# Should AI-generated images be labeled in 2026? What Article 50 of the AI Act demands

> Do AI-generated images need to be labeled? What Article 50 of the AI Act mandates starting August 2026, what edits are exempt, and what penalties exist.

- **Source (source):** https://aistohire.com/aipicflow-create-images-with-ai-and-edit-photos-53/blog/labeling-ai-images-ai-act-article-50
- **Platform (platform):** AIstoHire
- **Language (language):** en
- **Updated (updated):** 2026-11-20

**In summary:**

- Since **August 2, 2026**, Article 50 of the European AI Regulation (AI Act) applies: anyone publishing an image generated or manipulated with AI that **could be mistaken for real** must clearly and visibly indicate it ([European Commission, FAQ on Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)).
- There are **two distinct layers**: the invisible technical mark within the file (provided by the model supplier) and the **visible label** placed by the publisher. The first does not replace the second.
- The **standard retouching** that does not alter the substance of the image —exposure, color, minor cropping, removing a speck, increasing resolution— is **excluded** from the marking obligation.
- A **generated background**, a **virtual model**, or a scene that never existed **does** count: the criterion is whether the result appears authentic.
- Fines can reach **15 million euros or 3% of global revenue**; the deadline for the technical marking of systems that were already on the market is **December 2, 2026**.

---

We have been hearing for two years that "images generated by AI will need to be labeled." It is no longer a future concern: the obligation takes effect from the summer of 2026 and affects anyone publishing images in the European Union, from a content creator with 3,000 followers to an online store updating its catalog. At AIsToHire, we work daily with image generation and editing —that’s exactly what aiPicFlow does, our tool to [create images with AI and edit photos by writing](/aipicflow-create-images-with-ai-and-edit-photos-53)—, and the question we receive the most is not "how is it generated?", but "does this I just created have a label or not?" This article answers that, case by case, with the rule and official guidelines in hand.

> Notice: this is a practical explanation, not legal advice. If your case is sensitive (health, politics, minors, healthcare, or regulated advertising), consult a professional.

## When is it mandatory to label images generated with AI in the EU?

Since **August 2, 2026**. On that day, the transparency obligations of Article 50 of the AI Act began to apply, requiring generative AI providers to mark their outputs in machine-readable format and publishers to notify when the content is artificial and could be confused with real material ([European Commission, FAQ on Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)). There was no general extension: the package known as the digital omnibus —Regulation (EU) 2026/1744, effective from July 27, 2026— delayed the obligations for high-risk systems, but **left the timeline of Article 50 intact** ([AI Act Explorer, Digital Omnibus](https://artificialintelligenceact.eu/ai-act-explorer/digital-omnibus/)).

These are the dates that matter:

| Date | What happens |
|---|---|
| August 2, 2026 | The transparency obligations of Article 50 apply: notice of interaction with AI, marking of synthetic content, and labeling of ultra-fakes (*deepfakes*) |
| December 2, 2026 | End of the transitional period for machine-readable marking in generative systems that were already on the market before August 2, 2026 |
| February 2, 2027 | Committed date in the Code of Good Practices for interoperability in watermark detection |
| December 2, 2027 | The obligations for high-risk systems in Annex III apply, postponed by the digital omnibus |

Sources: [European Commission FAQ on Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act) and the analysis by [Paul Weiss on the final transparency rules](https://www.paulweiss.com/insights/client-memos/eu-finalises-transparency-rules-for-ai-generated-content) (July 20, 2026).

One detail that reassures many people: **content published before August 2, 2026, does not need to be retroactively labeled**. The obligation is not retroactive, according to the Commission's own FAQ.

## What does Article 50 say exactly and who is it binding?

Article 50 divides the burden between two figures: the **provider** (who develops and markets the AI system) and the **deploying party** or *deployer* (who uses it under their authority, except for purely personal use). The provider must label outputs in a machine-readable format; those who publish must notify the public in a perceptible manner. If you use an image generation tool for your brand, your client, or your feed, **you are the latter**: the one who has to put the visible label ([text of Article 50](https://artificialintelligenceact.eu/article/50/)).

