In brief
Whenever a person chats with a service, organisation, or account, the interface should clearly state whether the current participant is a human being, an automated system, or a human using AI assistance.
Disclosure should appear before the conversation begins and whenever control changes. It should not be hidden inside terms of service or revealed only when the user asks the correct question.
Context
The same conversational interface can now represent customer-support staff, scripted bots, generative systems, supervised assistants, and combinations of all four. A fluent answer can encourage a person to infer human judgement, attention, or accountability that may not exist.
The proposal is not that every use of software must be announced. It concerns systems that participate as a speaker in a conversation or materially generate the messages attributed to one.
The blueprint
Require persistent, plain-language identity labels for conversational participants:
- Human when messages are composed and sent by an identifiable person.
- AI system when the conversational response is generated and delivered automatically.
- Human with AI assistance when AI materially drafts or transforms the response before a human sends it.
The label should update if a conversation is transferred between an automated system and a human agent.
Implementation
The obligation could apply to customer service, sales, public services, recruitment, healthcare administration, education, and other consequential interactions, with proportionate rules for low-risk uses. Platforms should provide standard technical fields so the disclosure is accessible to assistive technologies and cannot be removed by visual styling.
Organisations should retain an auditable record of whether a human or automated system handled a consequential exchange without exposing unnecessary personal data to the user.
Trade-offs
The boundary between simple automation, AI assistance, and material generation will be contested. Labels that are too detailed may confuse users; labels that are too broad may become meaningless. The framework must also avoid exposing an employee’s internal tools or personal identity when role-level accountability is sufficient.
Open questions
- What level of AI assistance changes the identity label?
- When should a named human be required instead of an accountable organisation?
- How should multi-agent conversations display changes in control?
- Which conversations are consequential enough to require audit records?