Chapter 1. AI Ad Copy and Substantiation

Learning objectives

• Distinguish the four types of unfair labeling and advertising prohibited by the Act on Fair Labeling and Advertising, and identify each type in AI-written copy.
• Distinguish whether each sentence is a factual claim or a subjective expression to select the subject of substantiation.
• Explain the requirements that substantiation data must meet (method, independence, direct relationship) and judge whether given data satisfies the requirements.
• Determine the conditions under which superlative and exclusive expressions, comparative expressions, and price and discount expressions can be used.
• Design a review process to filter out figures and sources fabricated by AI through comparison with the original source.
• Rewrite ad copy to match the scope verified by substantiation data.

Introduction

A junior marketer at a side dish subscription service asked generative AI, "Write detail-page copy introducing our regular side dish delivery." The draft that appeared in three seconds was smooth: 'No. 1 Nationwide Side Dish Subscription', '92% Resubscription Rate', 'Balanced meal plan designed by a nutritionist', 'Just like Mom's home cooking'. Because the copy was good, the marketer was about to post it as was, but stopped at a single question from the team leader: "Who tallied that No. 1, and when?"

There was no one who could answer. The company had no ranking tally, the resubscription rate was actually 71%, and the meal plan was designed not by a nutritionist, but by a menu planner. The AI was not trying to lie. It had simply learned the 'form' of writing known as a detail page well, and because rankings, figures, and experts usually appear in such text, it merely filled in something plausible. However, the moment they appear in an advertisement, those sentences become the company's claims, and the company must be able to prove those claims.

This chapter deals with the rules of that proof. First, we look at what the Act on Fair Labeling and Advertising considers unfair advertising, and by what criteria it judges the 'likelihood' of consumers being misled. Next, we examine the scope of substantiation liability and the requirements for substantiation data through the Fair Trade Commission (FTC) notifications. Then, we address superlative, comparative, price, patent, and AI technology expressions that frequently cause problems in practice one by one, and finally learn the procedure for classifying and rewriting AI drafts on a sentence-by-sentence basis.

Key concepts

Labeling, advertising, and the responsible business entity

Under the Act on Fair Labeling and Advertising, labeling refers to characters and shapes written or affixed by a business entity to the containers or packaging of goods, business place displays, title deeds, etc., as well as packaging representing the characteristics of goods, and advertising refers to widely informing consumers of or presenting matters concerning goods through newspapers, broadcasting, telecommunications, etc. (Article 2). Online mall detail pages, search advertisements, social media advertisements, and app push notifications all fall within the scope of advertising.

The obligations under this Act lie with business entities, etc., namely the business entities and business entity organizations that advertise. It does not matter who wrote the sentences. Whether written by a copywriter, an agency, or generative AI, the company that released those sentences as an advertisement for its own goods bears the responsibility. Article 3 prohibits not only directly engaging in unfair labeling and advertising, but also "causing other business entities, etc. to do so." Therefore, advertising entrusted to an agency or influencer also does not escape the advertiser's responsibility.

Four types of unfair labeling and advertising

Article 3, Paragraph 1 cites four types of labeling and advertising that present a likelihood of deceiving or misleading consumers and a likelihood of undermining fair trade order. Article 3 of the Enforcement Decree defines the meaning of each.

The Examination Guidelines on Deceptive Labeling and Advertising define 'Omission' as not disclosing important facts from the beginning, and 'Concealment' as displaying them in a manner difficult for consumers to realistically recognize, such as through excessively small print or a fleeting duration. Because AI is adept at summarizing benefits concisely and attractively, it frequently creates deceptive errors that omit conditions and restrictions.

