Label It Before Someone Else Does: AI Image Disclosure Rules in 2026

For most of the short history of AI images, labelling was etiquette. Some people added "made with AI" to a caption, most did not, and nothing happened either way. In 2026 that changed in three places at once: European law, Californian law and the upload rules of the platforms where images are actually seen.

This is not legal advice. It is a plain map of who owes what, with links to the texts themselves, and an honest note on where a local tool leaves the work.

Do You Have to Label AI-Generated Images?

Increasingly, yes, if the image looks real and you publish it. The rules split the job between the people who build generators (providers) and the people who use them to publish (deployers, creators and uploaders).

The EU AI Act: Two Different Duties

Article 50 of the AI Act carries both. Under 50(2), providers of systems "generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated." The Act applies "from 2 August 2026". The Digital Omnibus amendment, in force since 27 July 2026, gives systems already on the market before that date until 2 December 2026 to comply with 50(2). The deep fake duty was not delayed.

Under 50(4), "Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated." Art is not exempt, but the duty is lighter: for "evidently artistic, creative, satirical, fictional or analogous" work, disclosure only has to happen "in an appropriate manner that does not hamper the display or enjoyment of the work."

Who counts as a deployer matters. Article 3 defines one as anyone "using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity." A hobbyist posting for fun is outside it. A studio, an agency or a freelancer delivering client work is inside it.

In June 2026 the Commission published a Code of Practice on marking and labelling AI-generated content. "The Code is voluntary", but once approved, signatories "will be able to demonstrate compliance" by following it.

California: Big Providers First, Platforms Next

California's AI Transparency Act, as amended by AB 853, became operative on 2 August 2026. It reaches "covered providers": generative systems with "over 1,000,000 monthly visitors or users" that are publicly accessible in the state, with penalties of $5,000 per violation per day. The same bill adds duties for large online platforms from 2027 and for capture-device makers from 2028.

The Platforms Already Ask

• YouTube requires creators "to disclose AI-generated or altered content that seems realistic", and not for content that is clearly unrealistic or only minor edits.
• TikTok requires creators "to label AI-generated or significantly edited content that shows realistic-looking scenes or people", and says unlabelled content "may be removed, restricted, or labeled by our team."
• Meta applies "AI info" labels across its apps, using its own detection and industry signals.

The common thread is in the title of this post. If you do not label realistic content, the platform may label it for you, restrict it, or take it down.

Metadata Alone Will Not Carry the Label

The technical standard for marking is C2PA Content Credentials: a signed manifest whose assertions record an asset's origin, its modifications and "use of AI (i.e., how it was authored)." The Content Authenticity Initiative publishes open-source tools to attach them.

Metadata is fragile, and the standard says so. The C2PA explainer asks "Can the provenance metadata be removed?" and answers "Yes it can", which is why it pairs metadata with watermarking and fingerprinting. NIST's report on synthetic content (AI 100-4) notes that metadata "can similarly be stripped altogether, as it often is when files are shared (e.g., via social media platforms)." A visible label in the caption or the frame survives where a hidden one may not.

Where Vision Studio Leaves the Job

Here is the honest part. Vision Studio saves plain PNG images and MP4 video. It does not write a watermark, an AI mark or Content Credentials of its own into the file, and the prompt and settings stay in the app's local library and iteration tree rather than inside the image. Your files leave the app clean.

That is a privacy feature, since nothing about your prompt travels with the picture. It also means that when an image goes out into the world, labelling it is your step, not the tool's. For work that has to be labelled:

1. Decide whether it is realistic. Realistic people, places and events are where every rule above applies.
2. Label visibly. A caption, an on-image note or the platform's own AI toggle. Do not rely on file metadata.
3. Add Content Credentials before publishing if a client or platform asks for machine-readable provenance, using the open-source C2PA tools.
4. Keep your record. The iteration tree stores the prompt, model, seed and settings behind each render, so you can show how an image was made if anyone asks.

The Rule of Thumb

If it looks real and you are publishing it for work, label it where people will see it. The law now expects it, the platforms enforce it, and a label you add yourself is always better than one someone else adds for you.

Vision Studio is free, open-source (MIT), and runs locally on your GPU. What it saves, where, and what it sends over the network are documented with their source files at vision-studio-x.com/features. Download it at vision-studio-x.com/download.

-Rocky

#AIDisclosure #EUAIAct #C2PA #ContentCredentials #LocalAI #OpenSource #VisionStudio