Beyond the Chat BoxEssay 08 / 10
Social agents and representation
ESSAY 08

When Your Agent Meets My Agent, Who Is Actually Speaking?

Agent-to-agent interaction should be a consent-governed introduction, not a substitute for human identity.

AgentDate began with a provocative premise: let agents represent people, interact with each other, and offer humans an explainable suggestion about whether a connection might be worth exploring. This shifts the first interaction from person-to-person messaging to agent-mediated representation. It is not just a new matching algorithm. It changes who is speaking, what counts as consent, and how a person can correct the story told about them.

Representation is not replication

A profile-derived agent is not a person. It is a model built from limited signals, assumptions, and design choices. Even public signals may not reflect how someone wants to be represented in a particular context. An interface should not say “this is what the person thinks” when it means “this is what the system inferred from selected data.”

A safer interaction sequence

A person explicitly opts in and chooses the sources and context the agent may use. The system shows a draft representation, evidence, and uncertain inferences. The person edits or excludes traits and topics. The agent explores compatibility within a narrow, disclosed purpose. It must not make commitments, disclose sensitive information, or impersonate the person. The person reviews the summary before any human-to-human contact is proposed.

Consent is ongoing

Consent to create an agent is not consent to every future use. Someone may approve compatibility exploration but not public messages, persistent profiling, or transcript sharing. The product needs purpose limitation, deletion, revocation, and visibility into what was shared. It should identify itself as an agent and never fabricate autobiographical details.

Where this can fail

A user may mistake the agent's inference for a person's statement. The system may amplify stereotypes, expose unexpected profile details, or rank people using proxies they never agreed to share. The product must guard against non-consensual agent creation and allow people to report and correct misrepresentation.

Questions for research

  1. Do people distinguish an agent's inference from a person's stated preference?
  2. What autonomy feels acceptable in an agent-mediated introduction?
  3. How should agents disclose uncertainty and provenance?
  4. Does reviewing an agent conversation improve decisions or create misplaced confidence?

Design hypothesis

The design problem is how humans can remain accurately represented, informed, and in control when agents speak on their behalf.

Research starting points

  1. Amershi et al. (2019), CHI. https://doi.org/10.1145/3290605.3300233
  2. Shneiderman (2020), Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy, International Journal of Human–Computer Interaction.
  3. NIST AI Risk Management Framework. https://www.nist.gov/itl/ai-risk-management-framework
← Back to all ten essays