Deepfakes and the Law — When a Fake Image Does Real Harm, What Does the Law Punish?

The Face of a Sixteen-Year-Old Junior

Between February and March 2023, a defendant living in Seoul transferred photos of a middle school junior's face, collected from Instagram, to a stranger the defendant met on Twitter. The junior was sixteen years old. The defendant made five requests to create sexually composite materials combining that face with another woman's body, and received the resulting photos. The prosecutor indicted this as the crime of producing sexually exploitative videos of minors. In May 2024, the court of first instance, the Seoul Western District Court, found this count guilty and sentenced the defendant to three years in prison suspended for five years, combined with other crimes. In November 2024, the appellate court, the Seoul High Court, punished the same act under a different crime in a different law (Seoul High Court 2024No1823). In August 2025, the Supreme Court upheld that judgment (Supreme Court 2024Do17801).

Deepfakes are videos, photos, and audio created by using artificial intelligence to synthesize a real person's face, body, or voice into another video or to invent them anew, making it look as though that person did something they did not do. In that same year of 2024, Korean society experienced the deepfake sex crime crisis. Following that crisis, there is research tracking the process by which non-consensual synthetic sexual images came to be recognized as crimes (DOI: 10.1080/01924036.2025.2596578). Multiple actors pushed to have this issue recognized as a crime, and the strengthening of legal provisions, victim support, and reforms to the artificial intelligence management system began rapidly. Cross-border detection and international cooperation still remained barriers.

The term deepfake spread in 2017 when sexually explicit videos manipulated by artificial intelligence were uploaded to a discussion site. Even by a 2020 paper, the majority of deepfakes circulating online were sexual, yet public attention was focused on videos impersonating politicians to spread misinformation (DOI: 10.1080/23268743.2020.1757499). Even based on a 2021 Canadian policy review, most deepfakes found online were sexual, and it was rare for the people appearing in them to have consented to their production and publication (DOI: 10.1089/cyber.2020.0272).

A scoping review examining 54 English academic papers peer-reviewed from December 2017 to January 2026 notes the same mismatch. The framework prioritizing misinformation and the celebrity-centric framework have frequently obscured the harm suffered by ordinary people and the inequality of remedy measures (PMID 42745201). According to conventional wisdom, deepfakes are a matter of whether something is real or fake, and the problem is solved once detection technology sorts out the fakes. The places where harm actually accumulates lie outside what that framework clearly illuminates.

This article focuses primarily on Korean laws and judgments. Regulations from other countries such as the United Kingdom, the United States, the European Union, and Australia are compared only to the extent described in the shortlist papers. Because legal papers are frequently uploaded to public repositories before peer review, the grade of such materials is noted each time.

Even When Labeled as Fake

There is an experiment that showed deepfake videos to a representative sample of U.S. adults and asked how wrong they were. Non-consensual sexual deepfakes were rated as extremely harmful, and the vast majority of respondents wanted the creators to be criminally punished. Adding a "fictional" label to sexual deepfakes did not reduce the judgment that they were wrong. Non-sexual deepfakes were different. When they were labeled as fictional or did not contain actions that undermined original reputations, the judgment that they were wrong decreased significantly. In a follow-up experiment, respondents judged creating and spreading sexual deepfake videos to be as harmful and slightly more morally blameworthy than spreading actually recorded sexual videos without consent (DOI: 10.2139/ssrn.3781968, preprint). This study measured people's perceptions in a single country, the United States. Within that scope, the fact that a fictional label did not change judgments can be read as showing that the wrongness of sexual deepfakes is judged separately from whether the viewer was fooled.

Image-based sexual abuse is a concept that groups together the non-consensual filming, creation, or distribution of others' nudity or sexual photos and videos, viewing them as a form of sexual violence. A 2016 paper that newly proposed this concept describes the harm as follows. Physical and mental illness, alongside the loss of dignity, privacy, and sexual self-determination, befall individuals, and these combine into cultural harms affecting society as a whole (DOI: 10.1093/ojls/gqw033). A 2017 paper pointed out that the term "revenge porn" has made legislation piecemeal by casting light only on the retaliation of ex-lovers, and proposed viewing such acts as a single continuum connected to other sexual violence (DOI: 10.1007/s10691-017-9343-2). Sexual deepfake composites fall within this continuum. A research monograph based on surveys of over 6,000 people across three countries and more than 100 interviews places AI-generated sexual deepfake videos on that list, and identifies power and control as the dominant motives for the offense (DOI: 10.4324/9781351135153).

Evidence regarding what the harm looks like is most specific among adolescents. A systematic review combining 12 empirical studies covering ages 10 to 24 reports psychological distress such as fear, anxiety, depression, and suicidal ideation; bullying and victim-blaming; and disruptions in education or employment such as transferring schools or losing jobs. The authors conclude that these results show similarities between image-based sexual abuse and physical sexual abuse (PMID 40709226). This review does not look at deepfakes alone, but covers image-based sexual abuse as a whole.

A scoping review limited to deepfakes views harm as accumulating over time. It states that at each stage of creation, distribution, visibility, and resurfacing, additional suffering, stigma, potential losses to relationships and careers, self-silencing, and procedural burdens are added. The same review also notes two blanks. Empirical studies centered on victims are scarce, and studies directly evaluating the effectiveness of countermeasures are almost non-existent (PMID 42745201).

