Labeling AI Ads: Who Checks What Your Marketing Agent Puts Into the Image?

Picture an ordinary Tuesday morning in the marketing department of a mid-sized company. The ad account is running, the autumn campaign is live, and somewhere in the settings, automatic image and video features are switched on because someone recommended them weeks ago as a way to "save time". Overnight, variants have appeared: a background was swapped, a product photo was relit, a short video was assembled from three still images. The ads are running. And nobody has asked whether any of these images should be labeled as AI-generated or AI-altered.

That question is uncomfortable because it touches two things at once: the brand and the law. The brand, because customers increasingly notice when something looks "not real", and a badly placed AI image can quickly cost trust. The law, because there are rules on disclosing artificially generated content that you should read before a letter lands in your inbox, not after. This article is explicitly not a news piece. It deals with a lasting topic that does not hinge on a deadline: who checks what an agent puts into an image before it is published?

The occasion for raising it today is a figure from a study summarized on October 22, 2026 by the trade outlet NetInfluencer (author: Dragomir Stojkov). We put it in context, say openly what it proves and what it does not, and show how a marketing team with clear limits and an approval step can handle it. All details about the study come from that coverage; we have not checked the vendor's original study ourselves.

What the study measured, and what the source says

The study comes from Resemble AI, a company that sells software for detecting synthetic media. According to NetInfluencer's summary, 37,768 image and video files from 15,433 ads were analyzed. The files come from Meta's public ad library, pulled in September 2026, and cover 509 companies across 26 industries. The vendor's own detection tool was applied to those files.

The result in the source's short form:

• 3,379 files were flagged as AI-generated or AI-altered. That is 8.9 percent of all files analyzed, which the report phrases as "roughly 1 in 11".
• Looking only at files that brands ran through their own pages, the rate was 6.9 percent (1,971 of 28,463 files).
• 222 of 389 brands examined (57 percent) ran at least one ad with content flagged as AI.
• Only 2 of 3,379 flagged files (0.06 percent) disclosed AI use in the ad copy. The only brand the report names is Turkish Airlines, in the caption of a video for International Cat Day.
• For 971 files, the study also checked whether they carry so-called C2PA provenance credentials. Not a single one did, including the confirmed fakes.

Two remarks belong here right away. First: Resemble AI sells detection software. A vendor selling such software has an obvious interest in a message like "there is a lot of AI content, and it is barely labeled". That does not make the numbers wrong, but it is a reason to read them as an indication, not as a measurement from a neutral lab. Second: we could not find the original study directly. We rely on a trade outlet's summary with author and date. If you want to use the numbers in a presentation, check them against the original source first.

What the numbers mean, and what they do not

A rate like "8.9 percent" invites the conclusion that one in eleven ads is "cheating". The source does not say that. It helps to break the statement into four parts.

First: detection is not intent. A detection tool classifies a file as probably AI-generated or AI-altered. Whether a team deliberately used a generated image or a platform feature quietly produced a variant in the background cannot be read from that. As far as we can tell, the report does not distinguish between the two. Nor does a detection rate say how reliable the tool is. That, too, is a reason to stay cautious.

Second: the rate varies by industry. According to the report, it is highest for consumer packaged goods at 13.4 percent, 3.8 percent for banking and financial services, and 2.8 percent for pharmaceuticals. The study also finds that, regardless of company size, the rate sits at roughly six to nine percent. For mid-sized companies that means there is no size discount. Whether a business has 20 or 2,000 employees does not change the basic question.

Third: not every flagged file comes from the brand. The report also describes pages that use a brand name without being run by the brand, where the share of flagged AI files is far higher. That is a different problem, namely brand misuse, and it should not be mixed up with the daily reality of a legitimate ad account. That is why this article names no individual brands except the one that, according to the source, did disclose.

Fourth: the core message remains uncomfortable. Even if you relativize the numbers generously, a ratio remains that stands out: of several thousand files flagged as AI, almost none carried a notice in the ad copy. And even the technical provenance marking was practically absent from the sample.

"Only 2 of 3,379 AI-flagged files (0.06 percent) disclosed AI use in the ad copy." The figure comes from NetInfluencer's October 22, 2026 summary of the Resemble AI study. It does not say rules were broken. It shows that labeling hardly happens in practice, and that there is usually no responsible process for it.

That last half-sentence is our assessment, not part of the source. We still think it is the real point: where nothing is labeled, it is rarely because of ill will. It is usually because the question appears in no working step.

Three levels of labeling: text, file, law

When people talk about "labeling", three different things are often mixed together. It helps to keep them apart, because a different party is responsible for each level.

Level 1: the notice for humans. This is the visible notice in the ad or the ad copy, such as a sentence or a label saying that content was created or altered with AI. This is exactly the level the study counted when it looked for notices in the ad copy. Whoever is responsible for the text and the image is responsible here, so the marketing team.

Level 2: the notice for machines. This is about metadata in the file itself. The C2PA initiative (Coalition for Content Provenance and Authenticity) describes on its website an open technical standard meant to help track "the origin and edits of digital content". Content Credentials are compared there to a "nutrition label for digital content"; anyone can view a file's history. Those are the wordings of the C2PA site, and we claim nothing more at this point. In practice, one simple observation follows: such a record has to be written when a file is created or exported. Whether a tool does that, whether it is switched on, and whether a platform keeps the record later depends on the tools in use. In the study's sample, it was present in none of the 971 files checked.