The specific obligations:

1. **Article 50(1)** — notify that one is interacting with an AI, unless it is evident.
2. **Article 50(2)** — the provider labels synthetic outputs (audio, image, video, text) in a machine-readable and detectable manner, with "effective, interoperable, robust, and reliable solutions to the extent technically feasible."
3. **Article 50(4)** — those who publish reveal that the image, audio, or video has been artificially generated or manipulated when it constitutes an ultra-fake, and the same applies to texts on matters of public interest without human editorial review.
4. **Article 50(5)** — the information is provided "in a clear and distinguishable manner, no later than at the first interaction or exposure" and meets accessibility requirements.

The Commission published its final guidelines on these obligations on **July 20, 2026** ([European Commission press release](https://digital-strategy.ec.europa.eu/en/news/commission-publishes-guidelines-transparency-obligations-providers-and-deployers-certain-ai-systems)), and earlier, on **June 10, 2026**, the [Code of Good Practice on Transparency of AI-Generated Content](https://digital-strategy.ec.europa.eu/en/faqs/code-practice-transparency-ai-generated-content), which is voluntary. The Commission and the European AI Committee declared it adequate on July 8 and 9, 2026 ([Reed Smith](https://www.reedsmith.com/our-insights/blogs/viewpoints/102nbz0/transparency-obligations-for-ai-generated-content-the-code-of-practice-adequacy/)). By July 31, 2026, around **190 organizations** had signed it, including 82 providers and 152 deploying parties from sectors such as technology, telecommunications, education, and retail ([DLA Piper](https://www.dlapiper.com/en/insights/publications/2026/07/what-the-eus-new-code-of-practice-means) and [Morrison Foerster](https://www.mofo.com/resources/insights/260911-full-transparency-organizations-sign-eu-code)). Signing is not mandatory, but those who do not sign must demonstrate compliance through other means and are subject to more information requirements.

### The deciding criterion: Does it look authentic?

The key word of Article 50(4) is *deepfake*, and the guidelines interpret it broadly. It includes content that resembles something that **exists, could plausibly exist, or could have existed**, including photorealistic portraits of invented people. Manifestly fantastic or impossible content is excluded. What is evaluated is the degree of resemblance, the underlying message, the context of publication, and what the public expects to see ([Reed Smith](https://www.reedsmith.com/our-insights/blogs/viewpoints/102nbz0/transparency-obligations-for-ai-generated-content-the-code-of-practice-adequacy/), [Paul Weiss](https://www.paulweiss.com/insights/client-memos/eu-finalises-transparency-rules-for-ai-generated-content)).

Translated into the trade: it does not matter if you used AI, what matters is if **someone can believe that the photo was taken with a camera**.

## The two layers of labeling: the invisible mark and the visible notice

The European system operates at two levels and it is important not to confuse them. The **first layer** is technical and invisible: signed metadata and imperceptible watermarks within the file, provided by the model supplier. The **second layer** is the visible label added by the publisher, aimed at a human looking at the screen. The Commission is explicit: **machine-readable marking does not alone satisfy the disclosure obligation of the publisher** ([European Commission FAQ](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)).

![The same photograph printed twice: on the left with translucent sheets representing the invisible technical mark in the file, on the right with a visible plate in the corner representing the label of the publisher](https://cdn.aistohire.com/i/0754-dgst-dtey-paec-oimy.webp)

| | Layer 1: technical mark | Layer 2: visible label |
|---|---|---|
| Who puts it | The AI model supplier | The publisher of the image |
| Legal basis | Article 50(2) | Article 50(4) |
| How it looks | Not visible: signed metadata + imperceptible watermark | Perceptible text or icon without technical tools |
| For whom | Platforms, authorities, verifiers, detection models | The public viewing the image |
| If it fails | Can be removed when recompressed or when taking a screenshot | Disappears if you don't put it yourself |

The Code of Good Practices recommends a **layered marking**, because today no technique alone meets all four requirements of the standard: metadata with timestamp and digital signature that record whether the content is generated or manipulated, **plus** an imperceptible watermark embedded in the content itself as a backup when metadata is lost. The signing suppliers also commit to providing free detection tools and not promoting the removal of marks ([DLA Piper](https://www.dlapiper.com/en/insights/publications/2026/07/what-the-eus-new-code-of-practice-means)).