How consumer misleadingness is judged

The Designated Notification cites consumer misleadingness and harm to fair trade as criteria for determining unfair labeling and advertising. There are five judgment principles that practitioners must know:

1. The standard is an ordinary consumer with average care. However, an advertisement targeted at a specific group may be judged based on an average person of that group, and the Designated Notification cites doctors, patients, and children as examples.
2. The standard is the impression received by consumers. The intention of the business entity is not a criterion for judgment.
3. The overall and ultimate impression is examined. Even if each individual sentence is true, it becomes problematic if the overall impression created by arrangement and emphasis differs from the truth.
4. It is sufficient that there is a tendency and possibility to mislead, and proof that consumers were actually deceived is not required.
5. Obvious Exaggeration that anyone can recognize or socially accepted advertising expressions (examples from the Designated Notification: 'Prime countryside location', 'Top teaching staff'), as well as subjective judgments such as taste, feel, appearance, and smell, are difficult to view as having misleadingness.

Harm to fair trade considers the likelihood that an advertisement interferes with consumers' rational purchasing decisions. The Designated Notification explains that when matters important to purchasing decisions—such as efficacy and quality, health and Safe conditions, price and Withdrawn rights, and economic interests between advertisers and Recommendation endorsers—are subjected to Concealment or Omission, there is a high possibility that harm will be recognized.

Distinguishing factual claims from subjective expressions

The Operating Notification on Advertising Substantiation defines matters related to facts as "content claimed in labeling or advertising whose truth or falsity can be judged through reasonable grounds or objective data." Put simply, it means "sentences whose correctness or falsity can be determined with data."

There are also expressions on the boundary. 'The best taste' is close to a subjective judgment about taste, but 'No. 1 in taste evaluation' claims the result of someone's evaluation and is therefore a factual claim. Judge not by adjectives, but by whether "it contains a claim that can be determined by data."

The structure of substantiation liability

Article 5 establishes the framework for substantiation.

1. Business entities, etc. must be able to substantiate matters related to facts among the labeling and advertising they have conducted (Paragraph 1). This means that rather than preparing data after receiving a request, they must already be able to prove them at the time of advertising.
2. When there is a likelihood of violation, the Fair Trade Commission may request the submission of data by specifying details (Paragraph 2).
3. A requested business entity must submit within 15 days, and the FTC may extend the period if it recognizes justifiable grounds (Paragraph 3).
4. If advertising continues without submission, the FTC may order the suspension of advertising until the data is submitted (Paragraph 5).

The revision of the Operating Notification on Advertising Substantiation, which took effect on September 3, 2026, specifically lists reasons for extension—such as natural disasters, mergers and acquisitions, rehabilitation or bankruptcy proceedings, seizure of account books and evidentiary documents, and severe business disruption due to fire or disaster—and permits extensions only up to 15 days from the date such reason ceases to exist. Failure to submit within the period subjects the entity to administrative fines, and if advertising is being conducted at that time, it becomes subject to a suspension order. Circumstances such as "please wait until test results are available" are not among the reasons enumerated by the notification.

What constitutes substantiation data

The Operating Notification on Advertising Substantiation requires all three general requirements for substantiation data recognized as reasonable grounds to be met.

1. Method: Must be one of test results, survey results, expert opinions, academic literature, or other scientific and objective data.
2. Procedure: Must be prepared through objective and scientific procedures and methods.
3. Direct relationship: The content of the data must have a direct relationship with the advertising claim.

Article 4 of the Enforcement Decree stipulates that tests and surveys for substantiation must be conducted through objective methods generally recognized academically or in the industry, and by institutions established under statutes or institutions that are managed independently from the business entity. An institution operated by a business entity or its affiliated company is not an independent institution. However, exceptions are allowed if circumstances such as the absence of an independent institution, protection of trade secrets, or excessive costs are recognized.

The requirements by type of data are as follows.

Direct relationship: Does the data really prove the claim?

Even if data exists, if it does not have a direct relationship with the advertising claim, it does not constitute substantiation. The Operating Notification on Advertising Substantiation provides the following examples:

• Even if an efficacious ingredient is present, if the content falls short of the minimum standard required to produce efficacy, data on ingredient efficacy does not serve as grounds for product efficacy.
• Claiming the effectiveness of an antibacterial baby bottle while submitting test results only for the material, rather than the environment in which the baby bottle is used.
• When the performance in a test report is not realized in the finished product actually distributed.
• Claiming that 'All ingredients are organic' while proving only some ingredients is only partially related.

