagery, voice and likeness misuse, and biometric identifiers. These laws vary, but they share a theme: when synthetic media makes it look or sound like a real person said or did something, disclosure and consent become critical. Illinois’ Biometric Information Privacy Act, while not an AI labeling statute, has also become relevant in voiceprint and faceprint disputes because AI systems can process identifiers that privacy law treats as sensitive.

U.S. labeling standard for businesses

Because the U.S. is fragmented, the best standard is not “what is the minimum label in one state?” The better standard is: would a reasonable consumer, viewer, listener, employee, voter, creator, or licensee be deceived or materially misled without disclosure?

Use a visible or audible label when AI creates or materially changes:

• testimonials, endorsements, reviews, or case studies;
• realistic people, products, events, locations, screenshots, or evidence;
• voices, performances, music, or avatars resembling real people;
• legal, financial, health, safety, or employment guidance;
• political, issue-advocacy, or public-interest content;
• assets you intend to license, sell, or register as copyrighted works.

Use internal-only documentation when AI merely supports drafting, grammar, ideation, resizing, formatting, translation review, or non-expressive workflow steps and the final public content is clearly human-controlled.

China: explicit labeling for deep synthesis and generative AI

China has moved faster than the United States on explicit AI-content labeling. The key rules are the Provisions on the Administration of Deep Synthesis of Internet Information Services, which took effect on January 10, 2023, and the Interim Measures for the Management of Generative Artificial Intelligence Services, which took effect on August 15, 2023.

The Deep Synthesis Provisions require providers and technical supporters to add conspicuous labels to generated or edited information where the service may cause public confusion or misidentification. They specifically address technologies that generate or edit text, voice, images, video, virtual scenes, and other information.

The 2023 Generative AI Interim Measures add broader governance duties for generative AI services offered to the public in China. Providers must take steps around lawful data sources, respect for intellectual property rights, personal information protection, accuracy, and labeling of generated content in accordance with the deep synthesis rules.

For copyright teams, China’s approach matters for two reasons. First, labeling is tied to platform and provider responsibility, not just publisher ethics. Second, IP compliance is part of generative AI governance. A company offering AI-generated content services into China should assume that labeling, training-data rights, personality rights, and content moderation are reviewed together.

A practical label taxonomy

The biggest mistake companies make is writing one universal label and using it everywhere. “Made with AI” is sometimes too much, sometimes too little, and often too vague. A better system uses categories.

1. AI-assisted, human-authored

Use this when a human created the expressive work and AI helped with editing, research organization, formatting, brainstorming, translation suggestions, or minor non-substantive changes.

Suggested public label when needed: “Created by [team/author], with AI assistance for editing/research.”

Usually no public label is needed for ordinary business writing unless the audience expects a fully human process or the context is sensitive. Keep internal notes anyway.

2. AI-generated illustration or media

Use this when AI generated a visual, audio, or video asset that does not depict a real event or real person in a misleading way.

Suggested label: “AI-generated illustration” or “Synthetic background image.”

This is common for blog hero images, social graphics, ads, and moodboards. If you want copyright protection, record the human creative choices: prompt design, selection, editing, compositing, arrangement, and final modifications.

3. Materially altered real content

Use this when AI changes a real photo, recording, performance, screenshot, document, or event in a way that affects meaning.

Suggested label: “Image materially altered using AI” or “Audio enhanced and partially reconstructed using AI.”

This matters because the audience may rely on the content as evidence. A simple color correction is not the same as adding people, removing objects, changing spoken words, or reconstructing missing scenes.

4. Synthetic person, likeness, or voice

Use this when the content depicts or imitates a real person, public figure, employee, customer, performer, or private individual.

Suggested label: “Synthetic voice recreation used with permission” or “AI-generated likeness; not real footage.”

This is the category with the highest legal risk. Copyright may be only one issue. Right of publicity, privacy, biometric law, contract, labor agreements, and platform policies may all apply. Our analysis of YouTube’s AI likeness detection rollout explains why creator-facing platforms are treating this as a rights-management problem, not merely a moderation feature.

5. AI-generated legal, financial, health, or safety information

Use this when AI produces guidance that people may rely on for important decisions.

Suggested label: “AI-assisted informational content reviewed by [qualified role/date], not legal advice.”

