Appendix B. Self-check answers and explanations

Chapter 1

1-1. They are falsehood and Exaggeration (nonexistent rankings or inflated sales volume), deception (omitting conditions such as penalty fees and additional costs while emphasizing only benefits), unfair comparison ('2x compared to competitors' without a comparison target or criteria), and slander (expressions disparaging competing products without grounds). Because generative AI tends to fill in plausible numbers and omit conditions, errors of the falsehood-and-Exaggeration type and the deceptive type are especially frequent.

1-2. What requires substantiation is '500,000 cumulative members' (a figure that can be verified through membership records) and 'Developed by 12 professional chefs' (a fact that can be verified through development personnel and career experience). 'A scent you fall in love with after one use' is a subjective judgment about scent, and 'The fun of choosing from three flavors' is an emotional expression, so 'Fun' itself is not subject to substantiation. However, because the product configuration of having three flavors is a fact, whether it matches the product information must be checked.

1-3. They are the method (tests, surveys, expert opinions, academic literature, etc.), objective and scientific procedures, and a direct relationship with the advertising claim. What breaks down most frequently in AI drafts is the direct relationship. This is because AI broadens the general effects of ingredients, material tests, and results under laboratory conditions into the actual-use effects of the finished product.

1-4. The independence of the survey institution (company employees directly conducted the survey), the representativeness of the sample (store visitors may not represent all buyers), the objectivity of the survey (fairly conducted by a third party), and non-awareness of the survey purpose (respondents answered while knowing it was a company survey) can be problematic. Whether the questions have a direct relationship with the advertising claim must also be verified.

1-5. The Designated Public Notice regards the previous transaction price as the price at which the same product was actually sold for a substantial period recently (around the past 20 days), and if the price fluctuated during that period, it regards it as the lowest price among them. Therefore, the actual sales records for approximately 20 days immediately before the discount and the lowest price during that period must be checked, and a price at which the product was never actually sold must not be used as the 'Regular price' as if it were the company's own previous transaction price (the criteria for manufacturer's suggested retail price and market price follow the separate requirements and explicit statement of standards in the public notice).

1-6. ① Check the labeling and advertising designated in the request letter and the contents subject to substantiation. ② Gather the substantiation data in possession and prepare to submit it within 15 days from the date of receiving the request (Article 5, Paragraph 3). ③ Take down or modify expressions from the advertisement for which data is insufficient. If advertising continues without submitting data within the period, an order to suspend the advertisement may be issued until data is submitted (Article 5, Paragraph 5), and exceeding the period subjects one to an administrative fine. ④ Remember that an extension may only be requested when there are grounds specified by the Public Notice on the Operation of Substantiation (natural disasters, mergers and acquisitions, seizures, etc.).

1-7. Example columns are sentence number, original text, whether it is a factual claim and its type, type of required evidence, evidence materials (name, subject, conditions, period), decision (maintain, modify, delete), revised phrasing, reason for decision, and reviewer and date. The division of work is such that the author enters whether it is a factual claim and its type in the primary round, the person in charge of evidence materials enters the evidence materials and decision, and the final approver signs the confirmation column.

Chapter 2

2-1. (B) and (C) apply. (B) is a consumer evaluating a product based on their own usage experience, and (C) is an expert recommending a product, which can be perceived as an Independent opinion of a third party. (A) is phrasing written by the advertiser under the company's name, so it is the advertiser's own opinion.

2-2. That the compensation is not conditioned on writing a review, that the compensation is provided equally to all applicants or participants, and that the review is posted voluntarily. In this case, it can be viewed as an exception where economic interests do not affect the credibility of the Recommendation.

2-3. For blog posts, display it in the title or at the beginning so that it is distinguished from the main text. For photo-centric posts, in principle place it inside the photo; if the photo and the main text are connected and easily recognized, it may be placed at the beginning of the text, and if it is a hashtag, in principle place it as the first hashtag. For videos where only a portion of the video constitutes a Recommendation, insert the disclosure at the beginning and end of that section and display it repeatedly during the Recommendation (displaying it in the title is also an option).

2-4. 'Tester group' is an expression where the details of the compensation are not clear (content principle), '#BrandName' is an expression that merely mentions the brand and does not disclose the compensation relationship (content principle), and 'Thanks to ○○' is a foreign language expression not regarded as the same language as the Recommendation targeted at domestic consumers (language principle). The Review Guidelines cite all three expressions as inappropriate examples.

2-5. A Virtual persona has no experience of actually using the product. Therefore, experiential expressions such as 'It was great using it', no matter how well disclosed as being a Virtual persona, cannot conform to actual experience and can become unfair advertising. Separate from the disclosure, experiential lines must be changed to information-delivery lines.

