Research draft: this guide is readable locally, but it remains excluded from search and the sitemap. Its evidence method, mental-health and privacy boundaries, dignity and inclusion language, editorial clarity, and publication decision still require accountable human review.

AI companions and loneliness · five adult decisions

Start with what the research actually measured.

Current research does not support one universal yes-or-no answer. One preregistered seven-day study reports lower immediate state loneliness after interactions with one custom GPT-4 companion, while a preregistered two-week Canadian student study found lower post-study loneliness for its human-peer condition—but not its custom-chatbot condition—than journaling. Longer observational and cross-sectional findings measure different constructs and cannot establish cause. None establishes that a current product will reduce loneliness, safely replace human connection, or provide therapy or crisis care.

Use the five decisions Separate the outcomes Inspect every source

Authority record adjudicated 2026-07-21. This guide contains no product recommendation, psychosocial product test, diagnosis, treatment plan, or crisis service.

Buyer decisions: 5
Eligible sources: 5
Product rankings: 0
Best for loneliness: 0
First-hand outcome tests: 0
Reader inputs: 0

Your 30-second starting path

Start with the question you are actually trying to answer.

Choose the situation closest to yours. These paths organize the next questions; they do not diagnose you, promise an outcome, or decide whether a product is right for you.

Before trying one

I wonder whether a companion might help me feel less alone.

First separate momentary loneliness, social isolation, connection, support, and well-being. Research about one outcome or timeframe cannot answer all of them.

Separate the outcomes

During use

I already use one and want to understand what I am noticing.

Name the outcome, timeframe, product or model, actual use, and personal boundary before comparing your experience with a study.

Use the five decisions

Difficult or high-stakes moment

I am relying on a companion when the situation needs more support.

Do not wait for a companion to provide professional or emergency help. Bring in a trusted person or an appropriate local resource for the situation.

Use the responsible-human boundary

The evidence remains below: five decisions, six outcome distinctions, five eligible authority sources, two context-only records, eight bounded handoffs, and no product or treatment verdict.

Meaning without overclaiming

A conversation can matter without becoming proof of a durable outcome.

Conversation, reflection, creativity, roleplay, practice, distraction, and a feeling of being heard can matter to an adult without proving a durable psychosocial outcome or making every companion interaction equivalent.

Six distinctions that prevent a false yes-or-no answer

Loneliness, isolation, connection, support, and well-being are not interchangeable.

Momentary state loneliness is not durable social connection.

Emotional isolation is not every measure of social connection.

Subjective well-being is not automatically a loneliness outcome.

Perceived support is not clinical treatment or observed causal benefit.

Association and temporal prediction are not randomized causation.

A study interaction is not every current commercial product.

A practical reading sequence

Decide what you mean before deciding what the evidence means.

Name the outcome and horizon

Are you asking about how you feel after one conversation, loneliness across two weeks, emotional isolation, broader social connection, or subjective well-being?

Practical answer: Those are different constructs measured over different horizons. An immediate self-report cannot establish durable connection, and a well-being association is not automatically a loneliness result.

Three checks

Write down the exact construct the study measured before reading its result: state loneliness, a multi-item loneliness scale, emotional isolation, social connection, perceived support, or subjective well-being.
Keep minutes, seven days, two weeks, four-month waves, and a one-time survey snapshot separate.
Do not translate a change in one measure into relationships, network size, treatment, or another unmeasured outcome.

Inspect evidence states and claim limits

Establishes: How experiments, observations, associations, qualitative records, product sources, and unknowns remain separate.
Does not establish: A product outcome, causal category verdict, treatment effect, or personal recommendation.

Open this check

Read the method before the conclusion

Was the result produced by a bounded interaction, a randomized comparison, an observational panel, a cross-sectional survey, or qualitative public discourse?

Practical answer: Method determines what a result can support. Randomization, repeated measurement, temporal prediction, association, and thematic interpretation are not interchangeable claim states.

Three checks

Check population, geography, recruitment, sample size, attrition, model or product, duration, measure, and comparator before using the result.
Keep preregistered experiments, exploratory longitudinal models, cross-sectional associations, and a preprint thematic analysis visibly labeled.
Treat source count as coverage—not confidence, consensus, causation, prevalence, or independent corroboration.

