AI labeling with mintys: label it simply, prove it durably

What is mintys?

mintys is the AI labeling solution by on:mint. It labels AI-generated and AI-altered content visibly and machine-readably following the C2PA standard, backs the labeling with an invisible watermark, and records evidence of when labeling happened and by whom. If your AI tool has already written provenance information into the file, mintys carries it over. mintys comes as a web app, as an interface (API) and as an MCP connection for agent workflows.

mintys is made for everyone who publishes media professionally and wants to stay on the right side of the law: marketing teams and communications departments, online shops and manufacturers, agencies and photographers, journalists, influencers, restaurants and hotels, estate agents, trade businesses, clubs and associations. The AI Act, the EU's AI regulation, splits the duties across two roles, and labeling looks different in each.

A deployer is anyone using an AI system under their own authority. A deployer publishing convincingly real AI content professionally has had to disclose visibly, since 2 August 2026, that it was artificially generated or altered. Visibly means: the person looking at the image recognises it immediately.

A provider is anyone who develops an AI system, or has one developed, and offers or puts it into service under their own name. Providers have to mark AI content machine-readably so that software detects it automatically. This role catches more companies than expected: anyone building their own application on top of someone else's AI model, say an image generator inside their own shop or an automated workflow, can become the provider of its outputs. That holds even when the application only runs inside the company.

Many companies are both at once. Anyone running their own image pipeline and publishing the results carries both duties and needs the visible labeling just as much as the machine-readable one. mintys handles both in a single step. Who carries which role is covered in detail by Who has to label AI images?.

mintys reaches your work in three ways:

• Web app: label in the browser, no installation.
• API: labeling straight from your product data or editorial system.
• MCP: labeling from agent workflows via the Model Context Protocol.

mintys is needed wherever the AI Act calls for labeling. That puts you on the safe side if a warning letter over missing or incorrect labeling lands, or a fine is on the table.

Which problem does mintys solve in AI labeling?

AI labeling fails in everyday work at four points: at the judgement call, at the design of the label, at the durability of the file, and at the evidence.

As a deployer you have to disclose visibly, and you cannot count your AI tool's machine-readable marking towards that, because the viewer cannot see it. The small “cr” mark that signals Content Credentials under the C2PA standard is not enough either: anyone who wants to read it has to click it. Your duty is only met by a notice the viewer sees straight away.

Four problems come up again and again in practice:

• The judgement call is unclear. Out of caution teams label everything, including the real product photo in front of an AI background, or they miss the photorealistic AI avatar because it is “just a model”. Too much labeling devalues your own genuine shots, too little is a risk of fines. mintys pre-checks every image and suggests what needs a label and what does not.
• The design is open. The AI Act prescribes no format, it only requires the notice to be understandable at first glance. So every team develops its own labeling style, and when the format switches from landscape to portrait the label slides into the crop or over the subject. In mintys you import your own branding, meaning logo, colours and typeface, and set your own labeling standard within it. Which cases get a label, and whether it reads “AI-generated” or “AI-modified”, mintys detects automatically; you decide the placement yourself. The label survives the change of layout, sitting in the same spot in portrait and landscape alike, and it survives the change of file format consistently. Everyone on the team labels the same way.
• The file loses the marking. A screenshot wipes the metadata, many platforms strip it on upload, and a visible label can be cropped off. What you labeled correctly arrives at the audience unlabeled. That is why mintys puts a second layer into the pixels themselves, one that survives cropping, compression, screenshots and re-uploading.
• The evidence is missing. If an authority asks, or a third party reuses your image without the label, all that counts is what you can prove. mintys records every labeling act in a digital passport inside the file and in a register: which image, which label, set when, by whom.

mintys works on exactly those four points, in the order of your work: judge, label, prove, manage.

How does AI labeling with mintys work?

You label AI content with mintys in a single pass of four steps: judge, choose the label, label visibly, label machine-readably.

1. Pre-check: does the content need a label at all? An AI-based pre-check sorts every image: does it show something real, and does it look authentic? Or is it retouching, a background swap or a recognisably unrealistic scene, which under the European Commission's guidelines usually needs no label? mintys supports you in that judgement, the decision stays with you. So you label only what has to be labeled.
2. Deepfake scanner: which label fits? A scanner estimates the AI share of the content and proposes the matching EU label: “AI-generated” for generated content, “AI-modified” for altered content. What gets assessed is always the overall image, not the individual element. So anyone assembling AI-generated components into a collage decides on the finished image as a whole which label it carries.
3. Visible AI label: in your corporate design. mintys places the label directly in the image, as an icon with plain German or English text and, if you want, in your corporate design. You set the placement once. After that the label sits in the same spot in every format, from portrait to landscape and across every file export.
4. Machine-readable labeling: per C2PA inside the file. mintys writes that same labeling into the file as a Content Credential under the C2PA standard, meaning signed provenance information that checking software can read. The visible label itself is documented there too: which label was set, when and by whom. mintys carries existing metadata over, such as a photo's EXIF camera data or a Content Credential your AI tool already wrote. So everything sits in one place and is machine-readable. How Content Credentials are built is shown by Content Credentials: how digital content proves its origin.

