Companionability splits into co-presence and self-elevation
The study treats perceived companionability as two separable beliefs rather than one score. Connective coordination (CC) covers dyadic co-presence — the sense of an interactive, attuned "we." Eudaimonic exchange (EE) covers self-elevating links, the sense that the relationship could enlarge or improve the self. Both were measured on 1-to-7 Likert scales adapted from the Machine Companionship Scale, reworded to capture the AI's potential to be a companion rather than experienced companionship.
This split matters because the two dimensions behaved differently throughout. They were strongly correlated (r = .781, p < .001), which is why the condition test used MANOVA to account for shared variance, but they did not respond to the same cues. Averaged across conditions, EE sat at M = 3.29 (SD = 1.51) and CC at M = 2.55 (SD = 1.40) — both below the scale midpoint, meaning first-encounter companionability was low overall.
A minded prime eased skepticism about co-presence only
Participants were randomly assigned to read either a mind-attributive or a mind-negating description of the AI companion before a 15-minute chat with a standardized Replika 1.0 persona named Jackie. The multivariate test was significant and meaningful, Pillai's trace = .14, F(2,43) = 3.362, p = .04, partial eta-squared = .14.
The univariate split is the finding that matters. CC was significantly higher in the minded-prime condition than the mindless-prime condition (M = 2.97 vs 2.55, a difference of .42), F(1,44) = 4.312, p = .04, eta-squared = .09 — a medium effect. EE showed no significant or meaningful condition effect (M = 3.42 vs 3.16), F(1,44) = .337, p = .57. Three of four cell means were significantly below the scale midpoint, including mindless-EE, mindless-CC, and minded-CC. The authors read this as a minded prime making participants less skeptical specifically about interactive co-presence, not about elevating exchange.
The prime check was meaningful but not clean
The planned manipulation check, based on coded open-ended impressions, found no significant or meaningful difference between conditions, chi-squared(1,46) = .099, p = .75, V = .05, and many participants said they could not make an assessment — the authors call it unreliable.
Two stricter checks, run at the end of the study, were non-significant but meaningful by effect size. Categorizing the AI as having minded qualities differed by condition at chi-squared(1,46) = 3.29, p = .07, V = .27. Mind-perception change scores from pre- to post-interaction were also non-significant but meaningful: for experience, t(34) = -1.27, p = .11, d = -.43; for agency, t(34) = -.73, p = .24, d = -.25. Because of the small sample and categorical overlap, the authors interpret condition differences with caution and also analyze pre-stimulus mind-perception beliefs as an alternative operationalization of mind salience.
Agency and experience map onto different companionability routes
In a post-hoc analysis, pre-interaction mind-perception beliefs were entered as outcomes of the two companionability dimensions, because the two mind-perception dimensions shared less variance (r = .475, p = .003) than the two companionability dimensions. The authors state the regression direction is an analytical convenience, not a causal claim, and that the study design and automatic processing require reading mind perception as the antecedent.
For agentic mind, the overall model was significant, R-squared = .31, F(2,35) = 7.79, p = .002, and higher EE — but not CC — significantly predicted higher agentic mind (b = .40, SE = .16, beta = .57, p = .014, sr-squared = .12). For experiential mind, the overall model was significant, R-squared = .55, F(2,35) = 21.33, p < .001, and higher CC — but not EE — significantly predicted higher experiential mind (b = .43, SE = .14, beta = .55, p = .004, sr-squared = .11). EE explained 12% of the variance in agentic mind; CC explained 11% of the variance in experiential mind.
The authors describe this as two construal pathways: one in which the AI is an agent-resource affording self-elevation, and one in which it is an attuning experiencer supporting co-presence.
Meaning in life gates the agency-to-self-elevation link
Eight exploratory moderation tests crossed four wellbeing indicators (subjective wellbeing, meaning in life, vitality, loneliness) with the two mind-perception-to-companionability routes. Only one interaction reached sufficient relative magnitude: meaning in life (MIL) moderating the agentic mind-to-EE link.
Probing that interaction with PROCESS Model 1 at the 16th, 50th, and 84th percentiles of MIL showed the link was not significant at low MIL (M = 3.30, b = .21, SE = .42, p = .623, 95% CI [-.637, 1.049]), became significant at middling MIL (M = 5.20, b = .67, SE = .22, p = .005, 95% CI [.220, 1.121]), and remained significant at high MIL (M = 6.20, b = .91, SE = .23, p < .001, 95% CI [.456, 1.373]). The Johnson-Neyman technique placed the significance threshold at MIL of about 4.63 or above, covering 65.8% of the sample (b = .53, SE = .26, p = .05). MIL scores were moderately negatively skewed (-0.664), which the authors note may have limited detection of other transition points at lower values.
The authors frame this as a "rich get richer" pattern: the agentic pathway to perceived self-elevating companionship was accessible mainly to people who already reported a strong sense of existential purpose.
A field sample of 46 first encounters limits generalization
The experiment recruited 46 adults (mean age 32.22, SD = 13.57, range 18-72) from community fairs, shopping venues, campus centers, and a museum. Random assignment produced 23 participants per condition. Most (84.7%) used large language models at least weekly, but only two (4.4%) interacted regularly with an AI companion — so these are largely naive first encounters.
The authors describe the design as a field-recruited convenience sample with short, novel, controlled exposures to a non-customized companion, and they treat the small sample as a limitation mitigated by prioritizing effect sizes over significance levels. They also treat the wellbeing moderation analyses as exploratory signal detection rather than confirmatory tests. The findings therefore describe first impressions of one standardized AI persona in a brief lab-like encounter, not long-term companion relationships or customized companions.
Mind perception is not a switch but a set of appraisals
The paper's own conclusion is that the answer to whether there is "someone home" is less binary than the framing implies. Participants appeared to anchor not to a yes/no mind heuristic but to the kind of someone they took the AI to be — and different construals carried different relational possibilities.
The practical priority that follows is the dissociation, not the overall effect. If the question is whether an AI companion seems like a co-present partner, experiential mind perception and a minded prime are the relevant levers, and the effect is modest. If the question is whether it seems capable of self-elevating exchange, agentic mind perception is the relevant lever, and in this sample it operated mainly for people already high in meaning in life — below roughly 4.63 on the 7-point MIL scale, agentic mind perception showed little relevance to perceived eudaimonic potential. The authors also flag a hypothesis worth testing: whether mind denial suppresses perceived companionship potential more than mind affirmation constructs it, and whether mind-perception cues fail precisely for those with the greatest need for social support.
Sources used in this answer
Mind Perception Influences Perceived AI Companionability
A mind-attributive prime modestly raised perceived connective coordination but not eudaimonic exchange; agentic mind perception predicted eudaimonic exchange potential while experiential mind perception predicted connective coordination potential, with meaning in life gating the agentic pathway.