AI Ads and Fake Testimonials: What Disclosure Cannot Fix

Hero image: AI-generated fictional advertising review scene. The people and product are illustrative.

A fictional accessories brand is preparing a travel-pouch ad for several markets, including Korea. The presenter looks like a customer, holds the product and says: “I used this on my last three trips. Nothing ever leaked.” The team created the presenter with AI. There were no trips, and nobody tested the pouch for leaks.

The editor adds “AI-generated” to the caption. The video still tells a story that never happened.

That is a useful place to begin reviewing an AI ad. Who appears to be speaking? What does the ad say happened? What evidence supports the product claim? Those questions need separate answers before the team approves the finished version.

Korea offers a concrete case for this review. Its revised recommendation-and-endorsement review guideline took effect on June 1, 2026. It addresses realistic AI-generated endorsers and makes clear that disclosing a virtual person does not resolve an invented experience. For brands preparing campaigns now, a separate food-advertising amendment taking effect on November 27 adds another reason to check the creative before scheduling it. Source: KFTC endorsement guideline; KFTC implementation announcement; MFDS implementing-rule proposal.

This guide uses Korea’s rules as a worked regulatory example. Its production checks can help teams review international campaigns, but the legal requirements described here are Korean, not a statement of law in other markets. Check each target market’s requirements before reusing the creative. This is general editorial guidance, not legal advice for a particular advertisement. Sources checked November 12, 2026 (KST); confirm current rules and product-specific requirements before release.

First, identify what the AI person is doing

An AI background, a virtual presenter and a fabricated customer review create different questions. Start with the audience’s likely understanding of the finished ad, rather than the tool used to make it.

The June guideline’s virtual-person provisions concern AI-generated people who are difficult to distinguish from real people and who recommend or endorse products. The broader guideline describes endorsements as content that can be perceived as a third party’s independent opinion, rather than the advertiser’s own statement. It is not a blanket rule saying that every use of AI in advertising falls into this one category. Source: KFTC endorsement guideline.

For the pouch campaign, examine the voiceover, account name, caption and visuals together:

Does the presenter sound like a customer describing a purchase?
Does the setting imply a real trip, test or professional assessment?
Do a name, job title or uniform suggest a real person’s qualifications?
Would a viewer understand which parts are an illustration and which are documented facts?

Calling the character a “brand ambassador” in the production brief tells the viewer nothing. What matters in review is the impression the viewer actually receives.

Put the disclosure where the person appears

Under Korea’s guideline, a qualifying virtual endorser must be identified in a way consumers can clearly and easily recognize. Its examples include a statement in the title or opening of a text post, a label close to the person in a photograph, and a nearby label throughout the person’s appearance in a video. The KFTC’s explanation illustrates wording such as “가상인물 포함” (“contains a virtual person”) and “가상인물” (“virtual person”). Source: KFTC endorsement guideline; KFTC implementation announcement.

For a Korean-language video, use clear Korean wording that tells viewers the person is virtual. A production note such as “made with AI” can leave it unclear whether the tool created the presenter, cleaned up the audio or changed the background. Choose wording that describes the actual use.

Then inspect the exported version on a phone. Can the label be read against the background? Is it still beside the presenter after the vertical crop? Does it disappear during a close-up? Is a platform button covering it? These are practical checks, not a new set of statutory pixel or timing thresholds.

A profile biography or a caption hidden behind “more” is a poor substitute for the guideline’s clear, nearby disclosure examples. Do not assume a platform’s AI label automatically satisfies every requirement for the ad.

Rewrite the experience before polishing the label

The guideline separately says that a virtual endorser’s statement presented as an experience may constitute improper advertising when it does not match facts that actually occurred. The KFTC’s implementation announcement explicitly addresses this even where the person has been labeled virtual. Source: KFTC endorsement guideline; KFTC implementation announcement.

Return to the fictional pouch video. There are three claims inside one sentence:

A customer used this particular pouch.
That customer took it on three trips.
The pouch did not leak during those trips.

A convincing face and voice cannot supply those facts. Adding a label leaves all three unsupported.

The brand can instead build a factual demonstration around the actual product. If the finished pouch genuinely has two compartments and a removable, washable lining, a revised script could say: “Two compartments separate your toiletries. Here is how the lining comes out for cleaning.” Show the real product performing that action, or make any illustrative sequence clear and accurate.