There is also the question of how common it is. In a study surveying over 16,000 people across 10 countries, 2.2% of respondents answered that they had experienced harm from sexual deepfake composites, and 1.8% answered that they had committed offending behavior (DOI: 10.1145/3613904.3642382). The denominator for this percentage is all respondents across the ten countries, and the abstract does not disclose which those ten countries are. It should not be read as a figure for South Korea.

For U.S. respondents, the wrongness of sexual deepfakes lay in becoming a sexual object without consent, and it was not reduced by a fake label. Studies on harm also describe the damage as the loss of dignity, privacy, and sexual self-determination.

The Questions Law Picks — Does It Look Real, Did You Agree?

Faced with the same fake, different laws ask different questions. Some laws ask about appearances and deception, while others ask about consent. A paper comparing regulations in the UK and the US calls this split "Dignity versus Deception" (DOI: 10.2139/ssrn.7530380, preprint). According to this paper, the US federal Take It Down Act requires two things: there must be an intention to cause harm, and the forged material must be indistinguishable from a real video. The author criticizes the first requirement for opening an escape hatch where someone can claim it was "just satire," and the second for shifting the center of the wrong from the invasion suffered by the victim to the deception suffered by the viewer. In the author's view, the result leaves victims of crudely made composites outside of protection. As this paper sees it, the UK shifted toward treating sexual self-determination as a right through the Online Safety Act 2023 and Section 66B of the Sexual Offences Act 2003, making the unauthorized sharing itself the basis of liability. As the author sees it, both countries have a vacant seat: the user who uploads bears responsibility, while the side that develops tools for creating abusive material remains largely outside regulation.

Article 14-2 of South Korea's Act on Special Cases Concerning the Punishment of Sexual Crimes stands on the consent side of this split. False video material is a term in South Korea's Sexual Violence Punishment Act referring to video material containing a person's face, body, or voice that has been edited, synthesized, or processed against that person's will into a form that may arouse sexual desire or shame. The official name of the Sexual Violence Punishment Act is the Act on Special Cases Concerning the Punishment of Sexual Crimes. This provision was newly created in March 2020 because deepfake harm was increasing while previous regulations made punishment difficult or lenient (Seoul High Court 2024No1823). What the article of the law asks about is the subject's will and the form. The requirement that it must be indistinguishable from the real thing is not in the article. The pre-amendment provision applied in the earlier case punished a person who made such edits for the purpose of distribution with imprisonment for up to 5 years or a fine of up to 50 million won (Sexual Violence Punishment Act Article 14-2, pre-October 16, 2024 amendment provision — the wording is identical to the January 25, 2024 enforcement version). The revised edition enforced on October 16, 2024 removed this purpose requirement and raised the penalty to imprisonment for up to 7 years or a fine of up to 50 million won (Sexual Violence Punishment Act Article 14-2, enforced October 16, 2024).

If the victim is a minor, one more law overlaps. It is the Act on the Protection of Children and Youth against Sex Offenses, abbreviated as the Youth Sex Protection Act. Child and youth sexual exploitation material is a term in the Youth Sex Protection Act referring to video, imagery, etc. in which children or youth, or people or expressions who can be clearly recognized as children or youth, appear and engage in sexual acts. Until 2020, it was called "child and youth-utilized pornography." The name was changed to make it clear that this itself signifies sexual exploitation and sexual abuse (Seoul High Court 2024No1823). The person who creates this receives life imprisonment or imprisonment for 5 years or more (Supreme Court 2024Do17801). What this definition asks about is the appearance: whether the character is clearly recognized as a child or youth. Whether the subject agreed is not included in this definition. Even within South Korea, different laws ask different questions. Based on the pre-amendment provision applied in the earlier case, the upper limit for editing false video material is 5 years, whereas this side has a lower limit of 5 years. That is why which law to apply mattered greatly in that case.

Following a Case

We return to the Seoul High Court case. The court resolved two issues. First, can a photo combining a real child's face with another person's body be considered a sexual exploitation material "in which children or adolescents appear"? Second, if not, is it a sexual exploitation material "in which expressions that can be clearly recognized as children or adolescents appear"?

On the first question, the court acknowledged that the face is one of the most important body parts representing a person's identity. Therefore, it noted that there is room to see this as a case "in which children or adolescents appear." Then it gave five reasons for the opposite view. The first is the principle of legality. The principle of legality is the rule that what constitutes a crime and what punishment it receives must be determined by law in advance, and this principle also prevents reading penal provisions broadly to the disadvantage of the defendant. As the second and third reasons, the court cited the fact that the statutory provisions separate the two types, and that judging from the history of legislation and amendment, "in which children or adolescents appear" is limited to cases where real persons appear. The fourth reason was that composite photos lack direct sexual exploitation or sexual abuse of the victim child, making them qualitatively different from cases where a real child appears. The fifth was that since there is a provision punishing false video materials, there was no concern that a punishment gap would arise (Seoul High Court 2024No1823).

On the second question, the court applied the standard that, when viewed objectively from the perspective of an average member of society, the person must be recognized as a child or adolescent without a doubt in appearance. And it judged that these photos fell short of that standard. The composite body was difficult to see as a child or adolescent judging by its state of physical development. In several photos, the ratio of the face to the body and the skin tone did not match, making it relatively easy to tell the fact that parts belonging to different people were combined. Even in one more sophisticatedly composited photo, there was no information on the screen to show that the character was a child or adolescent (Seoul High Court 2024No1823). The part regarding the violation of the Act on the Protection of Children and Youths against Sex Offenses, which the first instance court viewed as guilty, was judged not guilty here. Because other offenses bundled into this case were recognized as guilty, a separate not guilty verdict was not written at the very front of the judgment.

