# Labeling AI-generated ads: what the law requires in 2027 and how to mark your creatives without hurting conversion

> Practical guide: Article 50 of the AI Regulation, official EU icons, Pennsylvania and California, rules from Meta, TikTok, and YouTube, and a 6-step checklist.

- **Source (source):** https://aistohire.com/ads-image-studio-create-ad-images-with-ai-166/blog/labeling-ads-ai-law-2027-mark-creatives
- **Platform (platform):** AIstoHire
- **Language (language):** en
- **Updated (updated):** 2026-11-20

**In summary:**

- Since **August 2, 2026**, Article 50 of the European AI Regulation requires two distinct things for an AI-generated advertisement image: a **machine-readable label** (provided by the model supplier) and, when the content is a deepfake, a **visible label** for the user (you provide this, the advertiser posting).
- The Commission published the **Code of Good Practices for Transparency on June 10, 2026** and, with it, **three official downloadable icons** for labeling AI content. Using them is optional; labeling is not. In Spain, the supervisor for Article 50 is the **AESIA**.
- In the U.S., there are two deadlines approaching: **Pennsylvania** (law signed on October 22, 2026, effective **December 21, 2026**, up to $3,000 per violation) and **California**, whose obligations for large platforms begin on **January 1, 2027**. New York already requires notification of *synthetic performers* since **June 9, 2026**.
- **Meta, TikTok, and YouTube** do not ask for the same: Meta labels only when it detects AI, TikTok requires disclosure in advertisements starting July 21, 2026, and YouTube only when the content is realistic. And Meta has **paused without a new date** the terms that were set to take effect on November 20, 2026.
- The fear of losing conversion is measured: a study by MediaScience with the Ehrenberg-Bass Institute (June 12, 2026, 900 people) found that the AI label **did not worsen brand recall or attitude towards the ad**; the format of the label did matter.

![Square advertising creativity with a transparency label in the lower right corner](https://cdn.aistohire.com/i/6414-wubf-xxmt-jyze-plik.webp)

If you generate the images for your ads with AI —and according to the IAB, **83% of advertising executives** already do this, compared to 60% in 2024, in their report [The AI Ad Gap Widens](https://www.iab.com/insights/the-ai-gap-widens/) from January 15, 2026—, labeling has ceased to be an ethical debate and has become an operational requirement with deadlines and fines. This guide outlines what each jurisdiction requires, what each platform asks for, and how to set up the process so that labeling a creative does not take you an afternoon. Tools like **Ads Image Studio**, the study of [static creatives for AI ads](/ads-image-studio-create-ad-images-with-ai-166) from **AIsToHire**, are already designed to keep track of the information these regulations will require: provider, model, prompt, and what was edited in each version.

---

## What counts as "synthetic content" in advertising: not just deepfakes

It includes any image, video, or audio that AI has **generated or significantly modified**, not just fake faces. The Pennsylvania law defines *synthetic content* as "any image, video, audio, or text that AI has significantly modified or generated," and the European Regulation speaks of content "generated or manipulated" with AI. An invented background for your product, a model that does not exist, or a beach scene where there was actually a set fits squarely; a crop, a reframing, or a color correction does not.

The practical line used by both regulators and platforms is **appearance of reality**. If a consumer can believe that what they see is a photo of a real place, person, or moment and it is not, they must be notified. If the image is obviously fantastic, illustrated, or impossible, they do not.

What falls inside and what falls outside, in practice:

| Typical case in a creative | Is it synthetic content for labeling purposes? |
|---|---|
| Real product photographed, photorealistic background generated with AI | Yes |
| Photorealistic invented person holding your product | Yes, and in New York, it is a *synthetic performer* |
| Complete scene generated from a prompt with a photo-like appearance | Yes |
| Clearly drawn illustration or obviously unrealistic 3D style | Usually no (evidently fictional content) |
| Crop, resizing, color or light correction | No |
| Ad copy written with an LLM | No in Pennsylvania (excludes plain text); in the EU only if it is text on public interest issues |

> Article 50 of the AI Regulation exempts visible labeling, among other cases, for manifestly artistic, creative, or satirical content and for content that has undergone human review with editorial responsibility. The exemption is narrower than it seems: it does not cover an ordinary commercial advertisement.