Generative AI easily connects general knowledge that "copper has an antibacterial effect" with the product claim that "our pillow cover is Bacteria-free." Reviewers must find the gap between the two sentences, namely the point where facts about ingredients, materials, and laboratories cross over into claims about finished products and actual use.

Superlative and exclusive expressions

The Designated Notification considers unfair any advertisement that misleads consumers into believing that a product is remarkably superior or advantageous compared to competitors through exclusive, absolute expressions such as 'Maximum', 'Best', 'First', 'Top', and 'Only'. However, they may be used if the business entity clearly proves them or they conform to facts based on objective data and do not cause damage to competitors or consumers.

In practice, three categories are distinguished:

1. Factual superlatives: Claim verifiable facts such as 'Korea's first', 'No. 1 in Sales', and 'Longest battery life'. There must be objective data specifying the period, scope, and compiling organization.
2. Evaluative superlatives: Claim the results of someone's evaluation, such as 'No. 1 in satisfaction'. The Operating Notification on Advertising Substantiation exemplifies ranking expressions as primary targets for substantiation requests.
3. Advertising rhetoric: Rhetoric with neither a comparison target nor a fact, such as 'For your best day'. This is close to socially accepted advertising expressions referred to in the Designated Notification.

Superlatives in AI drafts mostly take the form of No. 1 or No. 2. Set "remove if there is no supporting data" as the default, and if data exists, disclose the scope close to the copy.

Comparative expressions

Among the general principles of the Examination Guidelines on Comparative Advertising, there are three basic requirements for comparative advertising: comparison must provide useful and accurate information to consumers; the subjects and criteria of comparison must be clear and the content and methods appropriate; and it must be substantiated by the results of tests and surveys conducted through objective methods by institutions established under statutes or institutions independent from the business entity. The fourth item of the general principles adds that the examples in the guidelines are not exhaustive, and even if an advertisement falls under an example, it may not be an unfair comparison if there is no likelihood of misleading consumers or likelihood of undermining fair trade.

Even if a competitor's name is not explicitly used, if consumers can identify which company it is through metaphors or similar symbols, it is regarded as comparative advertising. The idea that writing merely 'Compared to competitors' makes it acceptable is mistaken.

Price and discount expressions

AI naturally attaches conventional phrases such as 'Regular price', 'Up to 70%', and 'Today only' to discount banners. The Designated Notification sets specific standards for price display.

• The previous transaction price is the price at which the business entity actually sold the same product for a considerable period recently (approximately the past 20 days), and if the price fluctuated during that period, it is the lowest price among them. Presenting a price at which the product was never actually sold as the 'Regular price' as if it were the company's previous transaction price to calculate a discount rate is unfair. Cases where a manufacturer selling through multiple dealerships uses the manufacturer's suggested retail price (MSRP) as the baseline, or where opening commemorative discounts use the market price as the baseline, must follow separate requirements and the obligation to specify criteria set forth in the notification.
• It is unfair to display discounts on some products or at some stores as if they apply to most, or to use Exaggeration to make it seem as though most products have a high discount rate when only a small portion do.
• Displaying items as if they are sold only for a limited period or quantity unlike the actual situation, or leaving advertisements intact after a discount event has ended, is also unfair.

When reviewing discount banners, place sales records (recent prices), discount rate distributions by item, actual event duration, and inventory side by side, and check them against the copy.

Patent, certification, and AI technology expressions

Patents. The Designated Notification cites as an unfair example advertising efficacy based solely on the fact of a patent, even though patent registration itself does not recognize performance, efficacy, or safety. 'Patented structure' is a fact, but the latter part of 'Patented structure reduces knee strain by 40%' must be substantiated separately.

