A year ago, labelling AI-generated content was a matter of good manners. Today it is a legal requirement in several of the world’s largest markets. The EU, India, China and South Korea all now expect AI-generated images, audio, video and in some cases text to be identifiable as such, and California has added its own disclosure rules.
The rules differ in who they bind and how much detail they demand. But they ask for the same three building blocks: a label people can see, a marker machines can read, and a plan for when labelled content is misused. If you build or deploy generative AI, these are now product features, not policy footnotes.
This guide explains what each major regime requires, where the obligations overlap, and how to design one system that works across markets.
What counts as synthetic content
‘Synthetic content’ means material created or substantially altered by AI. A ‘deepfake’ is the most sensitive kind: AI-generated or manipulated images, audio or video that resemble real people, places or events and could falsely appear authentic.
Most regimes exclude routine edits. Colour correction, noise reduction, accessibility features and other changes that do not alter the substance of the content are generally outside scope. A chatbot that tidies your grammar is treated differently from a tool that puts words in a politician’s mouth.
The practical test is simple. Could a reasonable person mistake this content for something real or human-made, in a way that matters? If yes, assume a label is expected somewhere you operate.
EU: Article 50 of the AI Act
The EU AI Act splits duties between ‘providers’, who develop an AI system or put it on the market under their name, and ‘deployers’, who use it in their own business. Article 50 has applied since 2 August 2026. For generative systems already on the market before then, the machine-readable marking duty applies from 2 December 2026.
Fines for breaching Article 50 can reach €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups the cap is whichever is lower. National market surveillance authorities enforce these rules in each member state.
In June 2026 the Commission published a voluntary Code of Practice on marking and labelling AI-generated content. It is the most detailed practical guide so far. It accepts that no single marking technique currently meets every requirement, which points towards layered solutions.
- Providers of chatbots and similar systems must tell people they are interacting with AI, unless that is obvious.
- Providers of generative systems must mark audio, image, video and text outputs in a machine-readable, detectable way.
- Deployers must disclose deepfakes. For clearly artistic, satirical or fictional work, the disclosure can be light enough not to spoil it.
- Deployers must disclose AI-generated text published to inform the public on matters of public interest, unless a human has reviewed it and someone takes editorial responsibility.
India: the IT Rules amendment
From 20 February 2026, amendments to India’s IT Rules define ‘synthetically generated information’. That means AI-created or AI-altered audio, visual or audio-visual content that appears real. Good-faith editing, accessibility improvements and similar uses are excluded.
The duties fall on ‘intermediaries’, the legal term for platforms and services that host or transmit content for users. Intermediaries that enable the creation or spread of synthetic content must label it clearly and, where technically feasible, embed persistent metadata or provenance markers. They must not enable anyone to modify, hide or remove those markers. ‘Significant social media intermediaries’, those with more than five million registered users in India, must also ask users to declare whether content is synthetic and take steps to verify the answer.
The same amendments cut the deadline for acting on lawful takedown directions from courts or the government to three hours. That changes staffing, not just code. Someone needs to be able to act at 3am.
China, South Korea and the rest of Asia
China moved first. Its Labelling Measures, in force since 1 September 2025, require ‘explicit’ labels that people can see or hear and ‘implicit’ labels embedded in file metadata. They apply to generative AI services offered to the public in China and to platforms that distribute AI content.
South Korea’s AI Basic Act took effect on 22 January 2026. It requires notice to users when a product or service uses generative AI, and labelling of generative AI outputs. Some foreign providers serving Korean users must also appoint a local representative.
Elsewhere in Asia the picture is lighter. Japan and Singapore rely mainly on guidance and voluntary frameworks rather than binding labelling laws. That can change quickly, so check before launching in a new market.
US and UK: rules aimed at harm
The US has no federal labelling law. California’s AI Transparency Act became operative on 2 August 2026. It applies to providers of generative AI systems with over one million monthly users. They must embed hidden ‘latent’ disclosures in AI-generated images, audio and video, offer users the option of a visible disclosure, and provide a free AI detection tool. Duties for large online platforms follow from 2027.
