Why people ask AI for relationship help: skills, energy and context

Beyond connection versus isolation

Do general-purpose chatbots help people connect, or draw them away from others? A preprint by Emily Tseng, Patrick Gage Kelley, Renee Shelby and Ashley M. Walker proposes a more specific question: what does a person need when they turn to AI for social or emotional support?

Posted to arXiv on November 6, 2026, the paper introduces the Relational Capacity Framework. It describes people’s ability to form, maintain, change or leave relationships through three resources: social skills, bandwidth and contextual information. The central argument is that similar chatbot interactions can serve different functions depending on which resource is missing.

What the researchers studied

The team recruited 100 English-speaking US adults through the dScout research platform. Participants had to be at least 19 and have used a general-purpose chatbot for an emotional or interpersonal conversation within the previous 30 days. Researchers deliberately selected participants from 527 qualified respondents to include varied demographics and usage frequencies, rather than drawing a representative population sample.

All participants completed an open-ended survey and recorded themselves thinking aloud while interacting with their chosen AI system about a typical social or emotional need. Twenty-three also completed hour-long interviews. Data collection ran from February through March 2026.

The researchers interpreted this material alongside existing theories from psychology, sociology and human-computer interaction. Their initial focus on social skills did not explain all the accounts, leading them to add bandwidth and information. This was an interpretive study of experiences and practices, not an experiment testing whether AI improves relationships.

Three reasons to seek help

Social skills included communicating, interpreting others’ behavior, judging how to handle relationships, asking for support and supporting others. Participants also sought help calming themselves, caring for themselves and understanding their emotions. Importantly, the authors distinguish seeking skill support from demonstrating that skills actually improved.

Bandwidth means the time, energy and emotional resources available to act, together with access to other people’s support. Someone may know how to communicate sensitively but struggle to do so while grieving, caregiving or exhausted. Some participants described turning to AI because friends or relatives were unavailable, or because they did not want to burden them.

Contextual information is the knowledge needed to apply an existing skill in a particular setting. Participants sought help navigating unfamiliar workplace expectations, parenting situations and bureaucratic systems. In these accounts, the problem was not necessarily an inability to relate to others, but uncertainty about the circumstances.

Different needs, different risks

Combining relatively high or low levels of the three resources produces eight states in the framework. These are temporary configurations, not fixed personality types. A person can move between them or occupy different states in different relationships.

One medical student, for example, described using AI to rehearse difficult patient conversations: the chatbot supplied scenarios, while she practiced her own responses. Another participant, a nursing professor coping with her mother’s dying, reported using AI to interpret medical information when grief made her professional knowledge difficult to access. The researchers interpret these as different functions: practicing a skill and temporarily supporting depleted bandwidth.

Other accounts exposed risks. A participant seeking help with her fiancé’s behavior received a gender-stereotyped explanation that the authors argue could normalize a potentially harmful dynamic. Another described relying on AI to direct many daily decisions. These cases illustrate concerns about misleading advice and surrendered agency; they do not establish how frequently those outcomes occur.

The framework also emphasizes metacognition: recognizing what one lacks, choosing appropriate help and evaluating the response. Yet even reflective users can encounter confident misinformation or advice that misses crucial relationship context.

What remains unproven

The study offers a conceptual model grounded in participants’ accounts, not a validated prediction tool or a causal estimate of AI’s effects. Its selected sample of existing users cannot establish how common these practices are across the public. Self-reports can overstate competence or omit difficulties; recorded interactions and interviews only partly address those limitations.

The research also does not establish lasting skill gains, dependence or relationship deterioration. Suggested risks—including reduced tolerance for disagreement or vulnerability—need investigation over time.

The authors recommend systems that encourage reflection, practice and informed action rather than pronounce judgments about relationships. They also warn against treating structural pressures, such as caregiving burdens or economic insecurity, as personal shortcomings that chatbot coaching can solve. Their proposed test is longer-term: does assistance leave people better able to manage relationships with other humans?