---

## European Union: Article 50 has been in effect since August 2, 2026

The transparency obligations of Article 50 of Regulation (EU) 2024/1689 **are enforceable from August 2, 2026**, according to the timeline outlined in the analysis of the [transparency rules of Article 50](https://artificialintelligenceact.eu/transparency-rules-article-50/). For generative AI systems that were already on the European market before that date, there was an extension until **December 2, 2026** to comply with the machine-readable marking of paragraph 2.

These are two separate obligations and it is important not to confuse them, as they fall on different actors:

| Obligation | Who fulfills it | How it materializes in an announcement image |
|---|---|---|
| Machine-readable marking (art. 50.2) | The **provider** of the generative AI system | Metadata of origin or embedded watermark in the file generated by the model |
| Visible labeling of deepfakes (art. 50.4) | The **deploying party**: you, who publishes the announcement | A noticeable notice or icon about the creativity or alongside it |

The Commission adopted its guidelines on Article 50 on **July 20, 2026** (document C(2026) 5054 final), which are not binding but, as noted in the [ICAM analysis on transparency obligations](https://web.icam.es/la-comision-aclara-las-nuevas-obligaciones-de-transparencia-del-articulo-50-del-ai-act/), will be the reference criterion for national supervisory authorities.

### The three official icons of the Code of Good Practices

On **June 10, 2026**, the Commission published the [Code of Good Practices on Transparency of AI-Generated Content](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content), which is voluntarily adhered to, with a section for providers (marking and detection) and another for deploying parties (labeling). By the end of July 2026, around 190 companies and organizations had signed it.

Along with the Code, the Commission released a set of [European icons for labeling AI-generated content](https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content). There are three, and they do not mean the same thing:

1. **Basic icon** — when AI has intervened in the creation of the content, or when you also provide your own text label or a second interactive layer.
2. **Icon for content fully generated by AI** — without elements created by humans or human editorial control beyond the prompt.
3. **Icon for content partially modified with AI** — pre-existing content made by humans that AI has altered.

They can be downloaded in SVG and PNG, in four variants (black, white, and both at 50% transparency), and can be used without attributing anything to the Commission. Their use is **optional**; the obligation to label is not. The guide does set conditions for format: the icon must be perceptible and distinguishable **at the latest upon the first exposure to the content**, clearly visible in size, and must remain visible if the content is downloaded or reshared.

### Who supervises in Spain: AESIA

In Spain, the competent authority for systems subject to the transparency obligations of Article 50 is the **Spanish Agency for the Supervision of Artificial Intelligence (AESIA)**, based in A Coruña and established by Royal Decree 729/2023, unless another authority is competent by sectoral distribution. That is to say: if you publish in the Spanish market, the regulator who can question you about your synthetic creations is AESIA, not a distant European authority.

---

## United States: two deadlines approaching and a state-by-state map

There is no federal law on AI advertising labeling; there are state laws with different triggers. The two that are tightening right now are **Pennsylvania, effective December 21, 2026**, and **California, whose platform obligations start on January 1, 2027**. New York has been requiring notifications for months when an advertisement uses a digital person.

| State | Regulation | Effective | What it requires | Penalty |
|---|---|---|---|---|
| Pennsylvania | Artificial Intelligence Deceptive Advertising Act (SB 806), signed on 10/22/2026 | 12/21/2026 | Clear and visible disclosure if synthetic content creates a false perception about what is being sold | Up to $3,000 per violation, plus costs and fees |
| California | California AI Transparency Act (SB 942 + AB 853) | Effective since 8/2/2026; large platforms and generative AI hosting platforms since 1/1/2027; capture devices since 1/1/2028 | Manifest and latent disclosures in generated content; platforms must detect and display provenance data and not remove it | $5,000 per violation per day |
| New York | Synthetic Performers Act (S.8420-A / A.8887-B) | 6/9/2026 | Visible notice when the advertisement includes a digitally created human | $1,000 for the first violation, $5,000 for subsequent ones |

The details from Pennsylvania are the most stringent in form and deserve careful reading. According to the [Reed Smith analysis on the effective date of December 21](https://www.reedsmith.com/our-insights/blogs/viewpoints/102o9i7/fake-content-false-impression-pennsylvanias-ai-ad-disclosure-law-takes-effect/), the notice must:

- State that the advertisement contains modified or AI-generated content.
- Appear **in the first instance in which the synthetic content is presented**.
- Be **in the same medium** as the synthetic content: if the piece is an image, the notice goes in the image.
- Be permanent or extraordinarily difficult to remove, when technically feasible.
- Be clear and visible, and be **in the consumer's language**.