The visible label and Content Credentials behave like a nutrition score and an ingredients list: one is readable at a glance, the other sets out in detail what is inside the file, and both sit on the product itself.

An example from an online shop: a fashion retailer prepares three image series for a new collection. The studio shots are real and only colour-corrected, they need no label. In the second series a sneaker stands in an AI-generated setting. Here what counts is not the individual element but the impression of the finished image: as long as the setting stays decoration and the sneaker looks the way it actually looks, the series usually needs no label. It is a different matter as soon as AI alters the product itself and shows, say, a sole that does not exist. Then the image carries “AI-modified”. The third series shows an AI model wearing the collection. That person does not exist, so the series usually counts as a deepfake and gets “AI-generated”. The retailer approves the suggestions, the label appears in the shop's design, and the machine-readable C2PA labeling stays in the files that go into the product data feed.

At the end of the pass you have a clearly labeled deepfake, visible for people and machine-readable for software. Your images circulate online without you having to give their labeling another thought.

How do you prove the labeling?

mintys delivers the evidence for labeling across two independent layers: the digital passport in the Content Credentials and the invisible watermark in the pixels. If one drops out, the other carries on.

The digital passport

The digital passport is the machine-readable labeling inside the file itself. When labeling, mintys records in it that and when the visible label was set, and by whom. Checking software reads those entries automatically. For people they are reachable through the small “cr” mark that many platforms and image programs show on the image: one click, and the passport opens.

What appears then works like a product's ingredients list. It sets out the image's chain of creation, meaning which AI tool was involved, what was changed about it, when that happened and which label the image carries. Every entry is cryptographically signed and therefore verifiable as genuine (C2PA explained).

The invisible watermark

mintys additionally embeds an identifier directly into the pixels. Behind that sits steganography, the craft of building information into a piece of media so that it cannot be seen with the naked eye. Metadata only travels along as long as the file keeps it. The watermark, by contrast, is part of the image and survives cropping, compression, screenshots and re-uploading. If the visible label is removed or a platform strips the metadata, the labeling remains provable through the watermark.

You can check it yourself. on:mint Verify is a free browser extension for Chrome and other browsers. A right-click on any image online shows whether it was AI-generated and whether it was labeled with mintys. To do that, Verify reads not only the Content Credentials in the file information but also scans the steganographic signal in the pixels. That is why it recognises an image labeled with mintys even when a platform has stripped the metadata.

Deepfake management: how do you manage and share labeled content?

Deepfake management in mintys is a dashboard with a register in which you record every labeling and every exception, and from which you share labeled deepfakes directly.

Exceptions. The register also holds the content that stayed unlabeled after the pre-check, together with the reason for the exception. That is the part teams regularly overlook: anyone documenting only the labeled images cannot show later that the unlabeled ones were reviewed. With a documented exception a deliberate decision is on record; without it there is a gap in your inventory.

Sharing on social media. From inside mintys you share the C2PA-labeled content with its visible label straight to your social channels. Every labeling carries on:mint's signature, issued with an officially recognised certificate. Platforms supporting the C2PA standard read the digital passport and show that the content carries AI labeling and who signed it. Your deepfake reaches the audience as verified content rather than a mere claim. When shared through on:mint it additionally carries the note “verified by on:mint”, evidence that the labeling was set and recorded in the register. on:mint has signed the Code of Practice on Transparency of AI-Generated Content and labels along its procedures. For you, proof of compliance becomes a formality. The European Commission's guidelines state explicitly that companies working outside the Code of Practice have to expect more frequent and more detailed requests for information.

Conclusion: AI labeling with mintys means label, prove, manage

The AI labeling duty is complex. mintys makes it simple: you get the right label, the invisible watermark and the evidence to go with it, and you can put deepfakes to work without worry. Anyone publishing AI images professionally has had to disclose, since 2 August 2026, that they were artificially generated or altered. mintys handles that and supplies the evidence with it.

On the cost side: the AI Act provides for fines of up to 15 million euros for breaches of the labeling duty or, for companies, up to 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises the lower of the two applies. On top of that comes a second risk that does not come from authorities: competitors and associations can issue a warning letter over missing labeling as unfair competition and demand a cease-and-desist. The Wettbewerbszentrale, a self-regulatory body of German industry for fair competition, holds that expressly possible and set up its own AI complaints office at the end of July 2026, with “Inadequate labeling of AI content/deepfakes” as one of four complaint types. Companies, associations and private individuals can file; a screenshot and the URL suffice as evidence.

Three approaches compared: what holds up when it counts?

Get started with mintys

mintys covers both sides of labeling: the machine-readable marking through Content Credentials under the C2PA standard, and the visible label including evidence. Content Credentials are more than box-ticking. They are readable quality signals showing the origin and creation of a piece of content, and when shared they carry on:mint's signature.

Start right away: label AI content.

Frequently asked questions