Certifications and marks. The Operating Notification on Advertising Substantiation cites certification marks regarding safety, environment, and new technology as primary targets for substantiation requests. Expressions that exceed the scope of certification (which items, when, and for which products) must be avoided.

AI technology expressions. The revision of the Operating Notification on Advertising Substantiation in September 2026 included expressions utilizing new technologies, such as "safer with artificial intelligence (AI) technology," in the examples of primary targets for substantiation requests. Phrases such as 'AI analyzes and recommends the perfect fit' and '95% AI accuracy' often claim not only the existence of a feature, but also its effect. Verify on what data and in what environment the accuracy figure was measured, and whether those conditions are identical to consumers' actual conditions of use.

Types of factual errors created by generative AI

Errors created by AI in advertising drafts generally fall into six types:

Grounds do not exist simply because AI cited a source. Generative AI can create non-existent titles and author names in plausible formats. A source becomes grounds only after a human opens the original text and verifies the figures and conditions.

Finding expressions requiring substantiation: Sentence classification procedure

When an AI draft is received, classify sentences in the following order:

1. Break the draft down into sentence units and assign numbers.
2. Mark whether each sentence contains a factual claim. Numbers, rankings, superlatives, comparisons, efficacy and performance, certifications and patents, evaluations by experts, organizations, or consumers, prices and discount rates, and Limited time and quantity are signals.
3. For each sentence containing a factual claim, write down the type of supporting evidence needed (sales records, test reports, survey reports, certificates, etc.).
4. The person in charge of evidence attaches data and notes the subject, conditions, and period of the data.
5. If there is no data, remove it; if data exists but the scope differs, rewrite it through the next procedure.

Subjective expressions are excluded from substantiation, but still undergo review for brand tone and rights infringement (Chapter 5).

Rewriting to match data

The principle of rewriting is "only what the data confirmed, within the scope confirmed."

Also check that the rewritten copy does not inadvertently expand the scope. Changing a comparison with one's own previous model to 'Best in the industry', or changing a fabric test to 'Bacteria-free', is not correcting, but creating a new error.

Going deeper

Why qualifying statements are not a solution

Qualifiers such as 'Individual results may vary' and 'Actual effects may vary depending on the environment' are frequently used in advertising practice. However, qualifiers do not change the claim itself. Because misleading potential is judged by the overall and ultimate impression, if '99% removal' is prominently displayed on the first screen and a qualifier is attached in small print at the very bottom, the impression received by consumers is still '99% removal'. The Review Guidelines on Deceptive Labeling and Advertising define displaying information in overly small text or for a brief period so that it is practically difficult for consumers to perceive as 'Concealment'.

A qualifier is meaningful when the claim is already substantiated, and the qualifier discloses the conditions of that substantiation close to the claim in a manner recognizable to consumers. In other words, a qualifier is a tool to accurately inform the scope of a substantiated claim, not a tool to revive an unsubstantiated claim.

Division of roles when using AI as a review tool

Generative AI can also be used for review. AI is fast and meticulous when finding and flagging numbers, rankings, superlatives, and comparative expressions across hundreds of copy phrases. However, the task of judging whether the flagged sentences are accurate cannot be left to AI. This is because the basis for judgment lies in materials that AI cannot view or lacks the authority to verify, such as the company's sales records, test reports, and survey reports.

The recommended division of roles is as follows. AI handles detection (flagging candidate factual claims) and organization (proposing the required types of evidence), while humans handle cross-checking (comparing materials and copy) and decisions (retain, modify, delete). Avoid structures where the same model judges copy written by the same model as "having no issues." A structure in which a system reviews its own results easily repeats the same misconceptions.