Content that is evidently fictional, parody, satire, commentary, and criticism, **and text-only content** are excluded. It is enforced by the state attorney general or district attorneys, and according to the [Kelley Drye summary before December 21](https://www.kelleydrye.com/viewpoints/blogs/ad-law-access/pennsylvania-enacts-an-ai-ad-disclosure-law-what-companies-need-to-know-before-december-21) there is no need to prove that the advertiser was aware of it.

From California, what affects you as an SME or ecommerce is not so much the direct obligation —it falls on providers with more than one million monthly users and platforms with more than two million— but its **ripple effect**: starting January 1, 2027, large platforms must detect embedded provenance data in the content they distribute, show it to the user, and not remove it. Translated: if your file has AI metadata, the platform will read it and display it, whether you decide to put a notice on it or not.

![Layers of provenance metadata on an advertisement image that remain attached to the file even if cropped](https://cdn.aistohire.com/i/6944-pgeo-klxe-mdsd-aoyv.webp)

---

## What each platform asks for: Meta, TikTok, and YouTube do not agree

All three ask for something different and none exempt you from the law. Meta applies its label alone when it detects AI and only requires self-declaration in political or social topic ads; TikTok requires disclosure in all ads with AI content starting July 21, 2026, and self-labels by reading the Content Credentials of the file; YouTube only requires it to be declared when the result is realistic.

| Platform | Do you have to declare it yourself? | What the platform does | What is exempt |
|---|---|---|---|
| Meta (Facebook/Instagram) | Only mandatory in ads about social issues, elections, or politics with altered photorealistic image, video, or audio | Applies the "AI info" label in the "About this ad" panel when its automatic detection finds signs of AI; starting 1/6/2026, it also detects third-party generative AI tools | Minor edits: cropping, rescaling, color correction |
| TikTok | Yes, in all ads with AI content, according to the policy update of 21/7/2026 | Automatically detects and labels via Content Credentials (C2PA, integrated since January 2025); a self-applied label cannot be removed | Politically generated AI content cannot be advertised, even with a label |
| YouTube | Yes, when altered or synthetic content is **realistic**: a person, place, scene, or event that the viewer could confuse with something real | Shows the notice of altered or synthetic content next to the video or in the description | Clearly unreal or animated content, special effects, inconsequential changes, and productivity uses (scripts, ideas, subtitles) |

Automatic detection is the part that small teams underestimate the most. TikTok claims to have labeled over 1.3 billion videos by combining Content Credentials, invisible watermarks, and detection models. If your file arrives with AI metadata, the label is applied **without asking you**, and on TikTok, it is no longer removed.

### Current news: Meta paused its new terms from November 20, 2026

On **November 5, 2026**, Meta announced that it **pauses without a new date** four documents that were set to take effect on November 20: the Commercial Terms, the Product Catalog Terms, the Self-Serve Ad Terms, and the Ad Creative Generative AI Terms. This is reported in the follow-up of [the pause of the Commercial Terms from November 20](https://www.relevantaudience.com/meta/meta-pauses-commercial-terms-20-november-2026/): the previous terms remain in effect while Meta "continues to talk with stakeholders".

Why it matters to you if you buy media on Meta: the Ad Creative Generative AI Terms made the advertiser **solely responsible** for what Meta's AI functions create or modify in their ads. That is, the legal responsibility for an edit you did not make. That clause is, for now, on hold —[Search Engine Journal](https://www.searchenginejournal.com/meta-pauses-updated-business-ad-terms/596493/) confirmed the pause on the same day—, but it is worth reading it as a warning: the market trend is to push the responsibility for synthetic content onto those who fund the campaign, not onto those who provide the model.

