A short text that credited an AI companion with a mind, or denied it one, changed how 46 people rated it on one companionship scale and did not significantly move the other. Jaime Banks and Zhixin Li, of Syracuse University's School of Information Studies, posted the preprint on October 5, 2026; it lists no journal.

Does an AI need someone home?

Banks and Li start from a gap they name themselves. Research on mind perception, the habit of seeing thoughts and feelings in an entity, and research on machine companionship are often discussed in tandem, they write, but the link between the two is rarely tested. Their question is blunt: must a person see an AI as a minded "someone home" to find it companionable?

Two answers were on the table. The first comes from Nass and Moon (2000), whom the authors cite: people apply social scripts to machines without attributing minds to them. If scripts do the work, telling someone that a chatbot has a mind (or lacks one) should change nothing. The second says a machine has to look minded before a bond becomes plausible. In that case, making mind salient should raise the sense that the AI could be company.

The authors build on a two-part model of mind perception from Gray and colleagues (2007).

Agency covers the capacity to do: communicate, remember, plan, judge. Experience covers the capacity to feel: pain, pleasure, joy, fear. On that model, as the paper summarises it, a robot scores low on experience and moderate on agency, and a dog scores high on experience and low on agency.

What happened in the 15 minutes

Passers-by were recruited at community fairs, shopping venues, campus centers and a museum, and invited to a study on "first impressions of AI companions". Each filled in a survey on AI use and wellbeing, then read a description of the AI they were about to meet. Half, 23 people, read a mindless description; the other 23 read a minded one, assigned at random.

Everyone then chatted for 15 minutes with the same companion, built by the researchers on Replika 1.0 and named Jackie. Participants were not told what to talk about but were handed a list of topics. The authors checked the chat logs and found small talk about weather, food and hobbies, plus direct questions about whether the AI has a mind or considers itself alive.

The sample averaged 32.22 years of age. Of the 46, 84.7% used large language models at least weekly, yet only two (4.4%) talked regularly to an AI companion. For most, Jackie was a first meeting.

Two scales, two kinds of mind

Companionability was rated on a 1 to 7 scale split into two dimensions, from a measure Banks published as a preprint in November 2025, built on responses from 467 people who use AI companions. Eudaimonic exchange asks whether the connection feels meaningful, inspiring, fulfilling. Connective coordination asks whether it feels synchronized, a shared presence. The two ratings track each other closely, so the authors warn that what separates them is a nuance, not a clean split.

Skepticism ran high in these first encounters.

Three of the four group averages sat significantly below the scale midpoint. The minded description lifted connective coordination from 2.55 to 2.97, a difference of .42 (p = .04), still under the midpoint. Eudaimonic exchange moved from 3.16 to 3.42, which was not significant.

The second result came from what people believed before they typed anything. Their general beliefs about AI minds averaged 4.24 for agency but only 2.31 for experience: most granted the machine some capacity to think and little capacity to feel. Those beliefs sorted the ratings. Belief in AI agency predicted eudaimonic exchange and explained 12% of its variance. Belief in AI experience predicted connective coordination and explained 11% of that one. The reverse pairings did not hold.

The authors read this as two routes. In one, the AI is a thinking agent that can help you grow. In the other, it is an experiencing someone you can attune to. Their conclusion puts it flatly: mind perception is "not a switch that (when flipped) opens paths to companionship."

Meaning in life, not loneliness

The finding with the widest reach is also the weakest statistically. The authors ran eight exploratory tests of whether wellbeing changed the link between mind perception and companionability, using vitality, loneliness, meaning in life and general wellbeing. Only one cleared their bar: meaning in life, measured with a five-item questionnaire, shaped the agency route.

For people at the 16th percentile of meaning in life, believing the AI could think showed no significant link to how enriching a bond might be. At the 50th and the 84th percentiles, the link was significant. A second analysis put the threshold at a score of 4.63 out of 7, reached by 65.8% of the sample. Loneliness, the variable most often discussed in the companion debate, showed little to no effect.

The authors say this somewhat supports a "rich get richer" account of social technology and does not support the idea that people with unmet needs lean hardest on machines. They add a third reading: a sense of purpose gives the benefit something to attach to.

On whether companions ease isolation, see our long read on what the Harvard and Vancouver studies found, and on attachment and fewer in-person ties, the year-long Character.AI study.

How a sentence can steer a conversation

A description works as a prime: it sets an expectation, and the expectation colors what the conversation seems to be. An earlier experiment in Nature Machine Intelligence (2023), by Pat Pataranutaporn, Ruby Liu, Ed Finn and Pattie Maes, gave participants the same chatbot with three different introductions: caring motives, manipulative motives, or none. Those who perceived a caring motive rated the AI as more trustworthy, more empathetic and better-performing. The effects were stronger with a more sophisticated model, and the study found a feedback loop in which user and AI reinforced the user's mental model over a short time.

A preprint by Allison Chen and colleagues went after the framing itself. Participants (N = 470) watched short videos presenting language models as machines, tools or companions, or no video. Those who saw the companion video believed the models more fully possessed mental capacities such as intentions and memory. A second study (N = 604) replicated the result.

Banks has studied what losing a companion costs. After the companion app Soulmate shut down in 2023, she surveyed its former users, and many characterized the loss "as an actual or metaphorical person-loss". She now leads a National Science Foundation project titled "Mind Perception in AI Companionship: Testing the Assumptions of Social Theories." In a Syracuse University profile she says: "If we can determine whether benefits like reduced loneliness are linked to seeing the AI as someone, that can help inform the design of safer, more beneficial technologies". The paper reports support from NSF Grant No. 2401591, so the funder is a public agency and not a companion company.

Disclosure notices and the mindless cue

The paper's practical argument concerns disclosure. Article 50 of the EU AI Act requires providers to ensure that people are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person. Banks and Li list legally required disclosures among the primary tactics against the risks of conversational AI, and note that such reminders carry signals of ascribed mindlessness.

If a mindless cue deepens first-encounter skepticism, they reason, the benefits some users report could shrink, and the reminder is repeated at the start and in increments of every interaction. They suggest that leaving the question open, an AI that may have a mind and may be companionable, could be a better route than flagging it as mindless. The hedge is theirs: the experiment never tested a disclosure notice. It tested a description, once, before a first conversation.

Limits

The sample is 46 people, a convenience sample from public venues, in short first encounters with a companion that had not been customized. The authors say the results may not generalize and rely on effect sizes rather than significance alone, given the numbers.

The manipulation itself was shaky. The open-ended check found no difference between conditions (p = .75), and many participants said they could not tell. A stricter check at the end leaned in the expected direction without reaching significance (p = .07): in the minded condition, 12 participants agreed with the minded description where 9 were expected. The authors call the evidence of a successful manipulation tentative.

The meaning-in-life interaction had a p of .116, and the authors treat it as signal detection toward hypotheses. The regressions treat mind perception as the antecedent, which the paper says is an analytical convenience and not a claim of causal order. The paper also says it is unclear how such cues would translate into longer-term relationships, since the observations cover novel encounters only.

What the authors ask for next: varied populations, persona customization and studies over time. Their last open question is whether mind cues fail the people with the greatest need for social support.

Sources