2-6. ① Show the proportions of positive, negative, and neutral according to the actual distribution; ② do not omit recurring major complaints; ③ enable direct navigation from the summary to the original review. In addition, it is recommended to add a procedure to record the summarization criteria and regularly cross-check them against a sample of the original texts.

2-7. Example: ① The details of compensation and the obligation to disclose in every post (principle of disclosing economic interests), ② disclosure wording, placement, and form by media (principles of placement, form, content, and language of disclosure), ③ writing based on the actual period of use and actual experience (actual experience principle), ④ not distorting the original content of the Recommendation by advertiser processing or restructuring (prohibition of distortion), ⑤ efficacy and figures only within the advertiser's scope of substantiation (Burden of proof), ⑥ prohibition of deleting or moving disclosures after posting (disclosure for each item), ⑦ contact method when the Recommendation is changed or Withdrawn (principle of change and Withdrawn).

Chapter 3

3-1. The contact information must have been received directly from the recipient through a transaction relationship, it must be within 6 months from the date the transaction ended, and the advertisement must be for goods of the same kind as those transacted. Even if an exception applies, the prohibition on transmission after refusal to receive or Withdrawn consent (Paragraph 2), separate Prior consent for night-time transmission (Paragraph 3), mandatory disclosure items (Paragraph 4), and prohibited acts (Paragraph 5) must still be complied with.

3-2. Article 50(3) requires obtaining separate Prior consent to send advertising information from 9:00 PM to 8:00 AM the following day, and Article 61(2) of the Enforcement Decree specifies e-mail as the excepted medium. Therefore, an e-mail at 10:00 PM can be sent to a person who has consented to receive advertising, but app push notifications can be sent only to those who have separately consented to night-time reception.

3-3. ① Immediately remove them from the transmission list upon receiving the intention, and exclude them from scheduled transmissions as well (Paragraph 2). ② Within 14 days from the date the intention was expressed, notify the sender's name, the fact and date of the refusal to receive, and the processing result (Paragraph 7, Article 62-2 of the Enforcement Decree). ③ Verify that the recipient does not incur any cost for processing the refusal (Paragraph 6). ④ Check whether it has been reflected in the transmission lists of other channels as well.

3-4. (B) and (C). (B) is the act of automatically generating contact information (Paragraph 5, Subparagraph 2), and (C) is the act of concealing the sender's identity or origin (Paragraph 5, Subparagraph 4). (A) is a desirable design that makes refusal easy, and (D) is the periodic verification required by Article 62-3 of the Enforcement Decree.

3-5. Consent to processing for promotional purposes (Article 22(1)7 of the Personal Information Protection Act (PIPA)) is consent to use personal information for promotion or sales solicitation, and Consent to receive advertising information (Article 50 of the Network Act) is consent to receive advertising messages. If there is only consent to receive, the basis for which data may be used for personalization may be insufficient; if there is only Consent to processing for promotional purposes, Consent to receive advertising information (or a transaction relationship exception) to send advertising messages must be separately verified.

3-6. Example of what can be done: Creating purchase trend statistics by age group and region using pseudonymized purchase data (Article 28-2, compilation of statistics). Example of what must not be done: Matching pseudonymized information back with original member information to send coupons to a specific member (Article 28-5(1), prohibition of processing to identify a Specific individual).

3-7. Example: ① Purpose of processing and the minimum items to be transferred, ② Prohibition of processing beyond the purpose of the outsourced work, ③ Technical and managerial protective measures and access restrictions, ④ Consent of the entruster in the event of re-outsourcing, ⑤ Condition that input data shall not be used for other purposes, such as the vendor's model training, ⑥ Methods for inspecting and supervising processing status, ⑦ Liability for damages, ⑧ Methods for return, destruction, and verification after the termination of outsourcing, ⑨ Website disclosure of the trustee and the outsourced work, and notification to data subjects regarding the outsourcing of promotional work.

Chapter 4

4-1. Example: "In an app push sent during the same week by randomly dividing active members from the last 90 days, does the copy written by AI increase the purchase conversion rate within 7 days after transmission by 0.3%p or more compared to existing copy?" It includes the comparison targets (AI copy vs. existing copy), the target population (active members from the last 90 days), the primary metric (purchase conversion rate within 7 days), and the decision criterion (0.3%p or more).

4-2. The two metrics have different denominators. Because the open-basis metric includes only people who have already opened the message in the denominator, it yields a higher result than the send-basis metric; mixing and comparing them means most of the difference stems from the difference in definitions. They can only be compared after recalculating them with the same numerator, denominator, and unit.