Level 3: the legal duty. This is where it gets delicate, and we explicitly do not want to give legal advice. We can only report what the regulation says according to the text on artificialintelligenceact.eu. Article 50 of the EU AI Act contains two rules that can be relevant for marketing. Providers of AI systems that generate synthetic audio, image, video or text content must ensure that outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated". And deployers who generate or manipulate image, audio or video content that counts as a deepfake must disclose that the content has been artificially generated or manipulated. The page gives the application date as August 2, 2026. How these rules affect a specific ad, from when they apply to your company, and whether your image counts as a deepfake at all are legal questions. Have them answered by a professional who knows your case. We have already published our own article that places Art. 50 alongside the ad labeling requirement under German unfair competition law; you will find the link under "More articles" below.

A practical consequence follows from the three levels: a platform can check Level 2, an ad account can support Level 1, but responsibility for Level 3 never leaves your company. Whoever publishes content is responsible for it, whether a person, a tool or an agent produced it.

Availability: a lasting question, not a product feature

At this point, a news article would talk about "generally available" or "beta". That does not fit here, and we want to say so honestly. This is not about a single new product feature. It is about a question that arises wherever a tool or an agent creates or alters images and videos before they are published.

That advertising platforms offer automatic creative features is well known. Which ones exactly are active in your ad account, which are switched on by default and which setting produces which variant, we cannot tell from the outside. That is in your account and in the respective platform's help pages. We therefore claim nothing about specific defaults or duties of individual platforms. For the DACH region, our honest answer is: the topic affects you to the extent that you use such features or use a service provider for them. The summary of the study that we read names no breakdown by country.

The report also points out that some countries and regions (South Korea, the US state of New York, China and India) have rules on labeling AI-generated content in advertising, while the US relies mainly on the FTC's general deception rules. That is NetInfluencer's statement; we did not verify it individually and cite it only as an indication that the topic is regulated differently around the world.

Who in the marketing team checks what the agent puts into the image

On amazing-agents.com we describe a social media team of agents that handle tasks separately. Three of them matter for this article. We stick to what is written on their respective pages and add nothing that is not there.

Sophia is the image designer in the social media team. She creates visual assets such as carousels, quote cards and reel covers following the brand board. Her page states explicitly that she does not deliver files ready for direct upload: every asset goes through the QA check for brand fit before it enters the editorial calendar.

Magdalena is the QA specialist in the Tech & Ops team. She checks content drafts against a fixed rule catalog: brand fit, tone, buzzword blacklist, platform rules and formal correctness. Her page says: "No visual goes live without a brand check." Her process is binary, passed or not passed, and a failed draft goes back with concrete feedback.

Natascha is the content agent for social media. Her page says that nothing goes live without your approval and that a brand check runs before approval.

What follows for our topic is a recommendation, not a promise about existing processes: the question "Does this image have to be labeled as AI content?" is a natural addition to the fixed rule catalog that a brand check works through anyway. It belongs at the same point as the question about buzzwords or character limits: before approval, not after. Whether an image has to be labeled is ultimately decided by a person in your company, with legal advice where needed. The agent can ask the question; you are responsible for the answer.

We deliberately refrain from examples "from our practice" or success figures. Instead of case stories, we would rather describe how we proceed in an audit.

How we proceed in an audit

An audit of automatically generated ad content consists of five steps, all of which are also possible in your own company without us.

1. Inventory: Which tools and platform features create or alter images, videos or texts in your company? This includes features in ad accounts, your agency's image tools, stock providers with an AI share, and your own agents. The result is a list, not an evaluation.
2. The file's path: How does an image get from draft to ad? Who sees it along the way? At which points can it change without a person seeing it?
3. The labeling question per content type: For each type (product photo with a new background, fully generated image, voice, video made from stills), you record who decides whether and how it is labeled. You obtain the legal assessment from a suitably qualified person.
4. Technical provenance records: Does the tool in use support Content Credentials under C2PA, and can the feature be switched on? And: do the details survive export? You best check that with a test file.
5. Fix the approval step: Who approves, using which checklist, and where is the decision documented? Without a documented decision, you cannot later show that a check took place.

At the end there is no certificate, but a process that your team understands and can follow.

What you can do this week

You need neither a project nor a budget for this. Three steps are enough to start.

1. Look into your ad account. Open the settings for automatic image and video features and note which are active. Also ask your agency or your employees which tools they use.
2. Check ten ads. Pick ten running ads and ask for each: Was the image or video created or altered? By whom? Who saw it before publication? If nobody can answer the question, you already have your most important finding.
3. Name one person. Decide who will ask the labeling question before publication in future and who will answer it. It can be the same person who already approves ads today. Record the decision in a simple document.

The difference: an agent with limits and an approval step

Let us sum up without making the situation more dramatic than it is. The study we discussed here comes from a vendor with a vested interest, we know it only from a summary, and the numbers show a detection rate, not intent and not a legal violation. At the same time, it shows something every marketing lead knows: when a question appears in no working step, it does not get asked.

The difference between two ways of working with agents can be described in a few sentences.

An agent without limits and without an approval step generates variants, publishes them, and the labeling question is never asked. If someone asks later, there is neither a decision nor a person who made it. The risk lies with your brand.

An agent with limits and an approval step works within a fixed frame: it may draft, but not publish. Sophia designs following the brand board, Magdalena checks against the rule catalog (and that catalog should contain the labeling question), Natascha submits for approval only after the check, and a person in your company makes the decision. The agent's work stays fast; the responsibility stays with you. That is exactly the difference that matters: not agent or human, but an agent that knows where its limit runs and when it waits for your approval.