# Labeling AI-generated content: what the AI Act requires and how to embed it in your workflow

Author: Marin Niehues, Niehues Consulting · Published: 2026-06-18 · Updated: 2026-11-16 · AI Law, Risk & Governance
Source: https://www.niehuesconsulting.com/blog/labelling-ai-generated-content/

The AI Act requires a visible notice for deepfakes, for AI-generated text on matters of public interest without editorial review, and for chatbots. Providers of generative systems also mark their outputs in a machine-readable format. For companies, compliance is decided in the content approval process. This article provides an overview and does not constitute legal advice.

The transparency obligations sit in Article 50 of the AI Act and apply regardless of a system's risk class, so they affect almost every company that uses AI in its communications. Official statistics show the scale: in 2025, around 26 percent of German companies with ten or more employees used AI, rising to 57 percent among those with 250 or more. The wider regulation is covered in our article on the [EU AI Act](/blog/eu-ai-act/).

**Labeling is a process question. A company that asks one fixed question about AI origin at every content approval step meets the obligation with little effort and can evidence it to the regulator.**

## Which AI content must be labeled under the AI Act?

Four categories must be labeled: chatbots, synthetic outputs at a technical level, deepfakes and published AI-generated text on matters of public interest. Each category is addressed to a specific role and requires its own form of notice.

**The four labeling obligations under Article 50**
Chatbots and assistants: systems that interact directly with people must let the person know they are dealing with AI. The obligation lapses where this is obvious to a reasonably well-informed person.

Machine-readable marking: providers of systems that generate synthetic content mark the outputs in a machine-readable format as artificially generated. Assistive functions for standard editing and systems that do not substantially alter the input are exempt.

Deepfakes: deployers who generate or manipulate image, audio or video content that resembles real people or events and would falsely appear authentic must disclose that the content is artificial. For evidently artistic or satirical works, a notice that does not hamper the display of the work is sufficient.

Text on matters of public interest: deployers who publish AI-generated text to inform the public on matters of public interest must disclose this. The obligation lapses where a human has reviewed the text and a person holds editorial responsibility for it.

The same form requirement applies to all four. The notice must be clear and distinguishable, appear at the latest at the first interaction or exposure, and meet accessibility requirements. A line in the legal notice of a website does not meet this standard, because the viewer does not see it next to the content.

## Provider or deployer: the role determines the duty

A company that uses third-party AI tools for its own content is a deployer, and deployers carry the duties on deepfakes and text. A provider is anyone who develops an AI system, or has one developed, and places it on the market or puts it into service under its own name. Providers carry the chatbot notice and the machine-readable marking.

In practice the line falls in two places. The first is the company's own customer chat: a business running a language-model assistant on its website under its own name checks whether this makes it a provider. The safe design is identical either way, because a sentence such as "You are chatting with an AI assistant" at the start of the conversation costs nothing. The second is proprietary generators: a company that offers customers a self-built application for creating images or text takes on a provider's marking duties.

Purely personal, non-professional use falls outside the definition of a deployer. As soon as employees use AI on behalf of the company, the company acts as deployer. This allocation belongs in the [AI policy for employees](/blog/ai-policy-for-employees/), so that everyone knows when their content needs a notice.

## When does the exemption for editorially reviewed text apply?

The exemption applies when a human has reviewed the AI-generated text for content and a natural or legal person holds editorial responsibility for its publication. Both conditions must be met together. The European Commission's code of practice on marking and labeling AI-generated content makes this concrete: the deployer names the responsible person and documents the organizational measures that ensure adequate review. The code does not require a record of each individual review.

This gives companies a clear dividing line. Press releases and expert articles that a named person approves before publication fall under the exemption. Automatically generated market reports published without human approval need the notice. Advertising copy for the company's own products generally does not concern a matter of public interest and falls outside this category. There is no comparable exemption for images and video: a deepfake needs the notice even after editorial review.

## How does machine-readable marking work?

Machine-readable marking means that an AI output contains a technical identifier that software can read even when a person cannot see it. The recitals of the regulation name watermarks, metadata and cryptographic methods among the techniques. Solutions should be effective and interoperable, as well as robust and reliable, as far as the state of the art allows.

The Commission's code of practice relies on several layers. Where no single technique meets the requirements, providers combine at least two layers, typically signed metadata and a watermark embedded in the content itself. Metadata alone rarely suffices, because a screenshot or recompression strips it out. Free-form text cannot carry metadata, so the code relies on watermarks within the text. Providers also make tools available to verify the marking.

An established open standard for provenance data is C2PA, from the Coalition for Content Provenance and Authenticity. It writes signed records into the file showing which tool created the content and how it was edited. For deployers, this means export routes must not remove that data. Many editing programs strip metadata on export by default, and that is where the provider's marking is lost.

## What labeling does to readers

A notice of AI origin lowers trust in content, even when the content is accurate. Two preregistered experiments with 4,976 participants from the United States and the United Kingdom show that headlines labeled "AI-generated" are rated as less accurate and shared less often, even when they are true or written by humans. The cause was readers' assumption that the content had been produced entirely without human oversight.

A second experiment with around 1,600 participants in the United States tested persuasive texts on policy issues. The texts shifted support by almost ten percentage points on average, and a notice of AI authorship did not measurably weaken this effect.

**What studies show about the effect of AI notices**
4,976 | Participants in two experiments on the effect of an "AI-generated" label on headlines
1/3 | The effect of the AI label on perceived accuracy, relative to a "false" label, even for true content
9.7 percentage points | Average shift in support produced by persuasive texts, with or without an AI notice

Two conclusions follow for companies. First, a blanket notice on all content costs credibility without being required by law. The notice belongs where the regulation demands it and where the viewer would otherwise be misled. Second, an explanatory notice works better than a bare tag. A wording such as "Created with AI, reviewed by our editorial team" corrects the assumption of full automation that drives the loss of trust. How marketing teams use AI beyond labeling is covered in our article on [AI in marketing](/blog/ai-in-marketing/).