Federal law targets the worst abuse instead. The TAKE IT DOWN Act criminalises publishing non-consensual intimate imagery, including AI deepfakes. From May 2026, covered platforms must remove such content within 48 hours of a valid request.
The UK has no labelling statute either. It relies on the Online Safety Act, enforced by Ofcom, to address illegal content on platforms. Since 6 February 2026 it has also been a criminal offence to create, or ask someone to create, a sexually explicit deepfake of an adult without consent.
One build that works across markets
You do not need a separate system per country. A layered design covers most regimes at once. Design to the strictest market you serve, then switch off what you do not need elsewhere.
For provenance, many companies use the C2PA standard, often branded ‘Content Credentials’. It attaches a signed record of how a file was made. Pair it with an invisible watermark, because metadata is easily stripped when files are screenshotted or re-uploaded.
- Visible labels where people see the content, in the feed, the player and the download.
- Machine-readable metadata using an open provenance standard your partners and platforms can read.
- An invisible watermark that survives compression, cropping and screenshots as far as the technology allows.
- Internal logs of what your system generated, so you can confirm or deny origin when asked.
- A user declaration step where you host content uploaded by others.
Where labels get lost
Labels fail at the joins. A file exported from your app may lose its metadata in a messaging app. An API customer may strip your watermark before publishing. A user may crop the visible label out.
Map the journey of one piece of content from prompt to public post. Ask, at every step, who controls the file and whether the label survives. Each gap is a design decision.
Some gaps are best closed by contract. If you supply models or APIs to other businesses, your terms should forbid removing provenance markers and make customers responsible for their own disclosure duties as deployers.
Labels do not fix prohibited content
A label does not make unlawful content lawful. From 2 December 2026, the EU’s AI Act amendments ban AI systems that generate non-consensual intimate imagery or child sexual abuse material. The US and UK rules above point the same way.
That means guardrails at the model and policy layer, not just a tag on the output. Your acceptable-use policy, prompt filters, output classifiers and trust-and-safety team need to work as one system. Fast takedown timelines, three hours in India and 48 hours in the US for intimate imagery, only work if escalation is already rehearsed.
Checklist
- List every feature that generates or alters images, audio, video or public-facing text.
- Decide your role for each feature: provider, deployer, platform or several at once.
- Add visible labels at every point users see, share or download generated content.
- Embed machine-readable provenance using an open standard such as C2PA.
- Add a robust invisible watermark alongside metadata.
- Test whether labels survive export, compression, screenshots and API delivery.
- Update customer and API terms to forbid stripping provenance markers.
- Block prohibited generation, including intimate imagery and child abuse material, at the model and policy layer.
- Rehearse a takedown process that can meet a three-hour deadline.
FAQs
We only use a third-party model. Are we still responsible?
Often, yes. Under the EU AI Act the model developer usually handles machine-readable marking, but you as deployer must disclose deepfakes and certain public-interest text. If you build a product on the model and sell it under your own name, you may become a provider yourself.
Does AI-assisted text need a label?
Not always. The EU rule targets AI-generated text published to inform the public on matters of public interest, and it falls away where a human reviews the text and takes editorial responsibility. Marketing copy a person has edited is usually outside it. Chatbots must still tell users they are talking to AI.
Is a watermark enough on its own?
Rarely. Regulators and the EU Code of Practice recognise that no single technique is reliable today. Combine visible labels, metadata and watermarking, and keep your own generation logs.
We are not based in the EU or India. Do these rules apply?
They can. The EU AI Act applies where a system is placed on the EU market or its outputs are used in the EU. India’s rules apply to intermediaries serving users in India. Where your users are matters more than where your company is incorporated.
What about satire and art?
The EU still expects disclosure of deepfakes in creative work, but in a way that does not spoil the work, such as a credit or caption. Check each market, because exemptions are not uniform.
Primary sources
Briefing · Hogan Lovells on India’s AI labelling rulesBriefing · Orrick on the AI Act Digital Omnibus (July 2026)General information, not legal advice.