PTN—25
AI self-identification
Make it unambiguous when a person is interacting with AI or reading AI-generated content.
People calibrate everything — trust, tone, patience, verification effort — based on who they think they are talking to. When a system lets someone believe AI output is human work, every downstream judgment is miscalibrated, and the eventual discovery reads as deception even when nobody intended it. Self-identification is the cheapest trust pattern to build and the most expensive to skip: one label versus a betrayal narrative.
Identification has two distinct surfaces. Conversational: the agent names itself as AI at the start of the relationship and never roleplays as human, even when asked. Provenance: AI-generated artifacts — summaries, drafts, images, meeting notes — carry a durable marker of origin, because content outlives the session that produced it. A summary pasted into a doc six weeks later should still be traceable to “AI wrote this,” which is why platform work like C2PA content credentials matters beyond any single product.
Do this with confidence, not apology. A quiet, persistent “AI-generated” label is honest infrastructure; a paragraph of disclaimers before every answer is throat-clearing that users learn to scroll past. And increasingly it is not optional — the EU AI Act requires disclosure when people interact with AI systems, and Microsoft’s HAX guidelines have listed “make clear what the system can do” as guideline number one since 2019. The design question is only how gracefully you comply.
When to use
- Any conversational interface where a person could plausibly assume a human is responding.
- AI-generated content that travels — emails, documents, images, meeting summaries.
- AI operating inside human channels: support queues, Slack, comment threads, code review.
- Voice interfaces, where the human-likeness of the medium raises the bar for disclosure.
Design considerations
- 01Identify at first contact, not in a settings page or terms of service.
- 02Never let the agent claim to be human, even under direct instruction or roleplay pressure.
- 03Attach provenance to artifacts, not just sessions — the label must survive copy, export, and forwarding.
- 04Design the hand-off seam: when a human takes over from a bot (or vice versa), announce the switch explicitly.
- 05Keep the marker calm and persistent — a small durable label beats a large dismissible banner.
- 06Label AI-suggested content that a human approved differently from raw AI output; provenance has grades.
Pitfalls
- ✕The buried disclosure: “AI” in fine print or behind an avatar tap. When people discover it later, the story becomes the deception, not the feature.
- ✕Human-mimicry theater — fake typing indicators, fabricated personal anecdotes, a bot with a human name and stock-photo face. Every mimicked signal is a small lie the product tells.
- ✕Seamless human hand-offs where the person never learns the first half of the conversation was a bot, poisoning trust in the human half too.
- ✕Provenance that dies in transit: content labeled in-app but exported clean, so the label exists exactly where it matters least.
In the wild
- Claude
- Consistently identifies as an AI made by Anthropic and declines to claim humanity even when a roleplay frame invites it.
- Intercom Fin
- Identifies as an AI agent in support conversations and makes the escalation to a human teammate an explicit, visible hand-off.
- Granola
- Meeting notes are framed as AI-enhanced from your own raw notes, keeping the machine’s contribution legible in the artifact itself.
- Adobe Firefly
- Attaches Content Credentials (C2PA) to generated images so provenance travels with the file, not just the app.
- GitHub Copilot
- Renders suggestions as visually distinct ghost text, so machine-authored code is never confusable with what you typed.