What the 2026 revised Operating Notification on Substantiation changed

The revision that took effect on September 3, 2026 gives two signals to AI marketing practice. First, expressions indicating the utilization of new technologies such as AI were included in the examples of primary targets for substantiation requests. As 'AI-tailored' and 'AI analysis' have become standard marketing qualifiers, if those terms imply effectiveness or safety, they must be substantiated with data. Second, the grounds for extending the submission period have been specifically narrowed. The approach of releasing an advertisement first with the expectation that materials "Can be prepared later" has become more dangerous. The notification also provided a checklist in an annex so that business entities can check substantiation methods and judgment criteria in advance.

At work

Soundgrain: Three sentences on the wireless earphone product details page

When launching a new model of wireless earphones, the audio equipment brand 'Soundgrain' created a product details page using generative AI. The review manager broke down the draft according to the sentence classification procedure and flagged three sentences containing factual claims: ① 'Industry-longest 40-hour battery', ② '97% sound quality satisfaction', ③ 'Snug and comfortable fit in the ear'.

① was a factual superlative. When the person in charge of supporting documentation checked, the 40 hours was a value measured by the in-house laboratory with the charging case included, volume at 50%, and noise canceling turned off. Among competing products, there was also a model that advertised 45 hours including the case. The in-house laboratory was not a testing institution independent of the business entity. The manager removed 'Industry-longest' and proposed withholding the battery life figure from the product details page until the results from an independent testing institution were available. The team leader suggested, "Since it is similar anyway, let's change it to 'Industry's highest level'," but the manager did not accept it, stating that the expression was also an exclusive expression and required the same substantiation.

The basis for ② was a survey directly conducted by the company among 200 members of a tester group prior to launch. The tester group consisted of people who received the product free of charge. Citing the survey result requirements (independence, sample representativeness) of the Operating Notification on Substantiation, the manager determined that this figure could not be used. The marketing team countered that changing it to '97% of testers satisfied' would be completely true to fact, but the manager viewed that the copy still ultimately gave the impression that the product's sound quality was superior, and that if the fact that the responses came from people who received free products was omitted, a likelihood of misleading would remain. The conclusion was to plan a survey by an independent institution after launch and not use the figure until then.

③ was a subjective expression, so it was not subject to substantiation. However, the second version from the AI, which included the phrase 'Good for ear health', was not adopted because it changed into a health efficacy claim. In the review sheet, the sentence number, classification, supporting documentation, decision, and reasons for the decision were recorded, along with the discarded alternatives (Industry's highest level, 97% of testers).

SoftMat: 'Korea's first eco-friendly material' and '99% slip prevention'

In the AI draft for the yoga mat brand 'SoftMat', 'Korea's first sugarcane-material yoga mat' and '99% slip prevention' were included. The marketer stated that both were fine because data from the company's research laboratory existed.

The review manager checked 'First' first. What the company had was only the release date of its own product, and there was no data showing that mats made of the same material did not exist in Korea before that. Upon searching online malls, there was an imported product that had advertised using a similar material two years earlier. Because it was an exclusive expression whose facts were unverified, it was deleted. The marketer suggested phrasing it ambiguously as 'Practically Korea's first', but this was rejected because it gave the same impression.

The basis for '99% slip prevention' was a friction test measured by the company's research laboratory on a dry mat. There was no data for wet conditions where sweat dripped during yoga. The manager noted two issues. First, because the company's research laboratory was not a testing institution independent of the business entity, it was necessary to first determine whether any exception under Article 4 of the Enforcement Decree applied. Second, broadening the results from dry conditions to a general claim of 'Non-slip' exceeded the test conditions. The conclusion was to commission an authorized testing institution to conduct tests under both dry and wet conditions, and until the results were available, to use expressions without factual claims, such as 'Grip that highlights surface texture'. The suggestion to add 'May vary depending on usage environment' at the very bottom of the product details page was not accepted on the grounds that a qualifier cannot substantiate a claim.

