The starting point is the familiar experience of discovering that apparently mutual recognition was one-sided. A partner may repeat the right words, mirror important values and appear to anticipate emotional needs, yet later prove unable to explain what the relationship had seemed to reveal. In the theory, the perceiver is the person who feels understood, while the responder is the partner or system whose understanding is being credited. The construct separates “felt understanding,” the perceiver’s appraisal that the responder grasps their thoughts, history and concerns, from “demonstrated understanding,” the content the responder puts on record in a form that can be checked against the perceiver’s own account. This separation is central because the effect concerns the distance between what someone believes and what the interaction can substantiate, not the possibility that people sometimes misread subtle behavior.

The proposed responder condition has two routes through which calibration can fail. Under the first, a responder produces specific, potentially falsifiable attributions about other people on matters of comparable intimacy, but makes such determinate statements about the perceiver below a coded threshold. Under the second, an attribution is explicitly and literally disconfirmed by the perceiver and then reasserted in its uncorrected form on at least two later, independent occasions. An attribution must be specific enough to be wrong: “You are someone who needs honesty” is broad and difficult to reject, whereas a statement about why someone became quiet after a painful family disclosure can be confirmed or corrected. A partner who is generally reserved cannot be classified through the first route, because the condition asks for evidence of selective failure to state what could be wrong. Likewise, a single failure to absorb a correction is insufficient; the proposed manual requires persistence, distinguishing a sustained relational pattern from an isolated conversational lapse.

Persistence, rather than immediate echo, is the defining feature. ELIZA could return transformed versions of a user’s earlier words, demonstrating continuity without modeling the person using it. The interpersonal proposal focuses instead on whether corrections remain available yet fail to alter the responder’s later conduct. For a reassertion to count, the correction must have been explicit, literal, delivered in a register the responder demonstrably understands and followed by a later occasion in which the corrected content was relevant. A responder who acknowledges the correction in the next turn and subsequently carries it forward has not met the condition. The paper likewise distinguishes a contradiction from a vacuum: the first route leaves the perceiver’s account unanswered, whereas the second repeatedly asserts something the perceiver has denied. Meeting either route does not by itself establish gaslighting, coercion or any diagnosis, because the proposed criterion describes contributions rather than a person’s motives or character.

The perceiver’s side of the configuration is described as “readiness,” a graded variable rather than a categorical trait. It has two proposed facets: motivation to be understood and the threshold at which an ambiguous contribution is accepted as evidence of understanding. The theory draws on research and clinical formulation involving attachment insecurity, complex post-traumatic stress and codependent relational scripts, but it does not present these as proven causes of the effect or divide people into fixed high- and low-readiness groups. In anxious or fearful attachment, a person may monitor a partner for signs of withdrawal and accept minimal cues as recognition. A trauma-trained perceiver may settle on an interpretation after little information because rapid judgment once served protection. The article explicitly labels the clinical extension to adult relationships as untested, even where evidence from childhood shows that altered cue processing can be more efficient rather than simply defective.

Several perceiver-side operations are proposed to create and maintain the gap. Returning someone’s vocabulary without making a specific inference can nevertheless be experienced as evidence that their state has been grasped. Indeterminate responses may be completed from the perceiver’s own emotional repertoire, while silence or a brief message may be assigned the most generous interpretation. Protective re-attribution can then preserve the original belief: a contradiction may be recoded as poor expression, hidden sensitivity or an unconventional form of care, and intermittent warmth may renew the impression that the partner possessed depth all along. The article separates credited interiority from felt understanding because a person may believe that a partner has a rich inner life while simultaneously doubting that the partner notices anything important. The bridge from imagined depth to felt recognition is treated as a conjecture rather than an established step.

The proposed maintenance loop connects four ideas. Attributional completion allows the perceiver to supply the inference the responder leaves open. Withdrawal of the accuracy demand occurs when the perceiver stops requesting clarification, testing interpretations or asking the partner to put understanding on record, even as felt understanding rises. Accuracy decay proposes that empathic accuracy is effortful and may remain low when it is neither requested nor consequential. Cue recycling describes a responder’s contributions becoming increasingly dominated by material already supplied by the perceiver, leaving fewer independent constraints on interpretation. The paper stresses that this sequence has not been directly observed in a dyad meeting the responder condition. A separate path-dependence conjecture proposes that dyads may continue differently depending on the early order in which accuracy was demanded and supplied, but mutual influence alone does not establish this historical effect.

To test the theory, the author adapts the empathic accuracy paradigm. During an interaction, the perceiver records a specific statement about their own state, and the responder independently records what they think the perceiver is conveying. Blind raters then compare the responder’s inference with the perceiver’s original report, while the perceiver rates how far the response appears to grasp the content. Both felt and demonstrated understanding are placed on a three-point correspondence scale, with model-implied differences reported in points rather than as raw coefficients from an ordinal model. The protocol distinguishes average discrepancy, the slope tracking demonstrated understanding and conditional bias at a predetermined criterion level. It also models contents within episodes within dyads and proposes equivalence bounds, including 0.3 points for major contrasts and 0.1 point per standard deviation for effects involving continuous readiness. These are planning values, not established clinical thresholds, and the method’s comparability with the original empathic accuracy procedure still requires validation.

Five hypotheses connect short-term analog studies with a six-wave longitudinal field module. Analog participants would encounter responder conditions in which corrections are carried, wrong attributions are reasserted, or determinate attributions are set at a low rate, with blocks matched afterward on the demonstrated understanding actually achieved. Predictions are that felt understanding will exceed demonstrated understanding more strongly in the latter two conditions, especially at higher readiness, and that feedback from outside the dyad will narrow the discrepancy more readily in comparison interactions. A vignette experiment asks whether unpredictable warmth produces a larger increase in attributed depth than warmth with the same average frequency but a more predictable order. A six-wave study would then test whether the excess of felt over demonstrated understanding fails to decline after a dyad meets the responder condition, rather than merely measuring a discrepancy at one occasion. Failure of the readiness interaction or the longitudinal maintenance test, provided the measurement probe meets predetermined evidential standards, is specified as grounds for withdrawing the construct.

The paper’s ambition is accompanied by substantial limitations. The coding manual, thresholds and anchor set remain provisional, inter-rater reliability has not been demonstrated, and the prevalence and predictive validity of the configuration are unknown. The adapted probe may itself elicit the accuracy the mechanism predicts will be absent, while asking the perceiver to state and lock in their internal contents may interrupt the very attribution process under study. Cultural and gender variation in readiness and in what counts as a determinate attribution is also unaddressed. Even if the proposed effect survives testing, it would not show that every feeling of profound understanding is false, nor that a partner who feels misunderstood is itself harmful; the account is intended to locate a specific failure of interpersonal calibration. Its most consequential test will therefore be whether objective records of what a responder demonstrates can explain when a perceiver stops checking, and whether correction consistently fails to reduce the gap between felt and demonstrated understanding over time.

Subject of Research: A dyadic theory of persistent over-attribution of interpersonal understanding

Article Title: The Interpersonal ELIZA Effect: a dyadic account of persistent over-attribution of understanding

Article References: Riehm, A. (2026). The Interpersonal ELIZA Effect: a dyadic account of persistent over-attribution of understanding. Frontiers in Psychology, 17, Article 1869060. https://doi.org/10.3389/fpsyg.2026.1869060

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1869060

Keywords: Interpersonal ELIZA Effect, empathic accuracy, felt understanding, demonstrated understanding, perceived responsiveness, attachment insecurity, AI companions, ELIZA effect, relationship psychology, projective attribution, Interpersonal, ELIZA