AI-generated images, video and audio are now ordinary parts of advertising, fundraising and social media. Since August 2026, the rules for labeling them have also become real law in more places. There is still no single standard. What you have to do depends on where your audience is, what the content shows and where you post it.
This explainer sorts the rules into three layers: the European Union’s AI Act, U.S. state laws, and the policies of the big platforms. It describes the rules as of November 2026. Several are new and still being interpreted, so check the official text before relying on any detail.
The short version
| Rule | Who it applies to | What it requires | In effect |
|---|---|---|---|
| EU AI Act, Article 50 | AI tool makers, and anyone who publishes AI content to people in the EU | Machine-readable marking by the tool; visible labels on deepfakes and on some AI-written public-interest text | August 2, 2026, with a grace period to December 2, 2026 for marking by tools already on the market |
| California AI Transparency Act | AI image, video and audio services with more than 1 million monthly users | Hidden provenance data in every file, an optional visible label, a free detection tool | August 2, 2026; platform duties from January 1, 2027 |
| State election deepfake laws | Campaigns, committees and others paying for political messages | A disclaimer on deceptive synthetic media, usually in a set period before an election | 30 states as of May 2026 |
| Platform policies | Anyone uploading to YouTube, Facebook, Instagram or TikTok | Disclose realistic AI content when you post it | Now |
The EU: two duties, split between the tool and the publisher
The EU AI Act mostly regulates high-risk uses of AI, but its Article 50 sets transparency rules that reach almost everyone who makes or shares AI content for European audiences. It splits the work in two.
Providers, the companies that build generative AI systems, must mark the audio, images, video and text their systems produce in a machine-readable way, so the content can be detected as AI-made. That usually means metadata, a watermark or both. This applies from August 2, 2026, with a grace period until December 2, 2026 for systems that were already on the market before August.
Deployers, the people and organizations who use those systems in their work, have a different duty. If you publish a deepfake, meaning AI-made or altered image, audio or video that resembles real people, places or events and would falsely appear authentic, you must disclose that it is artificial. If you publish AI-generated text to inform the public on matters of public interest, you must also disclose it, unless a person reviewed it and someone takes editorial responsibility for it.
Two points are often missed:
- Creative work gets a lighter rule, not an exemption. For content that is evidently artistic, creative, satirical or fictional, the disclosure can be done in a way that doesn’t spoil the work, such as a credit at the end, but it still has to be there.
- Purely personal use is outside the law. The AI Act doesn’t cover private, non-professional activity. A business, a campaign or a charity is acting professionally.
In June 2026 the European Commission published a final, voluntary Code of Practice on marking and labeling AI-generated content. It asks providers to use at least two layers of marking where needed, such as metadata plus a watermark, and to offer ways to detect them. For deployers, it points to a common EU icon or an equivalent label, and to spoken disclaimers where a visual label isn’t possible. Signing is voluntary, but following it is the simplest way to show you comply. Breaking the transparency rules can bring fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher, with lower caps for small businesses.
The United States: no federal rule, a growing set of state ones
There is no federal law requiring labels on AI content in general. As covered in who gets to regulate AI, the rules come from states, and the federal government is trying to limit them.
California’s AI Transparency Act is the broadest. It became operative on August 2, 2026 and puts the duties on large providers rather than on users:
- It covers generative AI services for images, video and audio with more than one million monthly users that are available in California. AI-generated text is not covered.
- Every file must carry a hidden (“latent”) disclosure that identifies the provider, the system and version, and the creation date.
- Users must be offered the option of a visible (“manifest”) label that is hard to remove.
- Providers must offer a free public tool for checking whether content came from their system.
- Violations carry civil penalties of $5,000 each, with each day counting separately, enforced by the attorney general. There is no private right to sue.
Further duties follow. From January 1, 2027, large online platforms must detect that provenance data and show it to users. From January 1, 2028, makers of cameras and phones sold in California must offer the option of recording provenance data in their own captures.
Election deepfake laws are the most common kind. Maryland became the 30th state with one in May 2026, according to Public Citizen’s tracker. Most require a disclaimer on AI-generated or materially altered depictions of candidates within a set period before an election, rather than banning them outright, and most exempt satire, parody and news coverage. Wording, time windows and penalties vary a great deal, and some laws have been challenged on free speech grounds. See AI deepfakes and elections for how these laws came about.
Separately, the Federal Communications Commission has ruled that AI-generated voices in robocalls count as “artificial” voices under federal law, so AI voice robocalls without consent are illegal regardless of any label.
The platforms: disclose when you upload
For most people, platform rules matter more day to day than any statute, because they apply the moment you post. The major platforms have similar policies, each with its own setting.
- YouTube requires creators to disclose content that seems realistic and was made or meaningfully altered with AI, such as making a real person appear to say something they didn’t, altering footage of a real event or showing a realistic scene that never happened. The setting is in YouTube Studio when you upload. Photorealistic content may get a label in the video player; other disclosures appear in the description. Clearly unrealistic content, beauty filters, color correction and captions don’t need disclosure. YouTube may add the label itself when a file carries content credentials, and creators who repeatedly fail to disclose can lose monetization or have videos removed.
- Meta (Facebook and Instagram) asks users to disclose photorealistic video and realistic audio that was digitally created or altered, and adds an “AI info” label. It also labels content automatically when it finds industry-standard signals from AI tools. Political and social issue ads must disclose digital alteration.
- TikTok requires creators to label AI-generated content that shows realistic scenes or people, with a setting in the post options, and automatically labels some content that carries content credentials. Unlabeled realistic AI content can be removed.
- Political ads on Google and Meta carry their own disclosure requirements for synthetic content that depicts real people or events, on top of any state law.
Platform policies change often. Treat these descriptions as a guide and check each platform’s help pages before a campaign.
What this means if you post AI content
The duty to label usually falls on whoever publishes, not only on the tool. If a local nonprofit uses an AI video generator to make a short fundraising clip with a realistic scene, the platform’s “made with AI” toggle and, for EU audiences, the AI Act’s deepfake label are the publisher’s job. The tool’s hidden marking helps, but it doesn’t replace a label people can see.
A practical checklist:
- Ask whether it looks real. Every set of rules focuses on realistic content that could mislead. A cartoon or an obviously impossible scene rarely needs a label; a realistic street scene, product demo or person usually does.
- Ask whether it shows a real person. Content depicting real, identifiable people, especially candidates, carries the strictest rules and, in many states, legal risk beyond labeling.
- Check where your audience is. Content aimed at people in the EU falls under the AI Act even if you are not in Europe.
- Use the platform’s setting, not just a caption. The built-in toggle is what the platform checks, and it puts the label where viewers see it.
- Keep the provenance data. Don’t strip metadata from files your tool marked. Some platforms read it and label the content automatically, and removing it can defeat the purpose of the law.
- Be stricter near elections. Anything political in the weeks before a vote may fall under a state deepfake law, whatever the platform allows.
Why labels have limits
Labels help, but they are not proof either way. Watermarks and metadata can be lost when a file is screenshotted, re-encoded or edited, and tools that don’t follow the rules won’t add them. A missing label doesn’t mean content is real. Researchers also warn about the “liar’s dividend”: once fakes are common, real recordings can be dismissed as AI. Checking the source still matters, and the claim checklist gives a quick way to do it.
What to watch
- December 2, 2026: the EU grace period for machine-readable marking ends.
- Commission guidelines on Article 50, which will clarify what counts as a deepfake and how the creative-work rule applies.
- January 1, 2027: California’s platform duties begin, and pending bills could change the law before then.
- Federal preemption. A federal law or court ruling could override some state rules; see who gets to regulate AI.