This article deals with non-consensual deepfakes, including those of a sexual nature, through their figures and their legal framework. No graphic descriptions. If the subject affects you personally, resources are listed at the end of the article.
The real scale of the phenomenon in 2026
For a long time, the word "deepfake" evoked viral videos of politicians, social media hoaxes, or historical reconstructions. The term "deepfake" comes from the contraction of "deep learning" (a machine-learning technique that uses multi-layer neural networks) and "fake". It refers to any audio, image or video content generated or altered by AI to make a real person say or do something they never said or did.
That technically neutral view is now overtaken by the statistical reality. According to converging figures from several studies (DeepMedia, Sensity AI, New York State Bar Association), 98% of deepfakes in circulation are pornographic. And 99% of these deepfakes target women.
The deepfake detection market tripled between 2023 and 2026, rising from $5.5 billion to $15.7 billion. Growth in detections in North America reached +1,740% over one year. In Asia-Pacific, +1,530%. These figures do not measure the increase in deepfakes themselves (which is far faster), but that of the industry trying to spot them. The gap between creation and detection is probably the worst signal in this whole story.
Who are the victims, really?
Celebrities were the first victims to receive media attention, and to some extent they still are. But the 2026 reality is very different. A study by the American Academy of Pediatrics (the leading professional organisation of paediatricians in the United States) showed that deepfakes are now used on a massive scale to blackmail teenagers aged 14 to 17. In August 2024, a case in South Korea revealed thousands of AI images of teachers and pupils, some of whom were 11 to 13 years old.
In France, several similar cases emerged in 2025 in middle and secondary schools. The method is almost always the same: a classmate takes a publicly available photo (Instagram, class photo, school profile), runs it through a "nudification" app (software that uses AI to generate a nude version of a person from a clothed photo), which sometimes costs less than 5 euros a month, and shares the result on social media or Snapchat. The victim discovers the images weeks later, sometimes after they have circulated throughout their whole school.
According to research cited by UN Women (the UN agency for gender equality and the empowerment of women) and UN Geneva, more than half of deepfake victims in the United States have had suicidal thoughts. This figure, chilling as it is, is not activist statistics: it is the result of clinical studies into measurable psychological consequences. Anxiety, depression, symptoms of post-traumatic stress, social withdrawal. What sets deepfakes apart is that the victim knows the images are not real, but that does not lessen the psychological impact. On the contrary, researchers note that the "unreal yet hyperrealistic" nature of the content creates a particular form of traumatic dissociation.
The law is slowly catching up with the technology
Several major legal frameworks come into force in 2026. Here is where things stand.
United States: Take It Down Act. Signed by Donald Trump on 19 May 2025, the Take It Down Act came into force immediately as regards its criminal provisions, but with a one-year grace period for platforms. Since 19 May 2026, all platforms hosting user content must have in place a report-and-remove system that deletes non-consensual content within 48 hours. Failure to comply exposes the platform to penalties from the Federal Trade Commission (FTC, the US federal agency for consumer protection and competition regulation). The law explicitly covers AI-generated deepfakes.
European Union: AI Act. The European regulation on artificial intelligence (in phased application since 2024) imposes transparency obligations on deepfakes: any content generated or modified by AI must be identified as such. However, the AI Act does not specifically address the criminal aspect of non-consensual deepfakes, which falls under national legislation.
France. The law of 21 May 2024 explicitly criminalised the distribution of non-consensual sexual montages, with penalties of up to 3 years in prison and a 75,000 euro fine where the victim is an adult, and 5 years in prison and a 100,000 euro fine where the victim is a minor. Creating such content, even without distributing it, is also punishable.
United Kingdom. The Online Safety Act prohibits the sharing of digitally manipulated sexually explicit images. Its limitation: it does not explicitly cover the creation of deepfakes, only their sharing, and does not always apply where intent to cause harm cannot be proven.
Brazil. Amendment to the criminal code in April 2025: increased penalties for psychological violence against women using AI to alter their image or voice.
Australia. 2024 law criminalising the creation or sharing of deepfake sexual content, with up to 6 years in prison.
Why the law is not enough (and probably never will be)
The UN and several legal experts agree on a harsh finding: fewer than half of countries have laws dealing with online violence, and even fewer have specific legislation on AI-generated deepfakes. Four structural obstacles limit the effectiveness of regulation.
Obstacle 1: the legal grey area. In most countries, "revenge porn" laws (the American term for the non-consensual distribution of intimate images for the purposes of revenge) were written before deepfakes existed. They often do not cover artificially generated images, creating a legal vacuum. Victims do not even know whether the abuse is illegal in their jurisdiction.
Obstacle 2: the difficulty of investigating. Even where laws exist, investigators need expertise in digital forensics (the scientific analysis of traces left by the use of computer devices), international cooperation and technical resources. Most judicial systems have none of these three elements in sufficient quantity.
Obstacle 3: the personal cost for the victim. To file a complaint, a victim must show their sexualised deepfake images to police officers, lawyers, moderators. They must see their name in official files, risk media attention, and sometimes face defamation complaints from their own abusers. Many choose to block, remove and try to survive without filing a complaint.