Punishment came through another path. The prosecutor also charged the same act in the second instance under the crime of editing false video materials. This is a method of asking the court to judge this if the first charge is not recognized. The defendant confessed to this criminal fact, and the court found guilty the fact that composite materials were made in a form that could arouse sexual desire or shame against the victim's will. Including other offenses bundled into this case, the sentence became 2 years and 6 months in prison with 3 years of probation (Seoul High Court 2024No1823).

An appeal is seeking judgment from the Supreme Court by objecting to the second-instance judgment. The prosecutor appealed, and in August 2025, the Supreme Court dismissed the appeal with the unanimous opinion of the participating Supreme Court Justices (Supreme Court 2024Do17801).

The Supreme Court took this opportunity to establish a standard. Photos or deepfake videos combining the face of a real child correspond to images of children created by the creator and cannot be viewed as involving the appearance of an actual child, and thus do not fall under sexual exploitation materials "in which children or adolescents appear." However, if the content expresses sexual acts of a child, it may fall under sexual exploitation materials in which "expressions that can be clearly recognized as children or adolescents" appear. That judgment must comprehensively consider the person's appearance and physical development, the person's actual age or identity, the source or production history of the composite material, and the background and situational settings.

The lower court is the judgment rendered in the tier immediately below, which is re-examined by the upper court. To the Supreme Court, the lower court was the Seoul High Court judgment. While the Supreme Court deemed it inappropriate for the lower court to explain another single piece—specifically, a photo retaken using the victim's photo as the background—as a case "in which children or adolescents appear," it considered the conclusion that it is difficult to view all photos as sexual exploitation materials to be justifiable (Supreme Court 2024Do17801).

If we trace this case again from the beginning, it is like this. The first issue of whether they "appear" was divided by legal doctrines such as the principle of legality, the distinction between provisions, and the legislative history. What divided the judgment on the second issue was a body looking like an adult and the absence of information indicating they were a minor. The signs of compositing easily seen in several photos also factored into those circumstances, but one sophisticated photo with almost no traces led to the same conclusion. The judgment factors established by the Supreme Court are also appearance and physical development, actual age or identity, source or production history, and background and situational settings. Composite materials that failed to cross the threshold of the Act on the Protection of Children and Youths against Sex Offenses in this case were punished under the false video material provision, which asks for the subject's consent. That provision did not ask whether they looked real. The weight of the sentence changed. Based on the pre-amendment provision applied to this case, it moved from a provision with a minimum of 5 years to a provision with a maximum of 5 years.

A Child That Does Not Exist — What the Constitutional Court Saw

On June 24, 2026, the Constitutional Court, which decides whether laws conform to the Constitution, issued rulings on two cases on the same day. Both cases dealt with cartoons, and none of the characters were real children. A petition for constitutional review of a statute is a procedure where a court presiding over a trial requests the Constitutional Court to determine whether a legal provision used in that trial violates the Constitution. 2022Heon-Ga-8 is a case where the Jeju District Court made this petition during the trial of a person who uploaded 82 cartoon files featuring child and adolescent characters engaging in sexual acts to earn points, and downloaded and possessed 13 of them (Constitutional Court 2022Heon-Ga-8). 2023Heon-Ba-263 is a case filed by a person who produced adult cartoons nine times featuring characters that can be clearly recognized as children or adolescents, and distributed and sold them by uploading them to a foreign website. This person applied for a petition during the trial, and when the court refused to accept it, directly filed for a review with the Constitutional Court (Constitutional Court 2023Heon-Ba-263).

Both cases were decided as constitutional by the unanimous opinion of all justices. The reasoning of the Constitutional Court was that the purpose of the regulation is not limited to preventing harm suffered by children in the process of producing pornography. The premise is that expressive materials that make children and adolescents sexual objects create distorted perceptions and abnormal attitudes about sex in viewers, and can lead to sexual crimes against children and adolescents. Freedom of expression is the right to speak or show my thoughts. The Constitutional Court stated that even sexually explicit expression falls within its protective scope, and judged that while the restrictions placed on expression by this provision are not significant, the public interest in protecting children is incomparably significant (Constitutional Court 2023Heon-Ba-263). Regarding the mere act of possession, citing the point that it encourages supply and diffusion as a form of creating demand, it viewed that a prison sentence of one year or more is not excessively heavy. The fact that the minimum sentence is one year, allowing for a suspended sentence, was also cited as a reason (Constitutional Court 2022Heon-Ga-8).

Both decisions contain passages that directly mention deepfakes. 2023Heon-Ba-263 wrote that deepfake videos and the like using images of real children infringe upon the human dignity and value of real children, and thus can be viewed in themselves as a record of sexual violation and exploitation, meaning that an infringement of personal legal interests exists (Constitutional Court 2023Heon-Ba-263). 2022Heon-Ga-8 wrote the following. Due to the development of artificial intelligence technology, anyone can easily create such images with simple commands, and deepfake videos and the like utilizing images of real children are surging. It also wrote that deepfake videos and the like infringe upon the human dignity and value of the real children whose images are used, and in themselves become a record of sexual abuse and exploitation (Constitutional Court 2022Heon-Ga-8).