## Embedding labeling in the workflow

Labeling works reliably when it is a fixed step in the content approval process and not a judgment left to each author. Five building blocks are sufficient:

- **Channel inventory:** every point where AI content leaves the company, from the website to client presentations.
- **Decision rule per content type:** the checklist below, mandatory before every approval.
- **Notice templates per medium:** a visible notice for visual content, a spoken disclaimer for audio-only content, an opening sentence in chat. Optional EU icons are available, and equivalent own labels are permitted.
- **Documentation:** the named person with editorial responsibility and a description of the review process, plus the placement of the notice for deepfakes.
- **Enablement:** a short session on the four categories for everyone who creates content. Article 4 of the regulation already requires AI literacy among staff.

**Checklist before every publication**
- Was the content wholly or substantially generated or altered with AI?
- Does an image or video show real people or events in a way that would falsely appear authentic? Then a notice is mandatory.
- Does a text inform the public on a matter of public interest? Then a notice is mandatory, unless a named person has reviewed it and takes responsibility for it.
- Does a system speak directly to customers? Then the AI notice appears at the start of the interaction.
- Is the tool's provenance data preserved on export?
- Is the notice visible next to the content and accessible, at the latest at first contact?

The choice of approach determines effort and effect. The comparison below sets out the usual routes.

**Routes to labeling in the company**
Criterion | Left to each author | Blanket notice on all content | With Niehues Consulting
Legal certainty | Gaps on deepfakes and chat | Mandatory cases covered | Mandatory cases covered and documented
Credibility of content | Unchanged until an error surfaces | Falls measurably, even for reviewed text | Notice only where mandatory, with explanatory wording
Evidence for the regulator | Barely possible | Possible, without proof of review | Responsible person, review process and placement evidenced
Day-to-day effort | Low, high risk | Low, negative effect on readers | One check question per approval

## Niehues Consulting as your partner for labeling

[Niehues Consulting](/) embeds labeling where content is created and approved. Marin Niehues is a management consultant with a delivery perspective. He builds organizations that deliver and brings AI into operations, in engagements ranging from VC-funded scale-ups to DAX 40 corporations. He has worked in 20+ countries on six continents, with the DACH region as his home base.

The approach follows the Niehues Delivery Model with its phases Understand, Build and Embed. Three of its check questions matter most for labeling. In Understand: who decides, who delivers, who uses? In Build: which roles, decision paths and interfaces does operation require? In Embed: who owns operation from day one? The answers produce the five building blocks, including named editorial responsibility. Labeling thus fits into the wider governance described in the article [AI management: what executive leadership must steer](/blog/ai-management-what-executives-must-steer/). The starting point is a free 30-minute initial consultation via [our contact page](/#kontakt).

**Three decisions for the executive team**
The executive team decides who holds editorial responsibility for AI-assisted publications and has the review process documented. It embeds the question of AI origin in the approval process of every external channel. It opts for notices in the mandatory cases, with wording that makes human review visible.

**Sources**
- Regulation (EU) 2024/1689 laying down harmonized rules on artificial intelligence (AI Act), Articles 50 and 99, 2024
- European Commission: Code of Practice on marking and labeling of AI-generated content, 2026
- Federal Statistical Office of Germany (Destatis): Use of information and communication technology in enterprises, 2025
- Altay, Gilardi: People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation, PNAS Nexus, 2024
- Gallegos, Shani, Shi, Bianchi, Gainsburg, Jurafsky, Willer: Labeling Messages as AI-Generated Does Not Reduce Their Persuasive Effects, arXiv, 2025

## Frequently asked questions

### Do I have to label AI-generated images?

AI-generated images must be labeled when they show real people, places or events in a way that would falsely appear authentic. A clearly artificial illustration does not fall under the deepfake obligation. The deciding factor is whether the viewer would take the image for a genuine photograph.

### Do I need to disclose that I used ChatGPT to write content?

Disclosure is required only for published text that informs the public on matters of public interest and has not been reviewed by a human. If a named person reviews the text and takes editorial responsibility, the notice is not required. Internal drafts and product copy fall outside this obligation.

### What counts as a deepfake under the AI Act?

Under the regulation's definition, a deepfake is AI-generated or manipulated image, audio or video content that resembles existing people, objects or events and would falsely appear authentic to a person. A synthetic video of the CEO counts, as does a cloned voice in an advertisement. The notice is required even if the person shown has consented.

### What are the penalties for failing to label AI content?

Breaches of the transparency obligations can be fined up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, the lower of the two amounts applies. The amount in each case depends on the gravity and duration of the infringement.

### Does a chatbot have to disclose that it is AI?

A chatbot must let users know they are interacting with AI, unless this is obvious from the circumstances. The notice belongs at the start of the conversation. A sentence in the chat window's greeting meets this requirement.

### Is the EU code of practice on AI labeling mandatory?

The code of practice is voluntary, but it is the recognized tool for demonstrating compliance with Article 50. Signatories can rely on its measures. Companies that do not sign must show the market surveillance authority themselves that their own measures are equivalent.

### Does the labeling obligation apply to internal content?

The deepfake obligation applies regardless of audience, so it also covers a synthetic video in an internal training session. The text obligation concerns only publications intended to inform the public. An internal disclosure rule for all other content is nonetheless good practice.

### How should I label AI content on social media?

On social media, place the notice directly in the image or video and also in the post text, so that it is visible without a click. Major platforms offer their own AI labels, which you activate in mandatory cases. The tool's provenance data is preserved if you upload the file without re-exporting it from an editing program.