---

## Does labeling an ad as made with AI hurt your conversion?

The available data says that **the label itself does not destroy performance, but its format does change the results**. A study by MediaScience with MediaPet and the Ehrenberg-Bass Institute for Marketing Science at the University of Adelaide, published on **June 12, 2026**, involving 900 U.S. residents, compared four ways of labeling against a control group without a label and concluded that AI labels did not negatively impact key metrics.

The numbers from the study, collected by [Marketing Dive on the effect of AI labels on advertising performance](https://www.marketingdive.com/news/ai-disclosure-labels-dont-hurt-ad-performance-heres-what-the-numbers-say/822711/):

| Metric | No label | Text 0-3 s | Text 4-6 s | Continuous text | Permanent icon |
|---|---|---|---|---|---|
| Unaided brand recall | 54 % | 60 % | 61 % | 61 % | 54 % |
| Brand recognition | 88 % | 92 % | 91 % | 87 % | 91 % |
| Positive attitude towards the brand | 51 % | 51 % | 49 % | 49 % | 44 % |
| Enjoyment of the ad | 63 % | 70 % | 70 % | 69 % | 56 % |

The useful takeaway: the **legible text notice fared better than the loose icon** in brand attitude (51 % vs. 44 %) and in ad enjoyment (70 % vs. 56 %). An icon without context leaves the user guessing what it means; a short and clear phrase does not. If you are going to choose a single way to notify in your static creatives, a brief and legible text has better empirical backing than an isolated symbol.

It is also worth not drawing the opposite conclusion: there is literature that finds negative effects on perceived authenticity, such as the review of [consumer responses to AI disclosure labels](https://journals.sagepub.com/doi/10.1177/21582440261417793) published in SAGE Open in 2026, which describes a dual effect —more novelty and interest, less sense of authenticity—. And the IAB measured a striking perception gap: only **45 %** of young consumers feel positive about AI-generated advertising, while **82 %** of executives believe they do. Labeling well is not just about compliance: it is about not staying on the wrong side of that gap.

---

## Operational Checklist: 6 Points to Mark Your Creatives Without Breaking the Flow

Compliance is not resolved on the day you upload the ad, but in how you produce the pieces. These six points are the minimum viable for a small business, an ecommerce, or a performance freelancer generating their images with AI, and they are ordered in the sequence in which they occur.

**1. Know which creatives are synthetic.** Before labeling anything, you must be able to answer, piece by piece, whether the AI generated or significantly modified the image. If your library of creatives is a folder of files without a history, that question has no reliable answer. The first thing is that each piece is born with that mark on it.

**2. Keep the provider, model, and prompt of each piece.** It is the record that allows you to demonstrate what happened, and it is exactly the type of data that California calls *system provenance data*: provider, system version, creation date, and identifier. Keep it even if no one asks for it today: in six months, reconstructing it will be impossible.

**3. Do not delete metadata when cropping or re-exporting.** This is where almost everyone fails. Passing the image through an editor, cropping it to the dimensions of each platform, or re-exporting it to another format can **eliminate the Content Credentials** that the model embedded. And Article 50 requires the provider to include them; if you destroy them in the next step, the file arrives clean and without traceability. Check what your export pipeline does with the metadata.

![The three orientations of an ad (square, horizontal, and vertical) with the AI notice placed in a different corner in each](https://cdn.aistohire.com/i/0734-pkjd-mdac-cldr-tgez.webp)

**4. Decide where the visible notice goes in each orientation.** A notice in the bottom right corner works in a 1:1 square feed and gets covered by the interface in a 9:16 Stories. Define a safe zone by orientation and respect it: Pennsylvania's regulation requires that the notice be in the same medium as the content and in the first presentation, and the European guideline asks that it remains visible if the piece is downloaded or reshared.

**5. Translate the notice to each market.** Pennsylvania explicitly requires the notice **in the consumer's language**, and the European label is localized ("AI" in English, "KI" in German). A notice in Spanish on a creative running in Germany does not comply. And a typo in the notice is worse than not putting it: it leaves you with non-compliance and also visual noise.