In practice, the standards that have emerged are the content credentials **C2PA** and **SynthID** type watermarks. Google has integrated SynthID into its generative products, OpenAI uses C2PA metadata and SynthID marks in images and audio, Anthropic announced watermarks in text based on SynthID-Text, and Synthesia incorporates provenance signals compatible with C2PA ([Morrison Foerster](https://www.mofo.com/resources/insights/260911-full-transparency-organizations-sign-eu-code)).

A very practical notice: **metadata is easily lost**. If you export, recompress in your CDN, or take a screenshot, layer 1 can disappear along the way. That’s why layer 2 —the label you put on— is the one that truly protects you.

## This does have a label, this does not: real cases

The boundary is not "I used AI", but whether the system **substantially alters** the image or its meaning. Article 50(2) expressly exempts systems that "perform an assistive function for standard editing" or that "do not substantially alter the input data or its semantics". The guidelines from July 2026 provide directly applicable examples for e-commerce: color correction, resolution scaling, background removal, and minor cropping are exempt; **generating a synthetic model showcasing a garment is not**.

![Split composition: on the left a cup photographed with a color chart and reflector, light and color retouching; on the right the same cup in a fully generated synthetic scene](https://cdn.aistohire.com/i/8464-pmmf-prdl-bkng-kmil.webp)

| Real case | Is there a visible label? | Why |
|---|---|---|
| Adjusting exposure, contrast, and white balance of a real photo | **No** | Standard editing, does not alter the substance of the image |
| Removing a speck of dust, a hair, or a reflection from the product | **No** | Minor cosmetic retouching, without change of meaning |
| Minor cropping or aspect ratio adjustment for a social media format | **No** | Assistive editing expressly cited in the guidelines |
| Increasing the resolution of an old catalog photo | **No** | Scaling does not add content that wasn't there |
| Cropping the product against a white background (removing the background) | **No** | Background removal is among the exempt examples |
| Replacing the background with a generated lounge that never existed | **Yes** | A new scene is generated that looks like a real photo |
| Virtual model, non-existent, showcasing your clothing | **Yes** | The guidelines exclude synthetic models from exemption; a photorealistic portrait of an invented person counts |
| Changing the color of a product to show another variant | **Depends** | If the result still looks like an authentic photo of the product, it falls into manipulation territory; moreover, consumer regulations require that the image does not mislead about what you are selling |
| Removing a tourist from the photo of your hotel terrace | **Yes, in general** | It is not a speck: it changes what the viewer understands about the scene (how crowded the place is) |
| Placing the face of a real person in another scene | **Yes, always** | It is the nuclear case of ultra-falsification; moreover, there are image rights at stake |
| Illustration, comic, or manifestly unreal render | **No** under Article 50(4) | It does not intend to pass as authentic. Note: if it is advertising, the rules against misleading still apply |

The three-second mental rule: **if removing the label would make someone think that is a photograph of something that happened, the label is necessary**.

## How to label on Instagram, TikTok, and YouTube?

The major platforms already have their own declaration system and, in part, they apply it on their own. **YouTube** requires the creator to declare upon upload any content generated or altered that appears realistic —a real person saying something they didn’t say, images of a real place or event altered, a realistic scene that never occurred— and displays the label in the description, or prominently over the video on sensitive topics such as health, news, elections, or finance; it does not require it for manifestly unrealistic, animated, or special effects content ([YouTube Help on altered or synthetic content](https://support.google.com/youtube/answer/14328491)). **TikTok** was the first video platform to adopt **C2PA** content credentials, which automatically recognizes and labels AI-generated content coming from other tools, and attaches those credentials to the content created on TikTok so that they survive the download ([TikTok Press Room](https://newsroom.tiktok.com/en-us/partnering-with-our-industry-to-advance-ai-transparency-and-literacy)).

