Korea’s AI Food and Cosmetic Ads: 166 Cases, and a Label That Often Isn’t There
On 4 Oct 2026 Kyunghyang reported MFDS data to Rep. Lee In-young: 166 AI or suspected-AI unfair ads in Jan–Jul 2026, above last year’s 93. Of 83 general-food cases, 61 had no AI label. Fake-expert food-ad ban takes effect 27 Nov.
TL;DR — On 4 Oct 2026, The Kyunghyang Shinmun reported data the Ministry of Food and Drug Safety (MFDS) gave Rep. Lee In-young: 166 unfair ads from January through July 2026 that were made with AI or suspected of it, already above all of last year (93). General food 83, cosmetics 30, health functional food 27, quasi-drugs 21, medical devices 5. Of the 83 general-food cases, 61 (73.5%) had no “AI-generated” label. MFDS says it has no tool that decides whether an ad was AI-made. A ban on AI fake doctors and pharmacists in the Food Labeling and Advertising Act takes effect 27 Nov 2026.
Last checked: 4 Oct 2026 (KST). Axortex regulation timeline for readers outside Korea. This is not medical advice, not a product review, and not a guide to making or hiding an ad. Every figure below is from Kim Chan-ho’s Kyunghyang report published 08:18 KST on 4 Oct 2026, citing MFDS material submitted to Rep. Lee.
Why a Seoul ad count matters if you shop K-beauty or Korean food online
Korean cosmetics, health foods, and “quasi-drugs” (a Korean category tighter than ordinary cosmetics, looser than a full medicine) are sold into feeds that do not stop at the border. Kyunghyang says the flagged ads mostly claimed disease treatment or made a product look like a different kind of product. Among cosmetics, 25 of 30 could be mistaken for pharmaceuticals. Among quasi-drugs, 15 of 21 dressed industrial products up as quasi-drugs.
The disclosure problem is separate from the medical claim. On the 83 general-food unfair ads, staff eyes — not a detector — put 61 in the “suspected AI” pile because there was no AI-generated label. 22 had a label. If you are outside Korea and a short video recommends a Korean cream, snack, or health food, the frame may not say it is an ad, may not say the speaker is virtual, and the Korean regulator that counted these cases says it cannot technically prove the speaker was generated.
This page does not cover India’s food-ad rules or any other country’s advisory. It is one Korean dataset and the laws timed to it.
Dated timeline
| When (KST) | What Kyunghyang reported | Why it matters | Source |
|---|---|---|---|
| Nov 2025 | MFDS has run AI Cops since last November. It reads phrases in online ad images and flags violations. It does not decide whether the ad was made with AI. | The tool people may assume is an “AI detector” is a wording monitor. | Kyunghyang, 4 Oct 2026 |
| Jan–Jul 2026 | 166 unfair ads made with AI or suspected AI. General food 83, cosmetics 30, health functional food 27, quasi-drugs 21, medical devices 5. Already above the 2025 full-year total of 93. | Seven months beat a full prior year. The largest bucket is ordinary food, not cosmetics. | Same |
| Jan–Jul 2026 (YouTube slice) | Cosmetics: 20 of 30 found on YouTube. Quasi-drugs: 18 of 21 on YouTube. | For those two categories, YouTube is where MFDS’s catch clustered. The story does not give a YouTube split for food, health functional food, or devices. | Same |
| April 2026 | Amendments to the Food Labeling and Advertising Act, the Pharmaceutical Affairs Act, and the Cosmetics Act passed, banning ads in which AI-made fake doctors, pharmacists, and the like recommend products. The government is preparing new administrative-disposition standards. | Three statutes, one idea: a generated “expert” endorsement is not a legal recommendation. Kyunghyang states an enforcement date for only one of them (next row). | Same |
| 1 Jun 2026 | Korea Fair Trade Commission review guidelines in force: if an AI virtual person recommends or endorses a product, the ad must say the person is virtual. | A disclosure rule already running, aimed at virtual recommenders, not only at fake white coats. | Same |
| 4 Oct 2026 | MFDS submitted the Jan–Jul counts to Rep. Lee In-young (National Assembly Health and Welfare Committee). Kyunghyang published them at 08:18. An MFDS official: human review cannot perfectly separate carefully made AI ads. Of 83 general-food unfair ads, 61 had no AI-generated label and were classed as suspected by visual inspection; 22 had a label. | This is the public dataset. “Suspected” is a staff judgment, not a lab result. | Same |
| 27 Nov 2026 (scheduled) | The clause in the amended Food Labeling and Advertising Act that bans AI virtual-expert ads takes effect. Kyunghyang notes that applying it still requires a way to confirm the person was created by AI. | The food-ad ban has a date. The detector does not. | Same |
The 166, as attributed
| Category | Unfair ads, AI-made or suspected AI (Jan–Jul 2026) | Extra split Kyunghyang actually printed |
|---|---|---|
| General food | 83 | 61 (73.5%) had no “AI-generated” label and were suspected by staff eyes; 22 had a label. No YouTube count given. |
| Cosmetics | 30 | 25 of 30 could be mistaken for pharmaceuticals. 20 of 30 were on YouTube. |
| Health functional food | 27 | No further split in the story. The photo caption says MFDS and Lee’s office supplied examples of AI images, voices, and videos of experts such as doctors and pharmacists related to health functional foods. |
| Quasi-drugs | 21 | 15 of 21 presented industrial products as quasi-drugs. 18 of 21 were on YouTube. |
| Medical devices | 5 | No further split. |
| Total | 166 | Compared with 93 for all of last year. |
Do not add the 61 and the 25. They answer different questions: missing labels on food, and drug-like claims on cosmetics.
