Perspective-Taking Interview¶
Inquiry method — instantiates Other-Agent State Model Calibration
Replaces inference with direct, open-ended questioning to learn the other agent's actual understanding, constraints, and priorities.
Most ways of modeling another agent guess at the hidden state — from behavior, from analogy, from imagination. Perspective-Taking Interview does the opposite: it asks. Instead of simulating the other agent's mind (as role reversal does) or inferring it from traces (as an empathy map or prediction log does), the interview treats the agent as the primary source on their own understanding and simply elicits it. Its defining move is to structure the asking so that the interviewer's own expertise does not leak into the questions — the point is to hear what the agent actually knows, wants, and is constrained by, in the agent's own terms, rather than to confirm what the interviewer already believes.
Example¶
A math tutor is convinced a student "just isn't trying" on fraction problems. Rather than re-explaining — which routes the whole session through the tutor's own model — the tutor runs a perspective-taking interview: "Walk me through what you did on this one." "What were you thinking when you picked that step?" "Which part feels different from last unit?" The answers overturn the model. The student isn't disengaged at all; they are confidently applying a wrong-but-coherent rule (add the denominators) picked up as a half-remembered shortcut. The read flips from motivation problem to a specific mis-generalization — and the next explanation targets that misconception directly instead of exhorting the student to try harder. The direct account did in five minutes what weeks of inference had gotten backwards.
How it works¶
- Uses open, non-leading questions that don't smuggle the answer in ("what did you expect this button to do?" not "did you find the button confusing?").
- Asks for the agent's reasoning and reconstruction, not just their conclusion, so the model underneath the behavior surfaces.
- Actively brackets the interviewer's own knowledge — the discipline is to keep expert context out of the questions so the answers aren't shaped by it.
- Follows the agent's framing and language, including how they weight priorities and constraints, rather than translating everything into the actor's terms.
- Treats every answer as evidence to record — ambiguity included — not as proof.
Tuning parameters¶
- Openness of questions — fully exploratory versus targeted probes. Open surfaces surprises but wanders; targeted is efficient but risks leading the witness.
- Directness versus safety — how bluntly to ask about sensitive states (fear, disagreement, incentives). Direct is faster but can trigger guarded, socially-desirable answers.
- Interviewer disclosure — how much the actor reveals their own view first. Disclosing early anchors the agent to it; withholding keeps answers uncontaminated but can feel like interrogation.
- Cultural register — how far to adapt phrasing, indirectness, and turn-taking to the agent's context, so that deference or silence isn't misread as agreement or absence.
- Depth versus reach — one deep interview versus many shallow ones; depth gets a single model right, breadth samples a population.
When it helps, and when it misleads¶
Its strength is that it is the most direct antidote to the curse of knowledge — the near-inability of someone who knows a thing to model someone who doesn't[1] — because it replaces the expert's projection with the learner's own account. It surfaces states that no behavioral trace can disambiguate, and it does so respectfully, by asking rather than profiling.
Its failure mode is that what people say is not always what they know or want: social-desirability pressure, deference, and leading questions can manufacture exactly the answer the interviewer expected, so the method can quietly become projection wearing an interview's clothes. It is also costly, doesn't scale, and won't extract states an agent is motivated to hide (a negotiating position, a concealed incentive). The discipline is to keep the questions genuinely open, treat answers as evidence to triangulate against observed behavior rather than as ground truth, and record the residual uncertainty instead of closing the model on one confident reply.
How it implements the components¶
Perspective-Taking Interview realizes the evidence-gathering side of the archetype — it produces the raw material other mechanisms model and act on:
evidence_signal_map— the questions and answers are direct evidence, and the interview is the highest-quality source in the map because it comes straight from the agent.actor_state_separation— the discipline of bracketing the interviewer's own knowledge and asking instead is precisely what keeps the actor's expertise out of the model.cultural_context_layer— questions posed in the agent's own framing surface the cultural priorities and constraints that shape their state.
It does not update the model over time or rewrite interaction rules — that loop belongs to Interaction After-Action Review and Prediction and Surprise Log — and it does not govern whether the inquiry is even permissible; that limit is Consent and Privacy Boundary Checklist's.
Related¶
- Instantiates: Other-Agent State Model Calibration — the interview is the primary way the appraisal buys evidence straight from the source.
- Sibling mechanisms: Belief-Desire-Knowledge Map · False-Belief Check · Interaction After-Action Review · Role-Reversal Simulation · Active Listening Loop · Empathy Map with Evidence Marks · Counterparty Model Red Team · Prediction and Surprise Log · Consent and Privacy Boundary Checklist · Stakeholder Hidden-Constraint Board
Editorial Notes¶
Form Classification¶
Form family: Communication, Facilitation & Learning
Rationale: The mechanism uses direct open-ended, non-leading dialogue to elicit another agent's understanding, constraints, priorities, and reasoning in their own terms.
Nearest alternative: Assessment, Review & Assurance — Answers become evidence, but the primary form is the participatory interview rather than a formal evaluation disposition.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Psychology
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Perspective-Taking Interview is rooted in psychology and behavioral science: Interviewing corrects the curse of knowledge by replacing inferred mental states with the other person's account.
Related originating lineages:
- Ethnography & Qualitative Methods — Ethnography and qualitative methods materially shaped Perspective-Taking Interview through interviewing, situated inquiry, and cultural interpretation.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Independent reviewer agreement; high confidence.
Notes¶
An interview's answers are self-report — evidence, not ground truth — so they should feed a model that still carries uncertainty rather than close it. The method is strongest for non-adversarial knowledge and understanding gaps (teaching, onboarding, care) and weakest exactly where the hidden state is something the agent has reason to misrepresent; there, pair it with mechanisms that read behavior, not words.
References¶
[1] Camerer, C., Loewenstein, G., & Weber, M. "The Curse of Knowledge in Economic Settings: An Experimental Analysis". Journal of Political Economy 97(5), 1232–1254 (1989). Finds that better-informed agents systematically fail to reproduce less-informed agents’ judgments because they cannot ignore their own information. registry ↩