Common mistakes

1. "The accuracy of copy written by AI is the AI company responsibility." Responsibility for advertising lies with the business entity that conducted the advertising. The terms of service of an AI service cannot shift responsibility toward consumers or the Fair Trade Commission. Regardless of the tool used to write the sentences, the company that advertised must substantiate them.
2. "Adding a qualifier at the bottom reduces the burden of substantiation." Misleading potential is judged by the overall and ultimate impression, and small print can constitute Concealment. Qualifiers are useful only for informing the conditions of a substantiated claim.
3. "If figures are removed and replaced with words, it is not a factual claim." Even without numbers, claiming an effect, such as 'Reduces knee strain', makes it subject to substantiation. It can even become a stronger claim, as in 'Industry's highest level'.
4. "Since data exists, expression is unrestricted." The data must be directly related to the advertising claim. Material tests, laboratory conditions, and results from other subjects serve as evidence only within those bounds.
5. "A survey directly conducted by the company is also a survey result." Surveys for substantiation must be conducted by an institution independent of the business entity, using a representative sample and appropriate questions. An in-house survey distributed to free trial users can hardly satisfy this requirement.
6. "As long as the competitor's name is not used, it is not comparative advertising." If consumers can identify which company it is, it is comparative advertising, and if it is compared with a subject that is not equivalent according to common commercial practice, creating a likelihood of misleading consumers, it can constitute unfair comparison.
7. "Since there is a patent, efficacy is also proven." Patent registration is not an endorsement of efficacy. The fact of patent registration and efficacy claims are evaluated separately.
8. "If the Fair Trade Commission requests materials, they can be prepared then." Substantiation must be possible at the time of advertising; the submission deadline after a request is 15 days, and the grounds for extension are narrow.

Summary at a glance

Check the official docs

• Act on Fair Labeling and Advertising: Article 2 (Definitions), Article 3 (Prohibition of Unfair Labeling and Advertising), Article 5 (Substantiation), Article 7 (Corrective Measures), Article 9 (Penalty Surcharges), Article 10 (Compensation for Damages).
• Article 3 of the Enforcement Decree of the Same Act: Meanings of the four types. Article 4 specifies the methods of testing and investigation for substantiation and the requirements for independent institutions.
• Operating Notification on Substantiation of Labeling and Advertising Contents: Examples of subjects of substantiation (including expressions of AI new technologies), submission period and grounds for extension, criteria for each type of data.
• Notification on Types and Criteria for Unfair Labeling and Advertising Activities: Criteria for determining misleading potential and impairing potential, price, quality and performance, exclusive expressions.
• Review Guidelines on Comparative Labeling and Advertising: Appropriateness of comparison subjects, standards, and methods, and examples.
• Review Guidelines on Deceptive Labeling and Advertising: Definitions and evaluation methods of Concealment, Omission, and reduction.

Self-check

1-1. State the four types of unfair labeling and advertising prohibited by the Act on Fair Labeling and Advertising, and provide one typical example of how each type appears in a generative AI draft.

1-2. Divide the following four expressions into those that require substantiation and those that do not, and provide reasons: 'A scent you fall in love with after one use', '500,000 cumulative members', 'The fun of choosing from three flavors', 'Developed by 12 professional chefs'.

1-3. State the three general requirements that the Notification on the Operation of Substantiation demands of substantiation materials, and explain which requirement is most frequently violated when reviewing generative AI drafts and why.

1-4. A health mattress brand intends to use in an advertisement the results of a survey conducted directly by its employees on 150 store visitors. Name two or more items that could be problematic among the criteria for evaluating survey results under the Notification on the Operation of Substantiation.

1-5. Explain the scope of sales records to check when reviewing the 'Regular price' label on a discount banner, using the concept of previous transaction price under the Designated Notification.

1-6. Write in order what the company must do on the day it receives a request to submit substantiation materials from the Fair Trade Commission (FTC), including the relationship between the submission deadline and an order to suspend advertising.

1-7. (Design) Design the columns of a table for reviewing sentence by sentence a draft product detail page written by generative AI. Include at least five columns, and determine what is recorded in each column and who records it.

Hands-on tasks