The opposing argument was written by the Jeju District Court, which sent the case to the Constitutional Court. It noted that there are doubts about the causal relationship between contact with virtual sexual exploitation materials and actual sexual crimes, that cartoons and animation are sometimes legally distributed abroad, and that cartoons drawn on books or paper are not punished simply for being possessed. It also argued that if no real child appears, there is no room for individual harm such as direct sexual exploitation or secondary victimization, making the weight of the crime lighter (Constitutional Court 2022Heon-Ga-8). The Constitutional Court did not accept these arguments.

In academia, arguments also emerge to use this distinction as the boundary for punishment. This is in the opposite direction of the Constitutional Court. A paper analyzing Portuguese criminal law divides virtual child sexual expression materials into two. Those that use even part of a real child's image and those that are virtual from beginning to end. The author views that only the side using real children's images impairs children's developmental conditions and thus should be treated as a crime (DOI: 10.15388/teise.2025.134.4). A scoping review combining 28 studies on image-based sexual abuse created by artificial intelligence observes the same boundary. Composites featuring identifiable children overlap with a victim-centered harm framework, while complete composites without identifiable children are treated as a separate category outside that framework (DOI: 10.1016/j.chbr.2026.101306).

When placing the Korean materials of this section side by side, one spot where the narratives conflict becomes visible. The Seoul High Court wrote that edited photos lack direct sexual exploitation or abuse, making them qualitatively different from cases where real children appear (Seoul High Court 2024No1823). The Constitutional Court wrote that deepfake videos in themselves become a record of sexual abuse and exploitation (Constitutional Court 2022Heon-Ga-8). The two sentences answered different questions. One was a place of reading the wording of the penal provision narrowly, and the other was a place explaining why virtual image regulation is necessary. Still, it is a fact that the two institutions wrote down in different words what deepfakes do to victims.

In the area of children, Korean law does not ask whether a real child was photographed. Instead, it asks whether they are clearly recognizable as children or adolescents, provisions punishing even drawings without real children are constitutional, and the remaining dispute is how far it is permissible to expand them.

Where to Draw the Line on Punishment — The Creator, the Distributor, the Possessor, and the Viewer

When it comes to adult victims, the place where the law diverges the most is when someone only creates the content without distributing it. A British preprint published in 2025 criticizes the fact that while the 2023 Online Safety Act criminalized the distribution of non-consensual sexual deepfakes, the act of merely creating them for sexual gratification or future crimes remains legal. The authors propose a strict liability model that focuses on the absence of consent instead of the defendant's intent (DOI: 10.20944/preprints202507.1645.v1, preprint). This is a description of British law at the time of its publication. A paper in a New Zealand law review recommends creating separate criminal provisions based on the harm inflicted on individuals and society (DOI: 10.26686/vuwlr.v56i2.10758). The Portuguese paper mentioned earlier suggests a new crime, stating that for adult victims, existing provisions on computer fraud, illegal filming, and defamation leave a gap where they fail to properly capture this harm. The same paper writes that caution should be exercised against overly expanding criminal law or approaching sexual offenses with moralism (DOI: 10.15388/teise.2025.134.4).

South Korea changed this stance in October 2024. The pre-amendment provisions applied by the preceding cases punished only those who edited images "with the intent to distribute them" (Article 14-2 of the Act on Special Cases Concerning the Punishment of Sexual Crimes, provision prior to the October 16, 2024 amendment — the wording is the same as the January 25, 2024 enforcement version). In the revised version enforced on October 16, 2024, that intent requirement was removed, meaning that creating such content against the subject's will, even without the intent to distribute, is punishable by imprisonment for up to 7 years or a fine of up to 50 million won. The penalty for distributors was also raised from up to 5 years to up to 7 years, and for those who distribute it using an information and communications network for for-profit purposes, the penalty changed from imprisonment for up to 7 years to imprisonment with labor for 3 years or more. Newly created Paragraph 4 punishes anyone who possesses, purchases, stores, or views such edited or replicated materials with imprisonment for up to 3 years or a fine of up to 30 million won (Article 14-2 of the Act on Special Cases Concerning the Punishment of Sexual Crimes, enforced on October 16, 2024). Composites of adult victims are now subject to punishment simply for possessing or viewing them. This amendment overlaps with the timing when reforms to strengthen legal provisions started rapidly after the 2024 crisis (DOI: 10.1080/01924036.2025.2596578). That abstract does not state which provisions were changed or how.

There is also a view that we should enforce existing rules rather than making new punishment provisions first. A 2024 paper that systematically reviewed empirical studies and analyzed policy and legal documents from various countries finds that much of the anxiety around deepfakes stems from alarmism rather than evidence. That does not mean they take the harm lightly. The authors conclude that deepfakes technology currently amplifies existing problems like non-consensual sexual images and misinformation rather than creating entirely new regulatory problems. Therefore, they state that properly supervising and enforcing existing rules and carefully considering necessary adjustments are important (DOI: 10.1177/14614448241253138). Whether punishment produces a deterrent effect is also still open. In the survey of 10 countries mentioned earlier, respondents in countries with laws targeting this issue also reported perpetration and victimization, and the authors read this as a signal that those laws are not sufficient to stop perpetration (DOI: 10.1145/3613904.3642382). Since it is a survey at a single point in time, it does not compare before and after the laws came into being.

In cases of illicit video materials as well, the basic guarantees of criminal trials apply unchanged. In another defendant's case charged with editing and distributing illicit video materials, the Supreme Court quashed the lower court judgments and remanded the case because the first and second instances proceeded with the trial without a state-appointed defense counsel. The reason is that since the defendant was detained for another case, it was a trial that could not be held without a defense counsel (Supreme Court 2024Do5566).