![Fan of ten versions of the same ad with the same AI notice for different markets and languages](https://cdn.aistohire.com/i/7054-vugn-wngn-wpmj-ttoc.webp)

**6. Review the notice before uploading the batch.** The final check is visual and on the final piece, not on the original: that the notice is there, that it is readable against the background, that it hasn’t been left out of the crop, and that it says what it needs to say in the correct language. A batch of twenty creatives in three formats and five languages is sixty pieces: without an explicit review step, some may go out without a notice.

---

## Frequently Asked Questions About Labeling Ads Made with AI

### Do I have to put "made with AI" on all my ads if I use generated images?

Not on all, but yes on those that might make the consumer believe something false about what you are selling. The common criterion in the EU, Pennsylvania, and the platforms is **appearance of reality**: if the image looks like a photo of something real and it is not, you must notify. An obviously unrealistic illustration or a minor retouch (cropping, color) does not require it.

### What happens if I do not label an ad generated with AI?

It depends on where it is published. In Pennsylvania, starting December 21, 2026, the Attorney General or a district attorney can request that the ad be removed, and the judge can impose up to $3,000 per violation plus costs. In New York, it is $1,000 the first time and $5,000 for subsequent times. In the EU, the national supervisor — AESIA in Spain — can act on the obligations of Article 50. And on the platforms, the immediate risk is the rejection of the ad or an automatically applied label that you can no longer remove.

### Is it enough for the platform to put the AI label?

No, because the platform's label and the legal obligation are not the same thing. Meta, for example, places its "AI info" label within the "About this ad" panel, which the user has to open; Pennsylvania's rule requires a clear notice **in the same medium and at the first presentation** of the content. Relying on the platform's automatic detection also leaves you without control: the label may be applied to pieces that you did not consider synthetic.

### Is the official EU icon mandatory?

No. The three icons published by the Commission along with the Code of Good Practices from June 10, 2026, are for **voluntary** use and free of charge, with no attribution required. What is mandatory is the labeling obligation of Article 50; the icon is a convenient and recognizable way to comply, but you can use your own text label.

### Is a prompt written with AI or a copy generated by an LLM also labeled?

In commercial advertising, generally not. Pennsylvania **expressly excludes content that is only text**, and in the EU, labeling of text generated by AI is limited to texts on public interest issues. What is labeled is the image, video, and audio with a realistic appearance.

### Can cropping the image to the dimensions of each platform leave me non-compliant?

Yes, in two ways. It can **remove the provenance metadata** that the model embedded, leaving the file without machine-readable marking, and it can **leave out the visible notice** that you had placed. This is the most common mistake and the easiest to avoid: check the export pipeline and review the final piece, not the original.

### Will labeling lower my CTR?

The data does not support that fear in brand metrics. The study by MediaScience and the Ehrenberg-Bass Institute from June 12, 2026, found no negative impact on recall, recognition, or sentiment towards the ad, and the text notice performed better than the standalone icon. That said, there is research that detects a loss of perceived authenticity, so the form of the notice matters: brief, clear, and well-placed.

## The Compliance Trail is Built When Generating, Not After

Almost everything these rules ask for — knowing which piece is synthetic, with which model it was made, what was edited and when, and being able to stamp the correct notice in each format and language — is information that exists at the moment of generating the creativity and is lost if not saved there. Reconstructing it afterwards, based on a folder of loose files, is the expensive work.

**Ads Image Studio**, the studio of **AIsToHire** for [creating ad images with AI](/ads-image-studio-create-ad-images-with-ai-166), leaves that trail natively: each version saves the provider, the model, the prompt, and the instruction with which it was generated, and the version tree documents what was edited and when, because each iteration is a new version and the previous one remains in the history. For the visible notice, the **text overlay** mode allows you to stamp it with exact typography over the already approved visual, in the 10 available languages, without regenerating the art and without risking the typical typos of text drawn by the model. And as the final dimensions are resolved upon export with smart cropping — not asking the model for pixels — the notice is decided on the approved piece and is not left to the randomness of each recropping.

> Create your creatives with the compliance trail already in place: provider, model, and prompt saved in each version, and the AI notice stamped in exact typography in 10 languages on the already approved visual. Sign up for free at Ads Image Studio and generate your first campaign today.

