Disclosing that AI made something: how to design the label

An AI label changes what people believe. What a CHI 2025 study found, what Article 50 of the EU AI Act requires and how to label without misleading people.

info Quick Definition

An AI-generated content disclosure is the visible signal that a text, image, audio clip or answer was produced or modified by an AI system. Its job is to let the person know what they are reading so they can evaluate it with that information.

Why disclose?

There are three reasons. The first is the right to know who or what you are interacting with: someone talking to a chatbot should know it. The second is that origin changes how something is read: a text with an AI label is read differently from one without. The third is regulatory, because the European Union sets concrete obligations, summarized below.

But disclosure is not a neutral checkbox. How the notice is presented changes what people believe and how much they trust, which is why design matters so much.

What do studies say about labels?

A study presented at CHI 2025 randomly assigned 911 participants to one of ten label designs for generated, manipulated or deepfake images on social media, or to a control group. The designs varied in tone (a warning with red icons and the word “deepfake”, a neutral label that does not mention AI, a neutral label that says “Made with AI”), in icon and colour, in position and in level of detail.

The results have useful nuances for a designer:

Labels increased the belief that an image was AI-generated, but the effect was small. The text “Made with AI”, placed at the bottom of the image, produced the highest belief.
Trust in the label depended on design. The most trusted version was a detailed, neutral description that did not mention AI, placed on top of the image, similar to content credentials.
Trust in a label was strongly correlated with trust in the platform (r = .73): people who trust the platform trust its label.
Labels did not significantly change likes, comments or shares. The type of content, political or entertainment, mattered more.

This is a single study, on social media images; it cannot tell you how the same label would read in a chat or a document. What it does show is that label design is a variable, so it is worth testing.

What does European regulation require?

Article 50 of the EU AI Act (Regulation 2024/1689) splits the obligations between whoever provides the system and whoever uses it:

Who	What they must do	Paragraph
Provider of a system that interacts with people (chatbots, voice assistants, agents)	Design it so people know they are dealing with AI, unless that is obvious	50(1)
Provider of a system that generates synthetic audio, image, video or text	Mark outputs in a machine-readable format so they can be detected as generated or manipulated, as far as technically feasible	50(2)
Deployer of emotion recognition or biometric categorisation systems	Inform the people exposed to them	50(3)
Whoever publishes a deepfake, or AI-generated text on matters of public interest	Disclose that it is artificial; for clearly artistic or satirical works the disclosure can be limited so it does not spoil the work. Text is exempt if it went through human editorial review and someone holds editorial responsibility	50(4)

Paragraph 50(5) adds that the information must be clear and distinguishable, given at the latest at the first interaction or exposure, and meet accessibility requirements.

The Article 50 obligations apply from 2 August 2026. Regulation 2026/1744 (the digital omnibus on AI) sets 2 December 2026 as the transitional date for the machine-readable marking in paragraph 50(2), and according to the law firm Cooley it only benefits generative systems already on the market before 2 August. The Commission published its guidelines on Article 50 on 20 July 2026, and the final voluntary code of practice on transparency of AI-generated content was published on 10 June 2026. The Commission and the AI Board consider it an adequate tool to demonstrate compliance. According to Cooley, non-compliance can bring fines of up to 15 million euros or 3% of worldwide annual turnover. This is not legal advice: before deciding anything, check the text on EUR-Lex and consult whoever handles compliance.

How do you design a good notice?

Five rules that follow from the research:

At the start of the interaction. In a chat, the notice goes before or alongside the first answer, not in a footer or in the terms and conditions.
Neutral and descriptive. “Generated with AI” informs. An alarmist tone makes people stop reading and treat everything else as doubtful.
With detail that helps. A description of what was generated, by which system and from which sources gives more information than an icon.
Accessible. The notice must be readable by screen readers, have enough contrast and not rely on colour or an icon alone.
With a way to verify. If the content matters, show how to check it: sources, human review, or a statement that the information has not been independently confirmed. See citations and sources.

When a notice is NOT enough

When the content deceives on purpose. A label does not make a deepfake honest. That calls for moderation, not only a notice.
When the notice becomes routine. If every line carries the same label, it stops informing.
When the person cannot act on it. Flagging something that cannot be reviewed or corrected is a formality.
When it is used to avoid responsibility. Writing “may contain errors” does not replace designing an honest experience. See AI dark patterns for the misleading uses.

Example

Illustrative example: an internal communication platform that generates channel summaries. Each summary carries, in grey under the title, “Summary generated with AI from 48 messages”, with a link “View original messages”. If someone edits the summary, the label changes to “Summary generated with AI, edited by Ana”. A screen reader reads the same information when the summary opens. The label states origin and authorship, and leaves verification one click away.

Common mistakes

Hiding the notice in a footer or in the terms.
Using alarmist words that make everything look fake.
Putting the label only in a colour or icon that screen readers cannot convey.
Believing a label replaces human review for content that matters.
Forgetting that changing the notice changes trust, so it has to be tested with real people. Compare with trust calibration.

Sources

Gamage et al., “Labeling Synthetic Content: User Perceptions of Warning Label Designs for AI-generated Content on Social Media” (CHI 2025), arXiv 2503.05711: https://arxiv.org/html/2503.05711v1
Cooley, “EU AI Act: Transparency Obligations Take Effect 2 August 2026” (law-firm analysis, not the legal text): https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026
European Commission, AI Act Service Desk, Article 50 (consolidated text): https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
European Commission, “Guidelines on transparency obligations for providers and deployers of AI systems”: https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems
European Commission, code of practice on AI-generated content: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
EUR-Lex, Regulation (EU) 2026/1744: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng

Frequently asked questions

Why disclose that AI generated a piece of content?

Because people have the right to know what they are interacting with or reading, and because information about origin influences what they believe and how much they trust it.

Does an AI label always reduce trust?

Not necessarily. The effect depends on design. A study with 911 participants found that labels increased the belief that an image was AI-generated, and that trust in the label varied with how it was designed.

What does the EU require about AI disclosure?

Article 50 of the AI Act requires telling people when they interact with an AI system, marking synthetic content in a machine-readable way and disclosing deepfakes. It applies from 2 August 2026; marking for generative systems already on the market has until 2 December 2026.

Who has to disclose, the provider or whoever publishes the content?

Both, each for their part. The provider of the system makes sure people know they are dealing with AI and marks outputs in a machine-readable way. Whoever publishes a deepfake, or AI-generated text on matters of public interest without human editorial review, must disclose that it is artificial.

How should an AI-generated content notice be worded?

Clearly, neutrally and at the start of the interaction, not hidden in a footer, and with a way to verify the content when it matters.