A label alone is not enough. Sensitive advice requires human review, source checking, jurisdiction limits, and escalation paths.

Contract language: make labeling someone’s job

Many labeling failures happen because no one owns the duty. Vendor and agency agreements should require disclosure of AI-generated deliverables, identification of tools used, rights warranties for inputs, preservation of provenance metadata where reasonable, human review, limits on confidential data in training-retained tools, and cooperation with takedowns or platform appeals. Do not bury this in a generic “comply with law” clause; say who must disclose what, when, and in what format.

The copyright registration angle

Labeling and copyright registration are separate, but they should share records.

If a company wants to register an AI-assisted work in the United States, it must identify the human-authored elements and exclude more-than-de-minimis AI-generated material. The Copyright Office’s 2023 guidance and later decisions make this unavoidable. A public label will not decide registrability, but the internal labeling record may support the application.

For example, imagine a company publishes a 40-page white paper. AI was used to generate a first outline, summarize cases, and propose charts. Human lawyers wrote the analysis, selected cases, edited every section, and created the final structure. That may be registrable as human-authored text and compilation if documented properly. The public page might say nothing, or it might say “AI-assisted research tools were used; final analysis reviewed by counsel.” The internal record should be more detailed.

Now imagine the same company publishes 30 AI-generated illustrations with minimal human edits. Those images may need public labels under platform or AI transparency rules, but the copyright registration should disclaim the AI-generated image content unless there is enough human authorship in selection, arrangement, or modification. See our guide to proving human authorship in AI-assisted works for the evidence side of that problem.

A 2026 operating checklist

Use this before publishing any AI-involved content.

1. Identify the asset type: text, image, audio, video, code, dataset, or mixed media.

2. Identify the AI role: ideation, editing, generation, transformation, translation, cloning, or personalization.

3. Identify jurisdiction exposure: EU users, U.S. consumers, China users, children, voters, employees, regulated industries.

4. Identify deception risk: could someone think this is real, human-made, licensed, endorsed, or said by a real person?

5. Identify rights risk: copyrighted inputs, living artists’ styles, real voices, trademarks, confidential data, third-party datasets.

6. Choose disclosure level: none, internal record only, metadata, visible label, audible label, repeated in-product disclosure.

7. Preserve provenance: tool, model, date, user, prompt summary, source files, licenses, edits, approver.

8. Review contracts: vendor disclosure, indemnity, data retention, training opt-out, rights warranties.

9. Review platform rules: YouTube, TikTok, Meta, app stores, ad networks, stock marketplaces, and client channels.

10. Reassess after edits: if a human substantially transforms AI output, update records; if AI materially changes human content, update labels.

Recommended default policy

For most organizations, the best 2026 policy is:

• No public label required for ordinary AI-assisted drafting or editing where a human controls the final expression and no reasonable audience would be misled.
• Visible label required for AI-generated images, videos, audio, avatars, synthetic testimonials, or realistic scenes used externally.
• Strong visible or audible label plus written consent required for real-person likenesses, voice clones, political content, employee/customer depictions, and regulated advice.
• Internal record required for every public AI-assisted asset, even when no public label is used.
• Legal review required before registering copyright, licensing AI-generated assets, using third-party copyrighted inputs, or publishing synthetic media that could affect someone’s reputation, rights, or economic interests.

This policy is stricter than the bare minimum in some jurisdictions and more flexible than a blanket “label everything” rule. That is intentional. Over-labeling every grammar edit trains audiences to ignore disclosures. Under-labeling realistic synthetic media creates legal and trust risk. The right approach is risk-based and evidence-backed.

Bottom line

AI labeling law in 2026 is not one rule. The EU AI Act pushes transparency into product design. U.S. law uses deception, registration, publicity, privacy, and platform enforcement to reach similar problems from different angles. China requires conspicuous labeling for deep synthesis and connects generative AI governance to IP compliance.

The winning move is to build one global workflow: classify AI involvement, assess deception and rights risk, choose the right disclosure, preserve provenance, and make vendors contractually responsible for telling you when AI was used.

That workflow will not just reduce regulatory risk. It will make your copyright claims stronger, your takedown responses faster, and your AI governance far less chaotic.