- [Ads Image Studio — Create Your Ad Images with AI](https://aistohire.com/ads-image-studio-create-ad-images-with-ai-166)

## About the platform

**Ads Image Studio — Create Your Ad Images with AI** — an app by AIstoHire. Available at: https://aistohire.com/ads-image-studio-create-ad-images-with-ai-166/blog/labeling-ads-ai-law-2027-mark-creatives

### Other languages

- `es`: https://aistohire.com/es/ads-image-studio-crear-imagenes-anuncios-ia-166/blog/etiquetar-anuncios-ia-ley-2027-marcar-creatividades
- `fr`: https://aistohire.com/fr/ads-image-studio-creer-images-publicitaires-ia-166/blog/etiqueter-annonces-ia-loi-2027-marquer-creativites
- `it`: https://aistohire.com/it/ads-image-studio-crea-immagini-annunci-ia-166/blog/etichettare-annunci-ia-legge-2027-etichettare-creativita
- `pt`: https://aistohire.com/pt/ads-image-studio-crie-imagens-anuncios-ia-166/blog/etiquetar-anuncios-ia-lei-2027-marcar-criatividades
- `de`: https://aistohire.com/de/ads-image-studio-erstellen-anzeigenbilder-ki-166/blog/anzeigen-kennzeichnen-ki-gesetz-2027-kreativitaeten-markieren
- `bg`: https://aistohire.com/bg/ads-image-studio-izobrazheniya-za-reklami-s-ii-v-3-formata-166/blog/etiketirane-na-reklami-s-ii-zakon-2027-markirane-na-kreativi
- `ja`: https://aistohire.com/ja/ai-kokoku-gazo-sakusei-166/blog/ai-kokoku-hyoji-hourei-2027-creative-label
- `nl`: https://aistohire.com/nl/ads-image-studio-advertentiebeelden-met-ai-in-3-formaten-166/blog/ai-advertenties-labelen-wet-2027-creaties-markeren
- `ko`: https://aistohire.com/ko/ads-image-studio-ai-gwanggo-image-166/blog/ai-gwanggo-pyosi-beop-2027-gwanggo-sojae-pyosi
- `sv`: https://aistohire.com/sv/ads-image-studio-annonsbilder-med-ai-i-3-format-166/blog/markning-av-ai-annonser-lag-2027-markera-annonsbilder
- `da`: https://aistohire.com/da/ads-image-studio-annoncebilleder-med-ai-i-3-formater-166/blog/maerk-ai-annoncer-lov-2027-kreativer
- `ar`: https://aistohire.com/ar/ads-image-studio-sur-ialanat-baldhka-alastnaai-b-3-tnsiqat-166/blog/label-ai-ads-2027-law-mark-creatives
- `no`: https://aistohire.com/no/ads-image-studio-annonsebilder-med-ki-i-3-formater-166/blog/merke-annonser-ki-lov-2027-merke-kreativer
- `fi`: https://aistohire.com/fi/ads-image-studio-mainoskuvat-tekoalylla-kolmessa-formaatissa-166/blog/merkitse-tekoalymainokset-laki-2027-sisaltojen-merkinta
- `zh`: https://aistohire.com/zh/ads-image-studio-ai-ad-creatives-166/blog/ai-guanggao-biaoshi-2027-fagui-sucai-biaoji
- `pl`: https://aistohire.com/pl/ads-image-studio-obrazy-reklamowe-z-ai-w-3-formatach-166/blog/oznaczanie-reklam-ai-prawo-2027-jak-oznaczac-kreacje
- `he`: https://aistohire.com/he/ads-image-studio-tmonot-lmodaot-bbinh-mlakotit-b-3-pormtim-166/blog/letaveg-modaot-bina-melachutit-chok-2027-simun-kriyativim
- `cs`: https://aistohire.com/cs/ads-image-studio-obrazky-pro-reklamy-s-ai-ve-3-formatech-166/blog/oznaceni-reklam-ai-zakon-2027-oznacit-kreativy
- `id`: https://aistohire.com/id/ads-image-studio-gambar-iklan-dengan-ai-dalam-3-format-166/blog/melabeli-iklan-ai-hukum-2027-menandai-materi-iklan
