AI Interaction Disclosure
A compliance layer for any AI system that interacts directly with people (chatbots, voice agents, conversational assistants), ensuring users are clearly informed they are engaging with an AI rather than a human. It defines when the obligation applies, what the disclosure must say, how it is surfaced across channels, and when an exception is legitimate.
The Problem
Users lose trust, and regulators take notice, when a product creates the impression of human interaction where none exists. The harm is rarely obvious in the moment. It surfaces later, as a decision made on the assumption of human judgment or a dispute where the nature of the interlocutor turns out to be material. What the user really loses is informed consent: the ability to decide how much to share, how much weight to give an answer, and whether to ask for a human.
- A user submits sensitive information (health, financial, a complaint) believing they are speaking to a person.
- The product mimics human tone, names, or style without disclosure, eroding the user's ability to calibrate expectations.
- Disclosure arrives too late, after the interaction has produced consequences the user would have handled differently.
- Disclosure is technically present but practically invisible: buried in terms, shown in illegible microcopy, or hidden in a settings menu.
- Some channels disclose and others do not, creating a patchwork that undermines confidence in the whole product.
Intended Outcome
AI Interaction Disclosure is working when users never have to wonder who they are talking to.
- Known from the start: users know they are engaging with an AI before or at the very first interaction.
- Proportionate, not uniform: routine exchanges carry a first-turn statement and a persistent badge; a heavier acknowledgment fires only at a sensitive-disclosure moment.
- Legible handoffs: when a human joins or takes over, the current responder stays identifiable.
- Perceivable by everyone: disclosure reaches users relying on assistive technology, not only sighted users.
- Consistent across channels: web, mobile, voice, and API-mediated interactions follow the same standard.
- Demonstrable: disclosure is systematic, logged, and version-controlled rather than ad hoc.
Scope & Boundaries
- The legal baseline: Article 50(1) disclosure for AI that interacts directly with people; Article 50(5) delivery standard (clear, distinguishable, accessible, no later than first interaction).
- A session-start mechanism that surfaces disclosure at conversation start on every channel, not retroactively.
- An evidence trail logging timestamp, channel, session ID, copy version, whether the disclosure rendered, and how the user responded, with rendering and acknowledgment as separate fields.
- Copy version control so disclosure wording changes are tracked and auditable.
- Two separate exception registers: the 'obviously AI' carve-out and the law-enforcement exemption.
- A one-time checkbox: disclosure must be present at every new interaction context, not only at account creation.
- Single-language coverage: copy in one language does not satisfy the obligation in others where the product ships.
- Personas as a substitute: a human name, avatar, or tone does not discharge the obligation, and may increase the need for it.
- Relay surfaces you do not control: third-party clients rendering your responses are handled via a machine-readable AI-origin flag plus API terms, with residual risk reported separately.
How it Works
The product introduces a disclosure layer that intercepts the user's first contact with any AI-powered conversational surface and states the system's nature before any substantive exchange takes place.
Identify: map every surface where the AI interacts directly with a person (chat, voice/telephony, autonomous messaging, AI-assisted fields, and API relay surfaces).
Define: for each controlled surface, decide the trigger (session start, first message, channel switch, sensitive moment), the format, the copy, the persistence, and how it is exposed to assistive technology.
Log: record timestamp, channel, session_id, copy_version, disclosure_rendered, and acknowledgment_type, keeping passive exposure and active engagement distinguishable.
Document exceptions: record the basis, owner, and review date for the 'obviously AI' carve-out and the law-enforcement exemption separately.
System Interaction
Placement & Triggers
- At the point a user is about to disclose something sensitive (health, financial, identity, complaint), whatever the product's nominal category.
- Onboarding, where users build a mental model of what they are dealing with.
- Escalation paths, where users believe they have reached a human reviewer.
- Handoff points, where responsibility moves between AI and human inside a single thread, in either direction.
Supported Architectures
Experience Surface
Disclosure should feel like part of the interface, not an interruption. The goal is informed users, not friction. The opening statement sets expectations, a persistent responder badge sustains them across handoffs, and a sensitive-moment step concentrates friction only where the stakes justify it. In voice channels, a spoken announcement carries the same disclosure with no visual dependency.
