Guide · Safety
Disclose When Users Are Talking to AI
Hiding that a reply is AI-generated is a safety and trust failure—not a UX flourish. Production products need clear disclosure appropriate to channel and risk.
Undisclosed AI interactions fail Safety & Responsible AI—label the bot and offer a human path. Primary control: Safety & Responsible AI
"Is this a person?" is a safety question
APRF Safety & Responsible AI expects users to know when they are interacting with AI—especially for support, advice, hiring, lending, health-adjacent, or any flow where people may over-trust a confident answer. Undisclosed bots create brand, regulatory, and harm risk even when the model is "secure."
A support widget once used a human-sounding name with no AI label. Escalations spiked when customers discovered the pattern later. The fix was not a smarter model: a persistent "AI assistant" badge, first-message disclosure, and a one-tap path to a human.
What good disclosure looks like
1. Before or at first reply: clear language that the counterpart is AI (not buried in Terms).
2. Persistent affordance: badge or footer that survives multi-turn chat.
3. Escalation path: how to reach a human when stakes rise—wired to real HITL, not a dead link.
4. Generated-content labels for emails, tickets, or docs the system drafts for someone else to send.
5. Channel fit: voice/IVR needs spoken disclosure; APIs need machine-readable `ai_generated` flags for downstream UIs.
Disclosure is not a substitute for refusals, fairness testing, or tool gates. It is the transparency layer those controls sit under.
Evidence, not marketing copy
Keep screenshots or recordings of disclosure UX per surface, the policy that mandates them, and a release check that new channels cannot ship without the label. Pair with Explainability when users dispute an outcome—they should know that it was AI and how to reconstruct why.
Next: Safety & Responsible AI
Open the related pillar specification for mandatory checks, artifacts, and pass conditions. Self-attest is optional.
Related
Frequently asked questions
- Is a line in the Terms of Service enough?
- Usually no. APRF expects user-facing disclosure at the point of interaction, not only legal fine print.
- Do internal tools need disclosure?
- Yes when employees might mistake AI output for human or authoritative policy. Label and cite; keep a human escalation path for high-impact actions.
- How does this relate to decision traces?
- Disclosure tells users they are talking to AI. Traces and citations explain why a particular answer happened. You need both for contested outcomes.