Obstacle 4: the cross-border nature. An image generated in Russia, hosted in the United States, shared in France does not fall within any simple legal framework. International cooperation protocols on cybercrime remain embryonic.
Can this technology still be stopped?
This is the central question, and the answer is honestly no. Here is why.
Firstly, the technology is now open source (software and models whose code and parameters are publicly available, modifiable and redistributable). The models capable of creating deepfakes (Stable Diffusion, Flux, dozens of derivatives) can be downloaded for free. Once a model is publicly available, no government can recall it. As we explored in our article on the US-China AI war, Chinese and Western regulations on proprietary models do not apply to open source models circulating everywhere. We develop this point in detail in our article on the limits of regulation.
Secondly, the cost of entry is negligible. The first "deepnude apps" cost thousands of dollars in infrastructure. Today, any smartphone can generate a convincing deepfake via an app with a subscription of 5 to 20 euros a month. The technical barrier has disappeared.
Thirdly, detection structurally lags behind creation. It is a game of cat and mouse: every improvement in detectors is followed by an update to the generators to circumvent them. Security researchers consider this imbalance to be probably permanent.
Fourthly, viral spread is instantaneous. As we documented in our article on AI-generated school textbooks, the internet does not distinguish between verified content and generated content. A deepfake image can be seen by millions of people before a moderation system reacts.
What this concretely changes in our relationship with images
Beyond the legal question, the mass deployment of deepfakes is structurally transforming our relationship with images. Three consequences are already observable.
The presumption of authenticity disappears. For 150 years, a photograph was considered reasonable proof of reality. For the first time in modern history, this presumption is collapsing. An image no longer proves anything unless it is verified by other sources. This cognitive inversion is profound, and children born in 2020 will probably grow up with this new intuitive norm: what we see is probably not true by default.
"Plausible deniability" changes the political game. In French, the term translates as "déniabilité plausible": the possibility of credibly denying the veracity of a piece of information. If any image can be a deepfake, then a genuine compromising image can now be disputed as "probably AI". This new ambiguity paradoxically protects the guilty: Donald Trump and several public figures have already used this argument to contest authentic photos. The Pelicot case in France in 2024 showed, conversely, that in the other direction it is also doubt that slows down justice.
Protection becomes virtually impossible. For a person to be "safe" from deepfakes, no photo of them would need to be publicly accessible. That is impossible in a society where every adult generally has a few photos on Instagram, LinkedIn, or simply school photos. As a researcher at the Alan Turing Institute (the UK's leading research institute for data science and artificial intelligence) puts it, "a single photo is now enough to generate hundreds of deepfake images in a matter of minutes".
Many observers note that a gradual normalisation is taking place, particularly among generations born after 2005. For them, the fact that sexualised images can circulate about them without their consent is not so much a catastrophe as "a known risk of life online". This trivialisation is itself a serious problem: it could lead to a society where the systematic violation of one's image is no longer fought, but simply absorbed as a cost of digital modernity.
What could, despite everything, improve the situation
An honest analysis must also note what can concretely help.
Regulatory effort is converging. The Take It Down Act, the AI Act, national laws: for the first time, several major jurisdictions are acting simultaneously. Fragmentation remains a problem, but the political momentum is now aligned.
Technical provenance standards are making progress. The C2PA standard (Coalition for Content Provenance and Authenticity, a coalition bringing together Adobe, Microsoft, Google and several others to certify the origin of digital content), makes it possible to cryptographically sign authentic images. If widely adopted, this standard could invert the problem: proving that an image is true rather than proving that it is false.
Proprietary AI models are building in safeguards. OpenAI, Anthropic and Google now actively block the generation of sexual content depicting real people. This obviously does not solve the open-source problem, but it limits the scale of the phenomenon for mainstream users.
Platforms are starting to cooperate. Under regulatory pressure, Meta, X, TikTok and several others are improving their automated detection systems. Far from perfect, they now remove more deepfake content within a few hours than they did two years ago.
What to take away for 2026
Deepfakes are not a passing problem that will be solved by a good law or a good technology. They represent a structural shift in our relationship with images, comparable in magnitude to the move from drawing to photography in the 19th century, or from photography to digital in the 1990s. Each of these transitions took decades for society to invent new norms, new legal tools and new behaviours.
The deepfake transition will probably be faster, because the technology is evolving more quickly. But it will still take time. In the meantime, those who have already grasped the change (notably younger people, who are growing up with this reality) are intuitively developing new reflexes: doubting by default, checking sources, protecting their own image. That is probably the only response suited to the individual scale.
At the collective scale, much remains to be done: harmonising laws, funding detection, training investigators, making platforms accountable. This analysis is part of our series on the blind spots of AI. You can continue with our article on the hidden economy of AI companions, which explores another, less discussed side of the phenomenon, and our analysis of the impossibility of fully regulating AI.
In France, the PHAROS platform (Plateforme d'harmonisation, d'analyse, de recoupement et d'orientation des signalements, run by the national police) collects reports of illegal content online, including non-consensual deepfakes. Le 3018, the national helpline for victims of digital violence, is free and confidential (Monday to Saturday, 9am-8pm). The charity StopFisha specifically supports victims of image abuse, including deepfakes.