Meta 工程师撰文指 AI 标签无法说明广告声称真假，建议按风险分级披露

Monday, November 2nd, 2026 – 12:25 am

SHARE:

Sourin Chakrabarti Software Engineer

Disclosure: Sourin Chakrabarti works on AI labeling for advertising at Meta. The views expressed are his own and do not represent Meta.

In October 2025, three clips appearing to show police detaining and pepper-spraying protesters racked up millions of views across X, Instagram and TikTok. Created with OpenAI’s Sora, the videos carried visible watermarks, yet NewsGuard found that users still cited them as evidence that the events occurred. The footage crossed an unsettling threshold: It didn’t just look realistic; it functioned as visual proof. 

Most concern about deepfakes has focused on politics. But the same tactic is now appearing in advertising, where synthetic content can make a product claim look proven. 

In a recent campaign investigated by Full Fact, a supposed doctor claiming to be affiliated with Guy’s and St Thomas’ NHS Foundation Trust tested a weight-loss patch and appeared to grow thinner on camera. The trust called the posts “completely fraudulent,” while Google’s Gemini detected SynthID watermarks confirming AI involvement. Yet that watermark only revealed that AI was involved, not where the fiction ended. Was only the doctor’s face synthetic? Or were the medical authority, hospital connection, personal experience and physical transformation invented, too?

This exposes the fundamental blind spot of AI disclosures: they flag the production tool, not the illusion. A generic label tells consumers that AI was used to build an ad, but leaves them to guess where reality ends and fiction begins.

When a simulation becomes evidence

The fraudulent NHS weight-loss ad was not an isolated incident. In 2025, the Australian Medical Association warned that deepfakes impersonating prominent doctors were promoting unproven products, with clinicians discovering the fraud only when patients asked whether to swap prescribed treatments for advertised miracle cures. In 2026, Chinese regulators launched an enforcement campaign targeting businesses that mass-produced fabricated consumer-experience posts.

The legal principle predates generative AI. In FTC v. Colgate-Palmolive Co., a shaving cream commercial used a plexiglass prop in place of real sandpaper. Viewers were misled because the simulation was presented as a genuine product test. The case offers a useful test for AI advertising today: Does an ad merely illustrate a claim or does it ask viewers to treat a synthetic scene as real-world proof?

From AI labels to claim disclosures

Platforms and regulators are building substantial disclosure systems to track synthetic content. Google displays “How this ad was made” panels, Meta attaches “AI info” labels and TikTok offers “AIGC” self-disclosure controls. Governments are pushing in the same direction: The European Union’s Article 50 rules require transparency for synthetic media, China mandates explicit AI markers and New York requires advertisements to identify synthetic performers.

Yet current AI-labeling systems mostly tell consumers that AI was used, not what it changed or why it matters. The same label can appear on an ad with a decorative AI background and one with an entirely fabricated testimonial. Those uses are not equally misleading, but a generic label may not distinguish between them.

Some policies address the clearest cases. The Advertising Standards Council of India’s 2026 draft guidance identifies fabricated experts, endorsements and exaggerated results as high-risk uses that labeling alone cannot fix. TikTok Shop prohibits AI-generated professional identities used to sell health products and alterations that misrepresent product performance. But a harder question remains: What about AI simulations that are lawful but could still be mistaken for real-world evidence?

The disclosure gap is becoming urgent as AI ad generation reaches industrial scale. In 2025, advertisers created three times as many Gemini-generated assets on Google as in 2024, including nearly 70 million in the final quarter alone. Meta has reportedly aimed to automate ad creation and targeting with AI by the end of 2026. Because these tools are embedded directly into ad-creation systems, safeguards cannot rely on post hoc consumer complaints. They must operate systematically across millions of ads.

The answer is not to label every synthetic pixel. Building on the IAB’s January 2026 disclosure framework, regulators should adopt a risk-based approach requiring specific disclosures whenever AI material appears to provide evidence for a product claim. This approach would produce three outcomes:

• Routine editing: Basic edits and decorative backgrounds receive no additional warning.

• Lawful simulations: Synthetic media that could be mistaken for real evidence must state what it represents, such as an “AI reenactment of a verified customer testimonial” or “simulated demonstration; performance claim based on laboratory testing.”

• Outright rejection: Fabricated endorsements, invented expert opinions and unsupported results are blocked entirely. This is consistent with a recent legal analysis and mirrors the FTC guidance, which permits AI avatars but prohibits fake testimonials attributed to them.

Platforms could implement this approach by requiring advertisers to disclose three things: the role played by any AI-created person or scene, whether the generated material depicts product performance or results and what real-world evidence supports the claim. Meta, Google and TikTok already accept advertiser declarations, so these questions could fit into existing workflows. Platforms should also retain a record of the advertiser’s answers and the notice shown to consumers, creating an auditable trail of how each ad was reviewed and presented. An emerging industry proposal calls this a served disclosure receipt.

Consider the weight-loss patch video. Under this system, an “AI was used” label would be only the starting point. The advertiser would have to identify the doctor as a synthetic expert, classify the apparent transformation as a product-performance claim and clarify whether the testimonial represented a real customer. It would also have to provide evidence supporting the weight-loss claim. If the doctor was falsely presented as real, the testimonial was fabricated or the result lacked substantiation, the ad would not be approved.

The goal is labels that answer the question consumers actually face: not just whether AI made the ad but whether AI made up the reason to buy.

“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.

Follow Sourin Chakrabarti and AdExchanger on LinkedIn.