Kyunghyang also described the kind of creative MFDS had on file, without treating it as a script to copy: an AI-made figure presented as director of a Gangnam dermatology clinic promising a teenage-acne cure, and a white-coat figure promoting rhinitis treatment. The paper’s opening scene is a shopper, identified only as “A,” who bought after a short social-media video the paper calls a Reel: no “ad” label, a weight-loss story, a woman who was a virtual person. Axortex is not repeating a testimonial, a brand, or a method.
Both sides
Regulators (as stated)
- Count, then admit the limit. MFDS put 166 cases in front of a lawmaker and, through an official quoted by Kyunghyang, said: “Because it is human work, it is impossible to perfectly distinguish meticulously produced AI ads.” The ministry has no separate technology to tell whether an ad was produced with AI.
- Labels plus eyes. Staff use voluntary “AI-generated” labels from advertisers or platforms, and their own inspection. That is how 61 food ads became “suspected” and 22 stayed “labeled.”
- AI Cops is not that technology. Since November 2025 it has read phrases in ad images for violations. It does not judge origin.
- Rules on a calendar. April: three acts amended against AI fake doctors, pharmacists, and similar recommenders, with administrative-penalty standards still being prepared. Since 1 June: KFTC says a virtual recommender must be identified as virtual. On 27 November: the food-ad statute’s fake-expert ban takes effect. Kyunghyang says that ban, too, needs a way to confirm the person was AI-made.
Advertisers, platforms, and the lawmaker’s ask
Kyunghyang does not quote an advertiser or a platform. What the same article shows on that side of the table:
- Some advertisers or platforms already label. Twenty-two of the 83 general-food unfair ads carried an AI-generated label. The label did not make the ad fair; it only marked origin. The other 61 did not carry it, and still entered the total as suspected.
- The posts sit on platforms. For cosmetics and quasi-drugs, YouTube held most of the cases MFDS logged (20/30 and 18/21). The paper’s scene-setting also points at short social video, including what it calls Reels, sometimes with no ad label.
- Lee’s proposal is aimed at platforms, not at another round of manual clicks. As quoted by Kyunghyang: “As AI-generated video and audio become more sophisticated, there are limits to responding solely by checking ads one by one.” He wants a system that prioritizes high-risk ads by linking platforms’ AI labels and source information and tracking repeated postings.
Axortex reading: regulators have a seven-month count, a phrase-scanning tool, and two disclosure or ban dates (1 June already in force; 27 November still ahead for the food-ad fake-expert clause). They do not, on their own account, have a way to prove an unlabeled video was generated. Platforms and advertisers are the ones who either attach the label or do not. Until 27 November, a missing label on a food ad is a detection problem MFDS has already described, not yet the fake-expert ban that statute will switch on.
Still unresolved
- “Suspected” is not “proven generated.” For 61 general-food ads, the classification is visual inspection. MFDS says careful AI ads cannot be perfectly separated by people.
- No YouTube share for three categories. Only cosmetics and quasi-drugs have a platform split in this story.
- Enforcement dates for two of the three April laws are not in the article. Pharmaceutical Affairs Act and Cosmetics Act amendments passed in April. Kyunghyang gives 27 Nov 2026 only for the Food Labeling and Advertising Act’s virtual-expert clause.
- Penalty standards are “being prepared,” not published in this report.
- No brand list, no charged individuals. The shopper is “A.” The Gangnam “director” is described as AI-made, not named as a real doctor. This page does not add names.
- Health claims in the ads are allegations of unfair advertising, not findings about any real product’s effect. Nothing here is guidance on acne, rhinitis, weight loss, or what to buy.
FAQ
Does 166 mean 166 confirmed deepfakes?
No. Kyunghyang’s phrase is unfair advertising made with AI or suspected of having been made with AI. On general food, 22 had an AI-generated label and 61 were suspected by staff because the label was absent.
Can MFDS tell a careful AI ad from a real person?
An MFDS official told Kyunghyang it cannot be done perfectly, because the review is human work, and the ministry has no separate technology for the question. AI Cops does not fill that gap; it reads phrases.
What actually changes on 27 November?
The amended Food Labeling and Advertising Act clause banning ads in which an AI virtual expert recommends a product takes effect that day, according to Kyunghyang. The paper says using the ban still requires confirming that the person in the ad was created by AI. The KFTC virtual-person disclosure guideline has been in force since 1 June 2026.
Is this a reason to avoid Korean cosmetics?
This article is not a safety rating. The cosmetics line in the dataset is about advertising: 25 of 30 flagged cosmetics ads could be mistaken for pharmaceuticals, and 20 of 30 were on YouTube. That is a claim-and-label problem, not a test of a formula.
Sources
- Kim Chan-ho, The Kyunghyang Shinmun, “even the Ministry of Food and Drug Safety was fooled by ‘AI-generated content,’” 4 Oct 2026, 08:18 KST — MFDS data submitted to Rep. Lee In-young; all counts, dates, and quotations in this piece.
Image: “Woman reaches for a cosmetic product in a display” by Shixart1985, CC BY 2.0, via Wikimedia Commons. Shelf photo for illustration only; not an ad named in the MFDS data.
← Back to all posts