That wording still needs checking. “Washable” must match the care instructions. The compartments must be present in the product sold. An AI-generated demonstration must not invent a removable lining or make the pouch appear larger than it is.

If a real customer has a useful story, preserve the original account, the relevant product and the permission to use it. Do not turn a mixed review into an unqualified endorsement or present several people’s comments as one person’s experience. Using a synthetic presenter to retell a genuine account requires careful attribution and review of the overall impression; it is not an automatic safe route.

Keep evidence for the product claim, too

A genuine testimonial does not prove every broader claim a brand might attach to it. One customer saying a pouch stayed dry is different from advertising it as “leakproof.” A test of one material under specified conditions may not support a promise about every finished pouch in everyday use.

Article 5 of Korea’s Fair Labeling and Advertising Act requires businesses to be able to substantiate factual matters in their advertising. That obligation is separate from identifying a virtual presenter. Source: Fair Labeling and Advertising Act, Article 5.

Before approving a factual claim, keep the source beside the script:

For a product feature, retain the specification and check the actual item.
For a test result, retain the report, the tested product, conditions and limits.
For a customer experience, retain the original account and the approved use.
For an expert statement, verify who made it, their relevant expertise and what they actually concluded.

Treat these as a working review record. They do not guarantee that a claim is lawful or that a particular test is sufficient. Where the evidence supports a narrower statement, narrow the copy. Where no relevant evidence exists, remove the claim rather than asking AI to make it sound more credible.

Check payment disclosure separately

The audience may need to understand both that the person is virtual and that the recommendation has a commercial relationship behind it.

The endorsement guideline separately addresses economic relationships that can affect an endorsement’s content or credibility. Payment, free products and employment can matter. The appropriate disclosure depends on the actual arrangement and medium. Source: KFTC endorsement guideline.

For a paid creator collaboration, write down who is paying, who is posting and what the audience will see. “Virtual person” explains the identity. It does not explain a paid relationship. Equally, “advertisement” does not tell the audience that the apparent customer was generated.

Agree on the disclosures in the brief, then check that they survive translations, cutdowns and paid-ad versions. Keep the factual review attached to those versions too. A shortened edit can create a new problem if it removes the sentence that qualified a claim.

Food brands need a November 27 check

A separate amendment to Korea’s Act on Labeling and Advertising of Foods, promulgated on May 26, takes effect on November 27, 2026. It addresses ads using realistic AI-generated sounds, images or video that may mislead viewers into believing relevant experts, such as doctors, pharmacists or professors, endorse, recommend or use the food product. This is a specific expert-impersonation provision, not a blanket ban on AI-assisted food ads. Do not approve a realistic doctor-style endorsement merely because an “AI” label has been added. Source: MFDS implementing-rule proposal; Food-advertising amendment effective November 27.

Check the rules for the campaign’s actual launch dates. The MFDS’s August 7 implementing-rule notice was a proposal; do not present its proposed sanctions as final requirements. Separately, the KFTC’s October 22 consultation concerned a nonbinding draft explaining existing advertising law for AI-related claims, with comments invited through November 11. It did not create a new law or replace the June endorsement amendment. Source: MFDS implementing-rule proposal; KFTC October consultation announcement.

Use one approval record for the finished ad

A short review record can keep the creative and evidence together:

Final asset: filename, version, platform, language, target markets and intended campaign dates.
Presenter: real or virtual, their apparent role, and the disclosure text and placement.
Claims: each factual statement and any claim conveyed by the visuals.
Evidence: the source supporting each claim, with its limits and date checked.
Commercial relationship: who paid or provided value, and the relevant disclosure.
Market and product-specific review: local requirements checked, applicable sector rules, unresolved questions and the person responsible for clearing them.

Watch the final export, not just the script. A caption can be accurate while a visual invents a result. A presenter can be disclosed while their testimony remains fictional. A real test can be cited while the headline promises more than it showed.

For the pouch brand, the stronger ad is the one the team can explain plainly: the presenter is virtual, the demonstrated features are real and the claims match the evidence. Make those decisions before spending on distribution.

Sources

PUT IDEAS INTO PRACTICE

Turn insight into opportunity.

Talk with PIN-AGENCY KOREA about creator partnerships, content production and reaching international audiences in Korea.

Work with PIN-AGENCY KOREA