Whose Job Is It to Erase

Punishment targets the perpetrator. On the victim's side, getting spread images taken down is cited as the reason that drives them to seek help. A scoping review combining 81 studies on help-seeking by victims of image-based sexual abuse lists the desire for justice along with the deletion of non-consensually shared images as motives. These studies were conducted mainly in English-speaking Western countries (PMID 39427323). There is an audit study that measured deletion speeds by reporting channel on X (formerly Twitter). The researchers uploaded 50 AI-generated nude images, reporting half as "non-consensual nudity" violations and half as copyright infringements. Images reported under copyright were all taken down within 25 hours. Images reported as non-consensual nudity did not have a single one taken down even after over 3 weeks (DOI: 10.1145/3816990). This is the result measured with 50 images on a single platform. Under those conditions, copyright worked faster than the violation of the victim's consent. The researchers conclude that targeted legislation regulating the removal of non-consensual sexual images is necessary.

The US Take It Down Act mandates removal within 48 hours, while the UK Online Safety Act imposes prior safety duties on platforms. The author recommends a hybrid model combining the UK's substantive justice with the US's processing speed, while adding responsibilities that extend to cryptographic provenance standards and tool developers (DOI: 10.2139/ssrn.7530380, preprint). A paper examining Australia's eSafety Commissioner—dubbed the world's first online safety regulator—as a case study concludes that this agency's approach of receiving reports and deleting content has substantial problems (DOI: 10.1080/10383441.2025.2504791).

There is also the matter of the very place where the harmful material is hosted. In a study that found 88 public sites hosting AI-generated non-consensual sexual images, 38 of them dealt exclusively with synthetic images. The infrastructure for these sites, such as web hosting and content delivery networks, was provided by a few of the world's largest internet companies, and most were discoverable through search engines (DOI: 10.54501/jots.v3i4.348). Another study found roughly 35,000 fine-tuned models for deepfakes targeting real people across two public repositories. Since November 2022, downloads were close to 15 million, and 96% of those models targeted women (DOI: 10.1145/3715275.3732107). The proposal to hold the creating companies liable for product defects comes from here. Because users are hard to track down and platforms evade responsibility through Section 230 of the US Communications Decency Act, the idea is to hold generative AI developers who release dangerous tools without safety guards accountable for design defects (DOI: 10.2139/ssrn.6445584, preprint).

There are also the voices of the people who actually do this work in South Korea. In a design workshop involving 10 Korean activists and victim-survivors who have fought against deepfake sexual abuse, participants cited the ambiguity in classifying what a post is, the difficulty of gathering evidence, and how monitoring increases their workload and puts their personal safety at risk. They emphasized that the ownership of data and the responsibility of platforms must be discussed together (DOI: 10.1145/3772318.3790902). If harm accumulates at each stage of spreading and coming back (PMID 42745201), the removal pathways play a role alongside punishment clauses in determining the magnitude of the harm.

Political Deepfakes and Freedom of Expression

The axis of consent fits well in sexual deepfakes. The situation is different for a candidate's fake speech videos. This is because the subject the law deals with here is false political speech. As of 2024, 20 U.S. states have passed laws blocking the spread of fake photos, videos, and audio regarding candidates. A paper reviewing these laws summarizes that since the U.S. Supreme Court's Alvarez ruling, false political speech has received strong protection under the First Amendment. The authors find that most provisions such as prior restraint, campaign rules, parody and satire exceptions, and mandatory disclosure face substantial constitutional hurdles, and as alternatives, they propose digital watermarks by AI companies and labeling by distribution services (DOI: 10.2139/ssrn.4830972, preprint).

Counterarguments also emerge within the U.S. It is the argument that post-judgment injunctions blocking the reposting of election deepfakes judged to be defamation can pass the strictest constitutional review even where they are viewed as prior restraint, if they are narrowly tailored. The author distinguishes such orders from censorship, calling them a remedy to prevent proven falsehoods from distorting democratic choices (DOI: 10.2139/ssrn.6998321, preprint). The authors of the earlier U.S. perception experiment believe that bans on sexual deepfakes should receive the same constitutional treatment as bans on non-consensual sexual imagery, and that non-sexual deepfakes can probably be dealt with only through defamation law (DOI: 10.2139/ssrn.3781968, preprint).

The two types do not operate completely separately either. A 2020 paper examining the relationship between the two using Twitter data reports that sexual deepfakes and political deepfakes operate in similar ways to silence critical voices, and recommends addressing the inequality that causes this technology to disproportionately target women (DOI: 10.1080/23268743.2020.1757499). The legal grounds in this section are all American, and South Korea's election deepfake regulations are not in the data of this article.

Can You Tell with Your Naked Eye?

