South Korea identified 166 illegal advertisements made with AI, or suspected of being made with AI, between January and July this year, already exceeding the 93 cases found in all of last year. The Ministry of Food and Drug Safety said it does not have a technical tool to determine whether a person shown in an advertisement is AI-generated, leaving a large part of enforcement to human reviewers.
General food ads accounted for the largest share
Kyunghyang Shinmun reported exclusively on Oct. 4 that the 166 cases spanned all product categories. General food ads were the largest group at 83 cases, followed by cosmetics with 30, health functional foods with 27, quasi-drugs with 21 and medical devices with 5.
The ads commonly promoted product efficacy or disguised one class of goods as another. One example cited in the report involved an AI-generated “director” of a dermatology clinic in Gangnam claiming a product could cure adolescent acne. Among the 30 cosmetics cases, 25 made the products appear to be drugs. Among the 21 quasi-drug cases, 15 involved ordinary industrial products presented as quasi-drugs.
YouTube appeared repeatedly in the enforcement data. Twenty of the cosmetics cases and 18 of the quasi-drug cases were found on the platform.
Why companies use fake doctors and pharmacists
Under current South Korean rules, doctors and pharmacists are already barred from guaranteeing or recommending food products in advertisements. The report says some companies responded by creating virtual figures that are not real doctors or pharmacists.
In one case disclosed by the ministry in June, a middle-aged man in a video who introduced himself as the director of a plastic surgery clinic was AI-generated, while the product itself was an ordinary processed food. The product sold 650,000 units over nine months and generated KRW 8.1 billion in revenue. The business was referred to prosecutors on suspicion of violating the Food Labeling and Advertising Act.
Consumers struggle to tell, and so do regulators
The report described the case of a woman identified as Ms. A, who bought a product after watching a short social media video in which a woman shared a weight-loss story. She believed the speaker was an influencer, but the figure was actually an AI-generated virtual person. She said the facial expressions and tone sounded natural enough that she believed it.
The Ministry of Food and Drug Safety said it has no separate technology to determine whether an ad was produced with AI. A related official said, “Because this is something made by people, it is impossible to perfectly distinguish sophisticated AI advertisements.”
In practice, the ministry relies on two routes: AI-generation labels added voluntarily by advertisers or website operators, and visual review by human monitors.
More than 70% of general food cases were identified by human review
In the general food category, which had the highest case count, 61 of 83 advertisements had no AI-generation label and were classified as suspected AI-generated by human reviewers. That equals 73.5%. Only 22 carried a label, which means fewer than 30% were proactively disclosed.
The report points to a clear limit in those figures: ads with no label and no obvious visual signs would not appear in the statistics at all.
The ministry’s AI Cops system, scheduled to operate from November 2025, does not close that gap. The tool reads text embedded in advertising images and flags problematic wording, but it does not determine whether the people shown in ads were created with AI. As the report put it, the system may have “AI” in its name, but its actual job is to find words.
The legal ban is in place, but enforcement starts with detection
On April 23, South Korea’s National Assembly passed amendments to the Food Labeling and Advertising Act, the Pharmaceutical Affairs Act and the Cosmetics Act to prohibit ads in which AI-generated fake doctors or fake pharmacists recommend products. The ban is set to take effect in late November.
Separately, the Korea Fair Trade Commission began requiring from June 1 that advertisements using AI virtual figures to recommend products clearly disclose that the figure is virtual. On Aug. 7, the Ministry of Food and Drug Safety also preannounced revisions to enforcement rules. Under the proposal, if a food advertisement uses an AI-generated figure that makes consumers believe the person is an expert such as a doctor, pharmacist or university professor, a first violation would bring a 15-day business suspension, and a third violation could bring a suspension of up to two months.
The report contrasts those penalties with the revenue in the ministry’s June case. One business generated KRW 8.1 billion over nine months, while the proposed sanction for a first offense is a 15-day suspension. But before any sanction can be imposed, authorities still need to identify that the person in the ad is fake.
AI Basic Act does not directly impose a labeling duty on advertisers using AI tools
South Korea’s AI Basic Act took effect on Jan. 22 and requires images, videos and audio generated by generative AI and distributed externally to be labeled as AI-generated. But the report says the legal duty applies to companies that directly provide AI products and services to users, not to users who employ AI as a tool to make finished content.
Under that reading, advertisers using AI to create ads do not automatically bear a labeling duty under that law alone.
A separate bill has been proposed that would require individuals uploading self-made AI-generated content to label it and would place management responsibilities on website operators. As of mid-July, however, deliberation had been repeatedly delayed because the National Assembly had stalled.
YouTube and South Korean portal Naver currently rely on their own platform rules to recommend disclosure, but the report says those measures are neither fully implemented nor easy to verify.
Experts suggest linking platform labels and provenance data
As AI-generated images and audio become more refined, the report says a system that depends on people checking ads one by one has clear limits. Some experts argued that a more efficient approach would be to connect platform-level AI labels and provenance information, then prioritize screening for high-risk advertisements.