Beginner level6 min readTo learn at another level, choose it before you start the course.
After this lesson you canTell visible labels from hidden marks, say who must act on AI-made content, and explain how US law gets sexual deepfakes removed.
The short answerRules on AI-made content act in four places. Labels show people when AI made something, and hidden marks go inside the file. Platforms must remove sexual deepfakes fast, and sharing them is a crime in many places.
In simple words
- Deepfakes are fake pictures, videos or voices made with AI.
- Since August 2026, the EU has required deepfakes made for work to be labelled as AI-made.
- AI tools must add hidden marks that computers can detect.
- Sharing sexual deepfakes of real people is a crime in many places.
The short answerRules on AI-made content work in four layers: labels for people, machine-readable provenance in the file, duties for platforms, and consent rules and crimes that protect the people shown. The EU AI Act, California, Connecticut and the US TAKE IT DOWN Act each cover parts of this.
In simple words
- Labels: EU deepfake disclosure since August 2026, also for AI-written public-interest text.
- Provenance: machine-readable marks in the EU; hidden provenance data in California and Connecticut.
- Platforms: big platforms in California must show provenance from 2027.
- Abuse: a US federal crime, with 48-hour takedowns for intimate deepfakes.
The short answerSynthetic-media rules stack four tools: Article 50 transparency, provenance mandates, platform duties around provenance, and criminal and takedown regimes for non-consensual intimate imagery. Detection stays unreliable, so the laws lean on provenance.
Key points
- Article 50(2): machine-readable marking by providers; 50(4): disclosure duties for deployers.
- California: hidden provenance plus a free verification tool; platforms must surface provenance from 2027.
- Connecticut: provenance that is hard to remove, for example C2PA, since October 1, 2026.
- Intimate deepfakes: a US federal crime since May 2025; 48-hour platform takedowns since May 2026.
On the content Labels that people can see
- Label deepfakes and AI-written news-style textEU In force
- Label AI faces or bodies in paid posts “Images virtuelles”France In force
- Disclose AI-generated performers in your adsNew York In force
- Get consent or clearly label AI fakes of real peopleFrance In force
- Label automated bot posts on social networksGermany In force
Inside the file Hidden marks that machines read
- Mark AI-generated content so machines can detect itEU In force
- Embed hidden AI labels and offer a free checking toolCalifornia In force
- Add provenance data to AI images, audio and videoConnecticut In force
- Do not knowingly host AI models that skip hidden labelsCalifornia From Jan 2027
On the platform Show the marks, remove abuse fast
- Big platforms must show whether content is AI-madeCalifornia From Jan 2027
- Remove reported intimate deepfakes within 48 hoursUS federal In force
- Remove intimate deepfakes in 48 hours, or face state finesConnecticut In force
- Let users report sexual deepfakes and act within 30 daysCalifornia In force
For the person shown Consent, and crimes for sexual fakes
- Never post AI-made intimate images of real peopleUS federal In force
- Do not offer or use AI that makes sexual deepfakesEU From Dec 2026
- Do not make or help spread sexual deepfakesCalifornia In force
- Never share sexual deepfakes of a person without consentFrance In force
- Never make or share fake recordings to smear someoneHungary In force
- Get consent before using a person’s AI voice or likenessCalifornia From Jan 2027
- Get consent before using a person’s AI digital replicaIllinois In force
Words to know
- Deepfake
- An AI-made picture, video or voice that seems real, often showing a real person.
- Provenance data
- Hidden data in a file that records where and how it was made.
- Platform
- A website or app where users share content, such as a social network.
Visible labels
A deepfake is a fake picture, video or voice recording made with AI. Since August 2, 2026, anyone in the EU who uses AI to make deepfakes for work must say that they are AI-made.
In France, paid influencer posts with AI-made faces or bodies must say “Images virtuelles”. New York requires ads to disclose performers that were made with AI.
Hidden marks
Labels for people are easy to remove, so laws also ask for hidden marks inside the file. In the EU, AI tools must mark what they make so that computers can detect it.
California requires AI image, video and audio tools to add hidden data about where content came from, and to offer a free checking tool.
Sexual deepfakes
In the US, France and other places, sharing sexual deepfakes of a real person without consent can be a crime. Since May 2026, US platforms must remove reported images within 48 hours.
From December 2, 2026, the EU bans AI tools that make sexual deepfakes of identifiable people without their consent.
Try it yourself
Choose one rule from each layer of the diagram. For each one, say who has to act: the AI maker, the publisher or the platform.
Check yourself
A company in Rome uses AI to make a video of a famous singer dancing in its shop, and posts it. What does EU law ask of it?
A visible “AI-made” label is easy to cut off a picture. What do laws add for this reason?