Design Assets
A UI Kit that makes AI identity legible across every channel, from the opening statement through handoff badges to the voice script. All components meet WCAG 2.1 AA and never rely on colour or position alone. Two of these are shared primitives, maintained once and reused by the Body and Emotion Scanning Notice pattern.
AI Identity Message
State the system's nature before any substantive exchange.
Placement · First turn in the conversation stream, above the input. Not a dismissible banner, so it stays referenceable in the thread.
- Identity line: “I'm [Name], an AI.”
- Scope line: what it can help with.
- Escalation line: how to reach a person.
AI Responder Badge
Keep the current responder identifiable for the whole session, including after handoff.
Placement · On the message bubble or in the conversation header. Persistent and non-dismissible.
- Variant `ai`: 'AI-powered' or 'Automated assistant'.
- Variant `human`: shown once a person takes over, so the switch is visible within one message.
- Built on the shared AI Status Indicator primitive, also used by the Body and Emotion Scanning Notice pattern for its Persistent Scanning Indicator.
Accessibility · Announced by assistive tech as part of message identity, not marked decorative.
Sensitive-Moment Acknowledgment
Add friction only where the stakes justify it, immediately before a user shares something sensitive.
Placement · Triggered by a detected sensitive-disclosure moment in the conversation, not the product's domain.
- Restated AI identity.
- One line on what happens to what they share.
- Confirm action, which emits acknowledgment_type.
- Alternative route to a human.
Escalation Affordance
Give users a route to a person without ambiguity about who is currently responding.
Placement · Persistently available, and restated inside the sensitive-moment acknowledgment.
- Current responder state.
- Action to request a human.
- Expectation of what happens next.
Voice Disclosure (shared)
Carry the same disclosure into channels with no visual surface.
Placement · Call start, before any data collection begins.
- Identity statement.
- Escalation instruction.
- Self-contained delivery with no dependency on a paired visual cue.
- Shared component: the Body and Emotion Scanning Notice pattern uses the same Voice Disclosure for its Article 50(3) voice channel. Only the announced subject differs.
When to use
Risks & Guidelines
- Consent theater: a banner dismissed once and never seen again, especially on long threads where a handoff happens later.
- Over-disclosure friction: heavy disclosure at every turn trains users to dismiss it, often caused by domain-gated modals firing on routine exchanges.
- Exception creep: quietly widening the 'obviously AI' carve-out without documented review.
- Persona confusion: giving the system a human name, then arguing the name alone constitutes notice.
- Channel gaps: disclosing on web while mobile, API, or white-label deployments go uncovered.
- Visual-only disclosure: a badge assistive tech skips, or identity conveyed by colour alone, failing the Article 50(5) accessibility limb.
- Test with real users that they understand they are interacting with AI before the session begins, and track the result.
- Test with a screen reader, not only sighted review, before treating any surface as covered.
- Apply the legal threshold strictly: 'reasonably well-informed, observant, and circumspect' is a legal standard, not a UX preference.
- Assign a named owner for the disclosure standard across all channels.
- Version every change to when, where, or how disclosure is shown in the compliance change log.
How to Measure Success
How will we know this pattern is strengthening trust?
North Star
Disclosure Coverage Rate: the percentage of AI-interaction sessions where a qualifying disclosure was surfaced at or before the first AI-generated turn (clear, in the interface language, timely, and perceivable by assistive technology). Target: 100%, with relay surfaces reported separately.
Disclosure Comprehension Rate
whether sampled users actually registered the disclosure, the measure of whether it worked, not just whether it was delivered.
Acknowledgment vs Exposure
active acknowledgment set against passive exposure. Rising exposure with falling acknowledgment means the design is going invisible.
Accessibility Conformance
the share of surfaces passing assistive-technology exposure, contrast, and non-visual equivalence checks.
Exception Health
whether every claimed exception is documented, categorised, and reviewed.