4-3. Significance level (α), statistical power (1−β), and the minimum detectable effect (δ). Because the sample size is roughly inversely proportional to the square of δ, halving the difference to be detected increases the required sample size by nearly four times.

4-4. Permissible actions: Data quality checks, such as verifying whether the allocation ratio matches the settings and whether impression records are accumulating normally. Impermissible actions: Declaring a winner based on interim results of a fixed-sample test and stopping the experiment. This is because repeatedly peeking and making decisions causes the false positive rate to exceed the predefined significance level.

4-5. The opt-out rate is a representative guardrail metric. Under Article 50(2) of the Network Act, advertising information cannot be sent again to a person who has refused to receive it; therefore, an increase in the opt-out rate represents a cost that reduces advertising reach until the recipient gives Explicit consent again. Even if the click-through rate increases, if the guardrail metric worsens, the rollout should be put on hold and the causes (sensational expressions, sending frequency, etc.) must be investigated.

4-6. ① Write brand tone as a detailed, evaluable rubric. ② Receive the output in a concrete format, such as 'Pass/Fail' or a 1–5 score, and have it reason before scoring. ③ First verify the agreement of the LLM scoring against human-graded samples before expanding to the entire set. In addition, include the principle of scoring with a model different from the one that generated the copy.

4-7. Example: Question (Does AI-personalized SMS increase the 7-day purchase conversion rate more than generic SMS?), target (members with consent to receive advertising and Consent to processing for promotional purposes), allocation (50:50 randomized at the member level), metric (send-basis 7-day purchase conversion rate), guardrail metrics (opt-out rate, customer inquiries), sample (calculated from baseline and minimum meaningful difference), duration (until the calculated sample size is reached, checking weekly trends), decision criteria (greater than or equal to minimum difference, no deterioration in guardrail metrics), legal verification (substantiation of factual claims in copy, night-time rules, mandatory disclosure items, basis for personalization data, and outsourcing requirements).

Chapter 5

5-1. Any four or more of the following: clear and specific instructions (formats and constraints), reason and context for the instructions, 3–5 relevant and diverse examples, a structure separating instructions, context, and examples using tags, and telling what to do rather than what not to do.

5-2. Example: "Because our product is not a pharmaceutical, promising therapeutic effects may mislead consumers. Instead of 'Complete cure' or 'Cures', describe it using sensory application experience (absorbs lightly) and verified ingredient information." Providing both the reason for the prohibition and alternative expressions to use allows the model to generalize and avoid other similar expressions.

5-3. Example: High (claims of efficacy, Safe qualities, or AI technology; rankings, superlatives, comparisons; pricing and discounts; faces or voices resembling real individuals; Virtual persona; targeting children) requires verification by the person in charge of substantiating evidence or rights/disclosure, followed by approval from the final approver. Medium (sponsored/review content, advertising messages) requires verification of the disclosure and transmission checklist by the person in charge. Low (emotional copy without factual claims, general images) requires approval by the brand manager. Periodically review samples of the Low tier to check whether the classification is correct.

5-4. Because the Copyright Act defines a work as a creative production expressing human thoughts or emotions, the portion to which a human creatively contributed becomes the core of protection. Documenting AI drafts, records of human redrawing or revisions, composition and selection decisions, workers, and dates serves as the basis for determining and explaining the scope for which protection can be claimed. Other systems such as trademarks should also be reviewed together to protect brand identification.

5-5. Article 2, Subparagraph 1, Item (l) of the Unfair Competition Prevention and Trade Secret Protection Act (the act of using another person's identifying mark, such as name, portrait, or voice that is widely recognized in Korea, for business without authorization) may become an issue. The company may only proceed by not using that voice (re-creating it with a different voice) or by obtaining permission to use it from the singer. An 'AI-generated' label is not a solution.

5-6. ① Allow generation only within a pre-approved scope (response boundaries, templates, real-time connected factual data); ② Have a manager monitor in real time and maintain an emergency stop mechanism to halt immediately if abnormalities are observed (override/suspension of operation); ③ Record and evaluate generation outputs and user feedback to use them for incident response, recovery, and change management.

5-7. ① Halt: Take it down from all channels and cancel scheduled postings and transmissions (halt time, channel list). ② Assess scope: Exposure period, channels, impressions, and affected customers (scope log). ③ Rectify: Correction notice, customer response (notice, response history). ④ Identify cause: At which stage it failed to be filtered (root-cause analysis report). ⑤ Remediate: Revise checklists, instructions, approval workflows, and repositories (change history and approval). ⑥ Identify similar types: Inspect other assets created with the same cause (inspection checklist and results).