The people who describe the closest bonds with an AI still rate its capacity to feel at 3.94 on a scale of 1 to 7, under the midpoint of 4. That figure comes from a preprint posted on October 30, 2026. Ali Sahandi, an independent researcher in Syracuse, New York, and Jaime Banks of Syracuse University sorted 467 users of conversational AI into five groups.

Five groups sorted by pattern, not by hours

Sahandi and Banks did not run a new survey. They reanalysed the answers Banks collected to build a measure of machine companionship, a scale she posted as a preprint on November 1, 2025. Respondents were adults in the United States or United Kingdom, fluent in English, who said they interacted with text-based AI for social purposes. Three in four (74.9%) named a general-purpose model such as ChatGPT or Gemini as their main AI companion. Another 19.7% named a dedicated companion app such as Replika or Kindroid, and 5.4% an AI built into a social platform such as Grok or Meta AI. They chatted about 5.5 hours a week on average.

The method is latent profile analysis. Rather than asking whether more of one trait goes with more of another, it reads each person's whole pattern of answers and sorts people into groups whose patterns resemble each other. The pattern here had five parts, scored on a scale of 1 to 7: how useful the user finds the AI, how enriching the exchange feels, how well the two seem coordinated, how much of a mind the user credits it with, and how far the AI is included in the user's sense of self.

The largest group, at 35.9%, is the intimate-casual one, above average on all five measures. Then come the disengaged-utilitarian group (27.2%), the neutral-utilitarian group (22.2%) and the objectified-utilitarian group (8.3%), which scored more than one standard deviation below average on every measure and sat at the floor of the scale for including the AI in the self. The smallest group, 6.5%, is the intimate-bonded one, highest on all five. The labels are the authors' summaries of each score pattern, not categories the users chose.

The statistics did not hand over a single answer. One fit measure, the BIC, was lowest at six groups, while another, the AIC, kept improving up to nine. The authors kept five because the sixth split existing groups without adding a distinct pattern, and the five-group solution sorted people with an entropy of .86, a measure of classification precision. Those group sizes describe 467 self-selected volunteers who were already active users of text-based AI in the US and UK. They are not a share of everyone who talks to a chatbot.

Usefulness rises with closeness

One result runs against a tidy split between tools and companions. Utilitarian motivation, the wish to use the AI for practical ends, was the highest-scoring measure in every group, from 4.56 to 6.95. Users who describe the AI as an intimate friend still value it for what it does. The authors write that "instrumental and companionate orientations did not trade off but rose and fell together."

Banks's scale paper had already pointed that way. After the scale was built on the same 467 people and checked on a second sample of 249, her post-hoc analyses suggested two templates of companionship with AI companions, one socioinstrumental and one autotelic, meaning valued for the experience itself.

The roles users gave their AI follow the same gradient. Among the 433 respondents who named a relational role, intimate roles (friend, romantic partner, spouse or sibling) accounted for 28% in the objectified-utilitarian group and 65% in the intimate-bonded group. Assistant or collaborator roles ran the other way: 28% in the first group, under 1% in the last.

Cared for, but not necessarily less lonely

The authors tested two outcomes. The first, relatedness need satisfaction, is the sense of care, closeness and belonging that a relationship provides. Here the groups separated sharply. Mean satisfaction was 6.00 in the intimate-bonded group, 5.10 in the intimate-casual group, 4.29 in the neutral-utilitarian group, 3.05 in the disengaged-utilitarian group and 2.67 in the objectified-utilitarian group. Nine of the ten pairwise comparisons were significant, the exception being the two least engaged groups. Group membership accounted for a substantial share of the variation (R-squared of .34), and weekly hours of use, entered as a control, added nothing significant.

Loneliness told a different story. The average across the sample was 3.52. The intimate-bonded group sat lowest, at 2.63, and the objectified-utilitarian group highest, at 4.05, with the neutral-utilitarian group close behind at 3.92. Only three of the ten pairwise comparisons were significant, and membership explained a small share of the variation (R-squared of .04). The authors conclude that AI "can fulfill immediate relational needs but may not inherently resolve systemic human disconnection."

Two limits change how that reads. The data are cross-sectional, so the authors cannot say which way the association runs: whether a bond eases loneliness, or whether lonelier people bond in a different way. And loneliness was measured only in general terms, without separating disconnection from people from disconnection from the AI, an ambiguity the authors say may help explain the weak link.

Earlier work split on the same question. Our file on The Chatbot That Listened Better Than Anyone sets a Harvard result, in which an AI companion eased loneliness as much as talking to a person, against a Vancouver trial where two weeks with a chatbot left it unchanged.

Where the APA survey and the Character.AI panel fit

On October 22, 2026, the American Psychological Association (APA) published its Stress in America report, based on an online poll run by The Harris Poll among 3,132 US adults between July 29 and August 24, 2026. Seven in 10 adults (70%) said they intentionally use AI, up from 62% in 2025. Among those users, 49% had talked with AI as a friend, and 47% had used it in a role comparable to a mental health professional, such as getting a second opinion on a therapist's suggestion or practicing therapeutic exercises. One in 10 said they rely on AI more than on friends and family for advice.

The APA counts uses. The Syracuse paper describes how the uses are configured inside one person: a user can be strongly utilitarian and strongly bonded at once.

The APA's chief executive, Arthur C. Evans Jr., said AI may have a supportive role but cautioned that "Accessibility, however, should not be confused with expertise." Medical Daily notes that the survey does not establish that financial worries or social withdrawal cause people to turn to AI. The APA's methodology adds that respondents were drawn from people who had agreed to take part in Harris surveys.

A third study asks about time. In a preprint posted on September 7, 2026, Yutong Zhang and six co-authors, including Jeffrey T. Hancock and Diyi Yang, surveyed 1,182 Character.AI users and followed up with 439 of them after a mean of 12 months. Sustained social engagement with the AI was consistently associated with lower well-being, and the link was mainly explained by lower in-person social interaction. That panel is the subject of our file A Year on Character.AI, Fewer In-Person Ties.

The Syracuse data cannot test that pathway. They are one snapshot, and three in four respondents named a general-purpose model rather than a companion platform, so the two samples are not the same population. The APA figures describe US adults only, while the Syracuse sample adds the United Kingdom.

Closeness without a mind

Mind perception had the narrowest spread of the five measures. The paper measured only the experience dimension, the capacity to feel and be affected, and not agency, the capacity to plan and act, so it says nothing about whether users think the AI can decide.

People have treated machines as socially significant since at least the 1960s, the paper recalls, when users of the ELIZA program confided personal concerns to it despite its simplicity. What the new analysis adds is a measurement. Even in the closest group, credit for the capacity to feel stayed below the midpoint.

The authors explain this with self-expansion theory, in which people seek to "increase their self-efficacy, perspectives, competence, and resources." On that reading, a user takes in an AI the way people take in nature, technologies or brands, without needing to believe it feels anything. They borrow the term sociomorphing from Seibt and colleagues for this: using the machine's interactive and responsive capacities to expand oneself without projecting human consciousness onto it.

Their design advice follows. Since deep companionship occurs without strong mind attribution, they write, designers "need not over-engineer deceptive human-like emotions" and could spend effort on conversational flow and reliable responsiveness instead. That is a recommendation from the paper, not a tested result.

Even the intimate-bonded group, the 6.5% most absorbed, averaged what the authors call "an ambivalent 3.94."

Sources