That does not exempt you from anything. Three practical rules:

- **Activate the AI-generated content declaration switch** that each platform offers at the time of publication. It is the quickest and most reliable way for the label to appear where the user sees it.
- **Do not rely solely on automatic detection.** If you recompress the file or pass it through an editor that deletes metadata, the platform may not detect anything.
- **The label must be visible on the first exposure**, not after two taps. The Code of Good Practices recommends placing it in a top corner, next to the headline or at the bottom of the piece, "immediately recognizable without user interaction" and visible for a sufficient amount of time; in audio, a spoken warning at the beginning, and in long or live video, repeated at intervals ([DLA Piper](https://www.dlapiper.com/en/insights/publications/2026/07/what-the-eus-new-code-of-practice-means)).

In other words: a `#AI` lost among fifteen *hashtags* at the end of the text is not a clear and distinguishable label.

## And in an online store or advertising campaign?

In e-commerce and advertising, two layers of requirements accumulate: that of Article 50 and that of consumer regulations, which prohibit misleading about the characteristics of the product. The label must be **next to the image on the product sheet or in the creative**, not buried in the legal notice at the bottom of the page.

![Hand holding a mobile with a grid of three product photos, each with a small notice plate in the corner, in an e-commerce office](https://cdn.aistohire.com/i/3206-gdjm-muhw-hdef-lwhy.webp)

What works in practice:

- **Product sheet:** a short notice under the gallery ("Set image generated with AI. The product is shown in photos 1 and 2") and the real unaltered photo as the first image of the carousel.
- **Paid campaigns:** campaigns with AI-generated content must clearly signal this in the EU from August 2, 2026 ([MarketingDirecto](https://www.marketingdirecto.com/actualidad/nueva-normativa-europea-inteligencia-artificial-publicidad)). In Spain, the [Code of Conduct for advertising through influencers](https://spainaudiovisualhub.digital.gob.es/en/actualidad/autocontrol--asociacion-espanola-de-anunciantes--iab-spain---cod) promoted by Autocontrol, the Spanish Association of Advertisers, and IAB Spain, in effect since October 1, 2025, already incorporates the criterion of informing when promotional content has been generated or modified with AI.
- **Massive catalogs:** document the process. If your retouching flow is always the same —color correction and cropping—, being able to demonstrate it saves you the discussion about whether a label was necessary.
- **A virtual model is not a shortcut free of obligations:** it fully falls under Article 50(4) and must be stated.

## What penalties exist and who oversees in Spain?

Failure to comply with the transparency obligations of Article 50 can be sanctioned with up to **15 million euros or 3% of the total global turnover** from the previous fiscal year, whichever is greater, with a proportionality clause for SMEs and small mid-cap companies ([EU Commission FAQ](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act)). National market surveillance authorities oversee; the European AI Office intervenes in limited cases.

In Spain, the **Organic Law Project for the proper use and governance of artificial intelligence** was published in the Official Bulletin of the Cortes on **June 12, 2026**, but **lapsed on October 6, 2026** when the Cortes were dissolved, without being approved. It designated **AESIA** as the market surveillance authority and single point of contact, with the AEPD for biometric systems and prohibited practices, and the Bank of Spain and CNMC in their areas. Its sanctioning regime included tiers of up to 35 million euros or 7% of turnover for prohibited practices, 15 million or 3%, 7.5 million or 1%, and 500,000 euros or 0.5% for minor infractions ([finreg360](https://finreg360.com/alerta/el-congreso-inicia-el-tramite-del-proyecto-de-ley-organica-para-el-buen-uso-y-la-gobernanza-de-la-inteligencia-artificial/), [Economist & Jurist](https://www.economistjurist.es/zbloque-1/ley-organica-de-ia-espana-aterriza-el-ai-act-con-aesia-sanciones-y-sandboxes/)).

For an individual creator or a small store, the realistic risk is not the 15 million: it is the removal of content, the loss of reach due to the label applied automatically by the platform, consumer claims, and the reputational damage of being caught selling a photo that was not a photo.

