Interactive inference

Introduced in Theriault, Young & Barrett (2021), and best read as the constructive counterpart to Barrett’s mental-inference-fallacy. Barrett names the error of inferring an internal state from an observed action; this construct is the same lab’s positive answer to the problem the fallacy leaves open — if minds cannot be read off behaviour, how are they inferred at all?

The problem it solves

To conform to others’ expectations (a sense-of-should), you must first infer what they expect. But their expectations — the predictions issued by their internal models — are not observable. The paper reviews the three standard accounts of this “mental inference” (it avoids the term theory of mind and its propositional-representation baggage): simulation theory (use your own mind as a simulator), modular theory (an innate mind-reading module), and ‘theory’ theory (mental inference as a special case of causal inference, children as little scientists). Each, taken alone, is circular for the same reason: estimating what others expect now requires prior estimates of what they expected before, and none explains how the very first estimate was formed.

The proposal

You break the circle by experimenting. Formally (extending the model’s Equation 5), your prediction error from person i is proportional to |b − (pb_i^M + e)| — the discrepancy between your behaviour b, your estimate pb_i^M of their expectation, and the error e in that estimate. If your behaviour and estimate are known to you, the residual prediction error is a read-out of your inference error. So:

1. Form a hypothesis about what the other expects, from prior knowledge — a dispositional inference (what this person has expected before) or a situational inference (what people typically expect in this context), each framed as a Bayesian posterior. Subsets of the situational prior are where stereotypes enter the model.
2. Enact a behaviour — the “experiment.”
3. Read the resulting prediction error as evidence — the “data.” If, across iterations, the other’s behaviour becomes more predictable (your arousal falls), your estimate is converging on their expectation; if less predictable (arousal rises), you are diverging.

Across iterations you cumulatively construct a model of the other’s mind. The estimate is a hypothesis, the behaviour an experiment, the prediction error (and its affective consequence) the evidence — mental inference as scientific inference, but embodied and interactive. The name is after Shaun Gallagher’s interaction theory: understanding others is in part an embodied practice, not a detached reading of latent states.

Why it is an interoception construct

The evidence in the loop is not only behavioural. The prediction error carries an affective signal — rising or falling arousal, an interoceptive read-out — that tells you whether you are converging. So the felt body is part of the inference machinery, not a bystander to it: the same core-affect that makes the sense-of-should aversive also indexes whether your model of the other person is getting better. This connects to the developmental picture on social-origins-of-interoception — infants learn others’ expectations through interaction before they can represent them — and to the simulation-map as the internal model doing the predicting.

Consequences the paper draws

• A limit on machine learning of mental states (e.g. emotion-recognition algorithms): they receive data but do not interact with the humans it came from, so they miss the trial-and-error refinement that human mental inference depends on.
• Social cohesion / fluency. Interactive inference is cheap with familiar, similar others (your priors are good, so few experiments are needed) and expensive with unfamiliar or outgroup others (bad priors, prolonged trial-and-error, aversive arousal) — which can tip toward avoidance. Fluent interactions “feel right” because they minimize prediction error, arousal, and metabolic cost.
• The intentional vs mechanical stance (Dennett): simple non-biological systems (a clock, an automatic door) need the inference run only once, after which a fixed input–output model suffices; living systems keep changing, so interactive inference must be re-run continually.

Relation to the mental-inference-fallacy

The two are complementary, from the same lab. Barrett’s fallacy says you cannot equate an observed action with an internal state — the mapping is many-to-many. Interactive inference agrees, and says minds are nonetheless inferable — not by reading states off single behaviours, but by iteratively testing hypotheses against prediction error over an interaction. It is the positive account that the negative one implies is needed.