AI Trust Patterns

Trust and Transparency

AI Disclosure

Make it clear that users are interacting with AI, and what it can and can't do.

Phase
Design · Build
Applies to
End users
Oversight
Human in command
Key factor
Rights & data sensitivity
OverviewWhen to useQuestions to askExampleHow other systems do itAnti-patternAnti-patterns to pay attention toReferencesRelated

Overview

In the agentic experience, people calibrate trust differently for a machine than for a person, and they can only do that if they know which one they're dealing with. AI disclosure matters at the start, in plain words or an icon, and sets expectations for what the AI can do and how to reach a person. Of all the obligations in the EU AI Act, this is the one most teams must meet first.

When to use

Use when

  • Any AI system that talks with people directly: chat, voice, email replies, comments.

Questions to ask

Ask these at kickoff, when design work starts. Tick them off as your team answers.

Example

A clinic's booking assistant introduces itself.

Clinic assistantDo
AIBooking assistantAITalk to a person
Hi, I'm an AI assistant. I can help you book, move or cancel appointments. I can't give medical advice. A person from our team can take over at any time.

How other systems do it

Platforms take different approaches to labelling AI content, from creator declarations to machine-readable marking.

YouTube

Creators must declare when realistic content has been meaningfully altered or synthetically generated. The label appears in the expanded description; for sensitive topics such as health, news, elections and finance, it appears more prominently on the video player itself. Creators who don't disclose can face penalties, and YouTube may add the label itself. This is a mandatory declaration model, not an optional badge.

Meta (Facebook, Instagram, Threads)

Content Meta detects as AI-generated carries an “AI info” label people can open for more detail. For content that was only edited with AI tools, Meta moved the label from under the account name into the post's menu. That's a useful caution: a label people have to look for does less work than one they see.

Synthesia

AI-generated videos carry C2PA provenance signals by default, so machines can detect them as AI-generated. Customers can also switch on a visible AI label for every scene, to help meet their own disclosure obligations. It's a two-layer approach: machine-readable marking from the provider, visible disclosure chosen by the publisher.

The strongest examples put the label where people see or use the output, not only in settings or legal notices. The form varies (an inline label, a player overlay, a creator declaration, embedded credentials), but the principle holds: the person seeing or using the output should be able to tell it was made by AI.

Anti-pattern: Hidden AI

A human name and photo, no label. People assume they're talking to a member of staff.

Clinic assistantDon't
EEmma from Reception
Hi! I'm Emma 😊 How can I help you today?

Anti-patterns to pay attention to

  • No AI label where the assistant appears.
  • A human name or photo with no disclosure.
  • No way to reach a person.

References

Current guidelines · EU AI Act, ISO/IEC 420012 references

For orientation only, not legal advice. This obligation applies beyond high-risk systems. EU AI Act links open a readable version of the official text published on EUR-Lex; ISO/IEC 42001 is a paid standard.

  • EU AI Act Art. 50(1) AI systems that interact directly with people must be designed so those people are informed they're interacting with AI, unless that's obvious from the context.
  • ISO/IEC 42001 A.8.2 Give users plain-language information about the system's capabilities and limits.