## Checklist in seven points before publishing

Before hitting publish, these seven checks resolve 90% of cases:

1. **Does it look like a real photograph?** If the answer is yes and there is generative AI involved, label it.
2. **What exactly did the AI do?** Separate standard retouching (light, color, cropping, resolution, cut-out background) from generating new content (backgrounds, scenes, people, objects).
3. **Does someone recognizable appear?** If there are identifiable faces or places, in addition to the label, check image rights and data protection.
4. **Is the label immediately visible?** On the image or right next to it, not among the *hashtags*.
5. **Have you activated the platform's declaration?** It is the cleanest route on social media.
6. **Are you keeping the metadata?** Ensure that your export and your CDN do not erase the file's provenance.
7. **Can you prove what you did?** Keep the instruction, the source image, and the parameters of each piece.

That last point is more important than it seems: when the line between "retouching" and "manipulation" is debatable, what saves you is the record of what you requested and on which image. Tools like aiPicFlow keep the **complete recipe for each image** —instruction, style, format, source image, and parameters— so you can reconstruct and justify any piece months later.

## Frequently Asked Questions

### Do I have to label a photo where I've only changed the light and color with AI?

No. Article 50(2) exempts systems with "assistance function for standard editing" that do not substantially alter the input data or its semantics, and the Commission's guidelines explicitly cite color correction and scaling as exempt examples. If all you did was reveal a photo that exists better, there is no label.

### Is it enough to say "made with AI" in the caption or in the hashtags?

Not reliably. Article 50(5) requires that the information be provided "clearly and distinguishably, no later than at the time of the first interaction or exposure." The Code of Good Practices states that the label should be immediately recognizable without user interaction: a mention at the end of a long text or among *hashtags* does not meet that standard.

### If Instagram or TikTok puts the automatic label, have I complied?

It depends on whether the label truly appears and is visible at first glance. Automatic detection relies on C2PA metadata or detection models, and both can fail if the file is recompressed or screen-captured. It is prudent to activate the AI-generated content declaration offered by each platform and not wait for the system to guess it.

### What if the image is clearly unreal, like an illustration or drawing?

Article 50(4) points to content that can be confused with authentic material, and the guidelines exclude what is manifestly fantastic or impossible. An illustration that no one would confuse with a photograph does not trigger the obligation. It is different if it is part of commercial communication: there, the prohibition against misleading remains in force.

### Who is responsible if it is not labeled, me or the tool I used?

Both, for different obligations. The system provider is responsible for machine-readable marking (article 50.2); the publisher is responsible for the visible label (article 50.4). Just because your tool incorporates watermarks does not exempt you from putting the perceptible notice in your publication.

### Do I have to label the images I published before August 2026?

No. According to the FAQ from the European Commission, content generated before August 2, 2026, does not require retroactive labeling. The obligation applies to what you generate and publish from that date onwards.

### In my online store, does the product photo on a white background need a label?

If all you have done is crop the background, correct the color, and scale the real photo of the product, no. If you have generated a new environment, an ambient scene, or a model that does not exist, yes. And in any case, the first image of the listing should be an unaltered real photo of what you are selling.

### What happens on December 2, 2026?

The four-month transitional period for machine-readable marking of generative systems that were already on the market before August 2, 2026, expires. From then on, all generative AI tools operating in the EU should technically mark their outputs. Then comes February 2, 2027, the date committed in the Code of Good Practices for watermark detection to be interoperable.

## Create knowing which side of the line you are on

The rule does not punish the use of AI: it punishes passing off what is not real as real without saying so. Once you have a clear map —standard retouching out, generation of what looks real in, visible label always the responsibility of the publisher— working with AI images is as calm as before, and much faster.

That is the idea with which we built aiPicFlow at AIsToHire: you can **generate an image from a phrase** when you need a scene you do not have, and you can **retouch your photos by writing what you want to change** —remove an object, change or remove the background, extend the frame, improve resolution, adapt the ratio— knowing in each case whether that use falls under the obligation to label or not. Moreover, each image keeps its reproducible recipe, which is exactly the documentary trail you need when someone asks how it was made. If you want to try it, go to [aiPicFlow to create images with AI and edit photos](/aipicflow-create-images-with-ai-and-edit-photos-53) and start with a phrase or with one of your photos.

> Create and edit images knowing when labeling is necessary and when it isn’t: try aiPicFlow for free.