A woman in Ohio finds a sexual deepfake of herself on a social media site and reports it. What must the site do?
Words to know
- Deepfake
- An AI-made picture, video or voice that seems real, often showing a real person.
- Provenance data
- Hidden data in a file that records where and how it was made.
- Platform
- A website or app where users share content, such as a social network.
Disclosure
The EU AI Act requires anyone who uses AI to make deepfakes, except in purely personal use, to disclose that the content is AI-made. For obvious art, satire or fiction, a light disclosure is enough. This has applied since August 2, 2026.
AI-written text that informs the public on matters of public interest must be disclosed too, unless a person reviewed it and someone holds editorial responsibility. So a newsroom whose editors check AI-assisted text and take responsibility for it does not need the label.
Provenance
Providers of AI that generates audio, images, video or text must mark the output in a machine-readable way. Systems on the market before August 2, 2026 have until December 2, 2026.
California’s AI Transparency Act requires hidden provenance data and a free checking tool. Connecticut asks AI makers to embed provenance that is hard to remove, for example with the C2PA standard.
Platforms and takedowns
From 2027, large platforms in California must detect provenance data and show users whether content was made by AI or captured by a camera. They may not knowingly strip it.
Under the US TAKE IT DOWN Act, publishing intimate images of real people without consent, including AI fakes, is a federal crime. For adults, the post must be meant to cause harm, or must cause it. Since May 19, 2026, platforms must remove a reported image within 48 hours of a valid report.
Voices and faces
Several laws protect a person’s voice and face. Illinois requires consent for AI replicas of a person. California has long required consent to use a person’s voice or likeness in ads, and from 2027 its law names AI replicas plainly. New York, and California since 2025, require the heirs’ consent for replicas of dead performers.
France makes it a crime to share AI fakes of real people without consent, unless it is obvious or clearly stated that they were made by AI.
Try it yourself
Choose one rule from each layer of the diagram. For each one, say who has to act: the AI maker, the publisher or the platform.
Check yourself
A news site in Vienna uses AI to draft an article on new tax rules. When may it skip the EU’s AI disclosure?
From 2027, a large social platform in California gets a photo whose provenance data says a camera took it. What must the platform do?
Someone in Lyon shares a funny AI video of a real actor, without consent, with the caption “made with AI”. Is that a crime in France?
Article 50 in detail
Providers must make generative outputs detectable as AI-made in a machine-readable way, with exemptions for tools that only assist standard editing. Systems already on the market have until December 2, 2026.
Deployers must disclose deepfakes, with a lighter duty for evidently artistic or satirical work. Public-interest text needs disclosure unless a human reviewed it under editorial responsibility.
Provenance mandates
California’s AI Transparency Act, amended by AB 853 and SB 1000, requires hidden provenance in generated images, video and audio and a free verification tool. Since September 30, 2026, it has applied regardless of user count.
From 2027, hosting platforms may not knowingly offer generative models that skip the required provenance, and large platforms must detect and display it. Connecticut requires hard-to-remove provenance where reasonable.
Why provenance, not detection
Tools that guess whether media is synthetic can be wrong in both directions. Provenance records origin when content is created, so a valid record is much stronger evidence than a guess.
Its weakness is stripping: re-encoding or a screenshot can remove metadata. That is why laws add platform duties not to strip it, and ask for marks that are hard to remove.
Intimate images and likeness
The TAKE IT DOWN Act made publishing non-consensual intimate images, including AI fakes, a federal crime from May 19, 2025. For images of adults, the post must be meant to cause harm or must cause it. It requires 48-hour removal from May 19, 2026.
California presumes that services built mainly for sexual deepfakes know there was no consent. The EU bans such AI tools from December 2, 2026, and France makes sharing sexual deepfakes a crime even when they are labelled.
Try it yourself
Choose one rule from each layer of the diagram. For each one, say who has to act: the AI maker, the publisher or the platform.
Check yourself
An EU photo app adds AI features that only fix red eyes and brightness. Does the Article 50(2) marking duty apply?
A detector says a viral photo is 92% likely AI-made, but a valid provenance record says a camera took it. Which is stronger evidence?
In June 2025, someone in Texas knowingly posts an AI-made intimate image of a real adult without consent, to humiliate them. What is it under the TAKE IT DOWN Act?
Sources
- AI Act, Article 50: transparency (EU AI Act Service Desk)
- SB 1000, amending the California AI Transparency Act (California Legislature)
- TAKE IT DOWN Act, Public Law 119-12 (GovInfo)
- Public Act 26-15 (Connecticut General Assembly)
- Content credentials standard (C2PA)
- Civil Code section 3344: consent to use a voice or likeness (California Legislature)
This lesson was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it lists. The live parts come from our checked stories, trackers, model comparison and rules checker.