The other half of the common view is the belief that detection is the answer. A pre-registered behavioral experiment (N=210) put human eyes to a direct test. Participants failed to reliably spot deepfakes, and neither raising awareness of deepfakes nor offering financial rewards improved discrimination accuracy. The direction of errors also skewed to one side. More participants mistook deepfakes for real videos than saw real ones as fake, and participants overrated their own discernment. The authors conclude that people use the heuristic "seeing is believing" while overestimating their poor discernment, and that this combination makes them especially vulnerable to deepfakes (DOI: 10.1016/j.isci.2021.103364). These are the results of a single experiment within the conditions of its videos and participants. A narrative review published in a journal whose peer-review rigor is hard to verify scans the literature from 2020 through September 2026 from the perspective of legal evidence and summarizes it as follows. This review also notes that humans frequently fail at spotting deepfakes and are largely unaware of how bad they are at it. Context, repetition, prior beliefs, emotions, source attribution, and warnings shake accuracy and judgment bias. Machine detectors perform better than unaided humans in tests under the same conditions, but their accuracy drops when videos are compressed, made by unfamiliar generators, intentionally tampered with, or when the data distribution shifts. A score produced by a detector does not prove authenticity on its own (DOI: 10.33545/26649799.2026.v8.i9a.511). The authors' conclusion is that seeing and being convinced with one's own eyes cannot replace showing sources and undergoing verified analysis.

Provenance is a method that traces records of which camera recorded a video, what edits it went through, and how it spread, thereby confirming whether the video remains untouched. A 2021 article summarizing the activities of a human rights and citizen journalism network suggests that such foundations help people far away properly interpret on-the-ground videos. The same article notes risks as well. Provenance records can be used to target vulnerable witnesses, and videos lacking records due to privacy or technical access reasons may end up losing trust instead (DOI: 10.1177/14648849211060644).

The liar's dividend is the benefit where the widespread knowledge that fakes can be made allows someone faced with genuine evidence to claim it is fake and get away with it. A 2026 conceptual paper arguing without empirical work divides the harm into two parts. These are the harm of deception, where a fake is believed to be real, and the harm of cynicism, where the mere possibility of manipulation leads people to disbelieve the real. The author argues, synthesizing existing studies, that the harm of cynicism is greater and harder to fix. This is because it arises from the general awareness that fakes exist rather than any specific fake, and it persists even after retractions (DOI: 10.68854/mesw8r14). The legal review mentioned earlier also writes that while warnings heighten vigilance, they can produce a liar's dividend effect that disparages genuine material (DOI: 10.33545/26649799.2026.v8.i9a.511). A chapter from a Turkish book covering election cases in Slovakia, India, the United States, and Türkiye discusses how politicians also use this technology as a shield to cover up real scandals, and introduces findings that media literacy education can instead foster cynicism and institutional distrust (DOI: 10.58830/ozgur.pub1099.c4387).

Aside from a single experiment, the evidence in this section consists mostly of conceptual papers, narrative reviews, and a book chapter. The 2024 systematic review seen earlier notes that concerns about deepfakes often stem from alarm rather than evidence, and that there is still much we do not know (DOI: 10.1177/14614448241253138). Evidence stretching as far as claiming that nobody will be able to believe anything real anymore is absent from these materials. What is confirmed is that visual judgment is weak, and that there are multiple proposals to supplement that weakness with records and verification.

Factors cited by the Supreme Court to consider when judging synthetic media include the subject's actual age or identity, as well as the source or production history of the synthetic material (Supreme Court Decision 2024Do17801). These are facts that cannot be known by looking at the screen alone and must be verified through records.

What Has Not Been Decided

First is the edge of punishment. Since October 2024, South Korea also punishes simply making, possessing, or viewing false video materials of adult victims. This provision has never been judged by the Constitutional Court. Making and having drawings without real children is also punished, and in this area the Constitutional Court took the former Act on the Protection of Children and Youth against Sex Offenses Article 11 as the subject of review and saw it as constitutional. The court that sent that case doubted causality, and a paper from Portugal argues for excluding completely virtual materials from crimes. What to do with synthetic materials made only of adult victims varies by country. South Korea punishes them, and the United Kingdom as of the 2025 preprint point did not.

Second is what to call the harm. The Seoul High Court wrote that there was no direct sexual exploitation or abuse in composite photos, and the Constitutional Court wrote that deepfakes themselves become a record of sexual abuse and exploitation. The two descriptions answered different questions, but in the earlier case one branch of that judgment was used to separate a clause with a 5-year lower limit and a clause with a 5-year upper limit based on the provisions before amendment.

Third is whether the law reduces harm. Offense and harm have been reported even in countries with laws, and studies directly evaluating the effectiveness of countermeasures are rare. Deletion pathways in an audit study reacted faster to copyright than to consent.

Fourth is political deepfakes and freedom of expression. In the United States, the constitutional protection of false political speech is strong, so the logic of consent that works in sexual deepfakes does not carry over as is.

Fifth is authenticity determination itself. The human eye is weak, and machine detectors shake when conditions change. Provenance records are an alternative, but records can make real videos without records suspected, and a conceptual paper argues that the harm of skepticism—where people stop believing the real—arises from the very awareness that fakes exist and remains even after correction, so a provenance-based foundation is more important than detection ability (DOI: 10.68854/mesw8r14). The dividends of a liar insisting the real is fake exploit that skepticism.

If you view deepfakes only as a problem of authenticity, the problem should end the day detection technology wins. The law asks different questions about the same fake. U.S. federal law is tied to whether it is not yet distinguishable from the real by the evaluation of the paper cited in this article and to the intent to harm, and the South Korean Act on the Protection of Children and Youth against Sex Offenses asks whether it is clearly recognizable as a child or adolescent. British law and Article 14-2 of the South Korean Act on Special Cases Concerning the Punishment of Sexual Crimes created in 2020 ask for the consent of the subject. In the case followed by this article, whether it was a photo where synthetic traces were visible or a photo where they were hardly visible, punishment was handed down under that provision asking for the subject's intent. The remaining questions are how far to punish and what to leave as expression, and who bears the responsibility for deleting. You can also read along "Facial Recognition and Surveillance — What the Law Has Not Decided" and "AI and Copyright — Is Generative AI 'Learning' or 'Plagiarizing'?" in the same series.