- [aiPicFlow — Create images by typing and retouch your photos without editing skills](https://aistohire.com/aipicflow-create-images-with-ai-and-edit-photos-53)

## About the platform

**aiPicFlow — Create images by typing and retouch your photos without editing skills** — an app by AIstoHire. Available at: https://aistohire.com/aipicflow-create-images-with-ai-and-edit-photos-53/blog/labeling-ai-images-ai-act-article-50

### Other languages

- `es`: https://aistohire.com/es/aipicflow-crear-imagenes-con-ia-y-editar-fotos-53/blog/etiquetar-imagenes-ia-ai-act-articulo-50
- `fr`: https://aistohire.com/fr/aipicflow-creer-des-images-avec-ia-et-editer-des-photos-53/blog/etiqueter-images-ia-loi-sur-lia-article-50
- `it`: https://aistohire.com/it/aipicflow-crea-immagini-con-ia-e-modifica-foto-53/blog/etichettare-immagini-ia-ai-act-articolo-50
- `pt`: https://aistohire.com/pt/aipicflow-criar-imagens-com-ia-e-editar-fotos-53/blog/rotular-imagens-ia-lei-de-ia-artigo-50
- `de`: https://aistohire.com/de/aipicflow-bilder-mit-ki-erstellen-und-fotos-bearbeiten-53/blog/bilder-kennzeichnen-ia-ai-act-artikel-50
- `bg`: https://aistohire.com/bg/aipicflow-sazdavane-na-izobrazheniya-s-ii-i-redaktsiya-na-snimki-s-pisane-53/blog/etiketirane-na-izobrazheniya-ii-ai-act-chlen-50
- `ja`: https://aistohire.com/ja/aipicflow-create-and-edit-images-53/blog/ai-gazo-label-ai-act-dai50jo
- `nl`: https://aistohire.com/nl/aipicflow-afbeeldingen-maken-met-ai-en-foto-s-bewerken-door-te-typen-53/blog/ai-afbeeldingen-labelen-ai-act-artikel-50
- `ko`: https://aistohire.com/ko/aipicflow-ai-image-saengseong-pyeonjip-53/blog/ai-imiji-rabel-ai-act-jae50jo
- `sv`: https://aistohire.com/sv/aipicflow-skapa-bilder-med-ai-och-redigera-foton-genom-att-skriva-53/blog/markning-ai-bilder-ai-act-artikel-50
- `da`: https://aistohire.com/da/aipicflow-lav-billeder-med-ai-og-rediger-fotos-ved-at-skrive-53/blog/maerke-ai-billeder-ai-forordningen-artikel-50
- `ar`: https://aistohire.com/ar/aipicflow-insha-sur-baldhka-alastnaai-wthrir-alsur-balktaba-53/blog/label-ai-images-ai-act-article-50
- `no`: https://aistohire.com/no/aipicflow-lag-bilder-med-ki-og-rediger-bilder-ved-a-skrive-53/blog/merke-ki-bilder-ai-act-artikkel-50
- `fi`: https://aistohire.com/fi/aipicflow-luo-kuvia-tekoalylla-ja-muokkaa-valokuvia-kirjoittamalla-53/blog/tekoalykuvien-merkitseminen-ai-act-50-artikla
- `zh`: https://aistohire.com/zh/ai-image-generator-photo-editor-53/blog/ai-image-labeling-ai-act-article-50
- `pl`: https://aistohire.com/pl/aipicflow-tworz-obrazy-z-ai-i-edytuj-zdjecia-piszac-53/blog/oznaczanie-obrazow-ai-ai-act-artykul-50
- `he`: https://aistohire.com/he/aipicflow-ytsirt-tmonot-bbinh-mlakotit-varikt-tmonot-bktibh-53/blog/tiug-tmunot-bina-melachutit-ai-act-seif-50
- `cs`: https://aistohire.com/cs/aipicflow-tvorba-obrazku-s-ai-a-upravy-fotek-psanim-53/blog/oznacovani-obrazku-ai-ai-act-clanek-50
- `id`: https://aistohire.com/id/aipicflow-buat-gambar-dengan-ai-dan-sunting-foto-dengan-menulis-53/blog/label-gambar-ai-ai-act-pasal-50