Judgments

• Seoul High Court 2024No1823 (November 1, 2024) — Case involving commissioning the production of sexually explicit synthetic media made with the face of a 16-year-old junior — Ruling that synthetic media of a real child's face is not sexually exploitative material "in which children and adolescents appear," nor does it easily qualify as "an expression that can be clearly recognized," acquitting on those grounds in the reasoning, and finding the crime of editing false video material added in the appeal guilty (overturning the first-instance Seoul Western District Court 2024Gohap93, sentenced to 2 years and 6 months in prison, suspended for 3 years).
• Supreme Court 2024Do17801 (August 14, 2025) — Appeal of the above case — Presenting standards and judgment factors stating that photographs and deepfake videos synthesized with the face of a real child are not sexually exploitative material "in which children and adolescents appear," but can qualify as "an expression that can be clearly recognized," and maintaining the lower court's conclusion to dismiss the prosecutor's appeal (unanimous among participating Supreme Court justices).
• Constitutional Court 2023HeonBa263 (June 24, 2026) — Provisions punishing the production, distribution, and commercial sale of adult comics featuring characters that can be clearly recognized as children and adolescents — Ruled constitutional, stating it does not violate the principle of clarity, freedom of expression, or the principle of equality (unanimous among the justices).
• Constitutional Court 2022HeonGa8 (June 24, 2026) — Provisions punishing the commercial distribution and possession of the same type of comics, requested for constitutional review by the Jeju District Court — Ruled constitutional, stating it does not violate the principle of proportionality between responsibility and punishment or the principle of equality (unanimous among the justices). Describing deepfake videos and similar materials as constituting in themselves records of sexual abuse and exploitation.
• Supreme Court 2024Do5566 (July 11, 2024) — Case of editing and distributing false video material — Remanding the case on the grounds that if the defendant was detained in another case, it constitutes a mandatory defense case, making the lower court proceedings conducted without a state-appointed defense counsel unlawful (unanimous among participating Supreme Court justices).

Laws and Decrees

• Article 14-2 of the Act on Special Cases Concerning the Punishment of Sexual Crimes (Distribution, etc. of False Video, etc.) — Text before the amendment of October 16, 2024 (the wording applied to the court judgment case; the comparison version is the enforcement edition of January 25, 2024): Editing for the purpose of distribution, etc., up to 5 years; distribution, up to 5 years; distribution for profit, up to 7 years. Enforcement edition of October 16, 2024 (revised edition, wording identical to the current version effective October 2, 2026): Removal of the purpose requirement; editing, up to 7 years; distribution, up to 7 years; distribution for profit, at least 3 years; possession, purchase, storage, or viewing, up to 3 years or a fine of up to 30 million won (paragraph 4 newly established).

Source Papers

Grades (pre-publication public drafts, preprints) are indicated in the body as they appear.

• "From grassroots advocacy to AI governance: lessons from South Korea's 2024 deepfake sexual abuse crisis" (2025) — DOI: 10.1080/01924036.2025.2596578 — Following the 2024 South Korea crisis, non-consensual synthetic sexual imagery came to be recognized as a crime, prompting rapid legal strengthening, victim support, and AI governance reforms. Cross-border detection and cooperation remain barriers.
• "'A Deepfake Porn Plot Intended to Silence Me'" (2020) — DOI: 10.1080/23268743.2020.1757499 — While many deepfakes in circulation are sexual, public attention centers on political deepfakes. Both operate in similar ways to silence critical voices.
• "The Real Threat of Deepfake Pornography: A Review of Canadian Policy" (2021) — DOI: 10.1089/cyber.2020.0272 — Most online deepfakes are sexual, and the subjects' consent is rare. Canadian law provides no clear remedies.
• "Sexualized Deepfake Abuse as Image-Based Sexual Abuse: A Scoping Review" (2026) — PMID 42745201 — 54 papers. The misinformation and celebrity framing obscure everyday harm. Harm accumulates at each stage of the cycle. Victim-centered empirical evidence and intervention effect evaluations are rare.
• "Deepfake Privacy: Attitudes and Regulation" (2021, pre-publication draft) — DOI: 10.2139/ssrn.3781968 — Representative US sample. Labeling sexual deepfakes as fiction does not reduce judgments that they are wrong (unlike non-sexual ones). They are as harmful as revenge porn and slightly more blameworthy.
• "Image-Based Sexual Abuse" (2016) — DOI: 10.1093/ojls/gqw033 — Establishing the concept. Physical and mental illness, and the loss of dignity, privacy, and sexual self-determination merge into cultural harm.
• "Beyond 'Revenge Porn': The Continuum of Image-Based Sexual Abuse" (2017) — DOI: 10.1007/s10691-017-9343-2 — The revenge porn frame fragments legislation. It should be understood as a continuum with other sexual violence.
• "Image-based Sexual Abuse" (2020, research book) — DOI: 10.4324/9781351135153 — Surveys of over 6,000 people and more than 100 interviews across three countries. Includes deepfakes. The dominant motives are power and control.
• "Outcomes of image-based sexual abuse among young people: a systematic review" (2025) — PMID 40709226 — 12 papers involving ages 10 to 24. Psychological distress, bullying, and disruption to education and careers. Similarities to physical and sexual abuse.
• "Non-Consensual Synthetic Intimate Imagery: Prevalence, Attitudes, and Knowledge in 10 Countries" (2024) — DOI: 10.1145/3613904.3642382 — Over 16,000 people. 2.2% victimization and 1.8% perpetration. Reported even in countries with laws → insufficient for deterrence.
• "Dignity vs. Deception" (2026, pre-publication draft) — DOI: 10.2139/ssrn.7530380 — US Take It Down Act (indistinguishability, intent to harm, 48-hour removal) versus UK (consent-based, ex-ante safety obligations). A common tool gap in both countries.
• "Artificial Intelligence and Sexual Offences: An Analysis of Deepfake Pornography in Light of Criminal Law" (2025) — DOI: 10.15388/teise.2025.134.4 — Portugal. Only virtual materials using real children's images are criminalized. Adult victimization is a gap, and while new crimes are proposed, over-criminalization and moralism should be guarded against.
• "AI-generated image-based sexual abuse: A scoping review" (2026) — DOI: 10.1016/j.chbr.2026.101306 — 28 papers. Synthetic materials featuring identifiable children and fully synthetic materials are treated as different categories. (The review suspicion marks on the shortlist are judged by paper number page counts — the abstract is from the original work)
• "The Weaponisation of Artificial Intelligence (AI)… NCIDs" (2025, preprint) — DOI: 10.20944/preprints202507.1645.v1 — The UK Online Safety Act only penalizes distribution while making creation legal. Proposes a strict liability model focused on the absence of consent.
• "Fake Images, Real Harm: A Case for Criminalising Non-consensual Intimate Deepfakes" (2026) — DOI: 10.26686/vuwlr.v56i2.10758 — Recommends a separate criminal provision in New Zealand.
• "What we know and don't know about deepfakes" (2024) — DOI: 10.1177/14614448241253138 — Systematic review. Much of the anxiety is alarmism. It amplifies existing problems rather than creating new ones. Existing rule enforcement and cautious adjustment.
• "The Help-Seeking Behavior of Victims of Image-Based Sexual Harassment and Abuse: A Scoping Review" (2025) — PMID 39427323 — 81 papers. Help-seeking motives include definition and removal, while barriers include shame and fear of negative responses. Primarily English-speaking Western samples.
• "Reporting Non-Consensual Intimate Imagery: An Audit Study of Deepfakes" (2026) — DOI: 10.1145/3816990 — 50 images on X. Copyright reports resulted in 100% removal within 25 hours, while non-consensual nudity reports resulted in 0% removal after more than 3 weeks.
• "Online safety regulation of deepfake abuse: a case study on Australia's eSafety Commissioner" (2025) — DOI: 10.1080/10383441.2025.2504791 — Significant problems with the reporting and removal approach.
• "The Backbone of Abuse" (2026) — DOI: 10.54501/jots.v3i4.348 — 88 public sites (38 dedicated to synthesis). Major internet companies provide the infrastructure, and most are found through search.
• "Deepfakes on Demand" (2025) — DOI: 10.1145/3715275.3732107 — Roughly 35,000 public variation models, about 15 million downloads, with 96% targeting women.
• "Deepfake Liability" (2026, pre-publication draft) — DOI: 10.2139/ssrn.6445584 — Limitations of user and platform liability. Proposes products liability (design defects) for generative AI developers.
• "Deepfake, Real Harm: A Participatory Approach…" (2026) — DOI: 10.1145/3772318.3790902 — 10 South Korean activists and victim-survivors. Ambiguous categorization, evidence collection barriers, surveillance burden, and personal safety risks. Data ownership and platform responsibility require discussion.
• "The Right to Lie with AI?" (2024, pre-publication draft) — DOI: 10.2139/ssrn.4830972 — Laws in 20 US states. Strong protection for false political speech since Alvarez. Mostly constitutional barriers. Watermarks and labeling as alternatives.
• "Surviving Strict Scrutiny" (2026, pre-publication draft) — DOI: 10.2139/ssrn.6998321 — Injunctions issued after narrowly tailored rulings can pass strict scrutiny.
• "Fooled twice: People cannot detect deepfakes but think they can" (2021, single experiment) — DOI: 10.1016/j.isci.2021.103364 — Preregistered experiment N=210. Inability to discern, no effect from awareness-raising or monetary rewards, and a bias and overconfidence leaning toward authenticity. (Review suspicion marks are paper number pages — original work experiment)
• "Deepfake evidence and the reliability of perceptual judgement in court" (2026, narrative review) — DOI: 10.33545/26649799.2026.v8.i9a.511 — Human discernment is weak and uncorrected. Detectors are vulnerable to changing conditions, and scores cannot self-prove authenticity. The endowment effect of warnings. Provenance and verified analysis.
• "Deepfakes… authenticity infrastructure responses" (2021) — DOI: 10.1177/14648849211060644 — Provenance-based utility and risks (targeting vulnerable witnesses, declining trust in unrecorded video).
• "Synthetic Media and the Epistemics of Public Communication" (2026, conceptual paper) — DOI: 10.68854/mesw8r14 — Harms of deception and harms of skepticism. The author argues that the harms of skepticism are greater.
• "Deepfake Teknolojisi ve Siyasi Güven Krizi" (2025, book chapter) — DOI: 10.58830/ozgur.pub1099.c4387 — Election case studies. Introduces politicians using it as a shield and the counterproductive results of media literacy education.