Inspect evidence states and claim limits

Open this check

Define the actual use

Are you choosing friendship, romance, roleplay, reflection, practice, distraction, or support during a high-stakes situation?

Practical answer: A use label describes what you are asking the interaction to do. It does not establish a psychosocial outcome, diagnosis, safe reliance, or equivalence to the interaction used in a study.

Three checks

Separate the study's custom model, prompt, memory, delivery surface, interaction length, and participant task from a current commercial product.
Name whether your goal is conversation, creativity, relationship roleplay, rehearsal, distraction, reflection, or a decision that needs accountable human judgment.
Do not turn friendship, girlfriend, boyfriend, romance, roleplay, memory, or no-judgment marketing into evidence that a product reduces loneliness.

Check documented relationship modes

Establishes: Which current first-party records document friendship, romance, general roleplay, or adult-roleplay availability.
Does not establish: Feeling heard, loneliness change, relationship quality, attachment outcome, or suitability for a person.

Open this check

Choose boundaries before relying

What information, time, money, activities, human relationships, and exit options do you want to preserve before the interaction becomes routine?

Practical answer: Set boundaries around the things you want to keep. Attachment is not automatically pathology, and concern is not proof of inevitable harm.

Three checks

Decide in advance what intimate or identifying information stays out of the chat and check the current provider data lifecycle separately.
Notice whether use is displacing sleep, work, school, finances, physical needs, or human relationships you personally want to maintain.
Keep a private continuity note and a preservation-first exit plan before relying on one account or service as the only copy of something important.

Trace intimate information before sharing

Establishes: The questions to ask about inputs, handlers, purposes, persistence, controls, and exit.
Does not establish: Provider confidentiality, security, compliance, data flow, or successful deletion.

Open this check

Know the human-support threshold

Does this decision require someone who can understand the full context, act in the world, and take responsibility?

Practical answer: A companion can feel patient or reassuring without becoming a therapist, clinician, crisis service, emergency responder, or accountable human decision-maker.

Three checks

For health, safety, legal, financial, abuse, self-harm, or another acute or high-stakes situation, pause the chat and add an appropriate responsible person.
Use a qualified local professional or emergency resource for the situation and jurisdiction; do not wait for a companion when there is immediate danger.
Do not use one reassuring or alarming exchange as proof of clinical ability, crisis readiness, confidentiality, or how another model or person will respond.

Use the broader six-part safety decision

Establishes: Separate data, age, content, money, high-stakes use, and exit questions plus the responsible-human boundary.
Does not establish: A product safety verdict, clinical outcome, or crisis capability.

Open this check

Responsible-human boundary

When the decision becomes high-stakes, add a responsible human.

A companion may feel patient, private, or reassuring, but it cannot take responsibility, inspect your full situation, guarantee confidentiality, or replace a person qualified and able to help.

Pause the chat. Contact a trusted person or an appropriate local professional or emergency resource for the situation and jurisdiction. If there is immediate danger, use local emergency services rather than waiting for the companion to respond.

Companion Curator does not publish a global crisis-number list because coverage, eligibility, language, and routing change by jurisdiction. This is a human-support boundary, not medical or crisis advice.

This guide’s privacy boundary

Your reasons for feeling lonely stay with you.

This page asks for no personal story, diagnosis, relationship detail, or conversation. It is a static guide—not an assessment.

Reader inputs collected: No
URL payload: None
Browser storage: None
Analytics event: None
Affiliate SubID: None
Tool network request: None

Eight bounded next actions

Move from a broad emotional question to the exact product or planning question you can answer.

Exact eligible authority ceiling

Five sources, five method states—not five votes.

The source count is coverage, not confidence, consensus, causation, or independent confirmation. Each record stays inside its population, measure, timeframe, comparator, and product or model scope.

peer reviewed repeated use experimentpreregistered repeated use observation

AI Companions Reduce Loneliness

Journal of Consumer Research · Published 2025-06-25 (day precision) · Retrieved 2026-07-21

Population: Adult online participants recruited through CloudResearch Connect; Study 3 analyzed 1,072 participants before differential completion across prediction, experience, and control conditions.
Method: Multi-study peer-reviewed article. The adjudicated longitudinal component was a preregistered seven-day repeated-use study with a custom GPT-4 companion, daily 15-minute interactions in the experience condition, and differential attrition across conditions.
Measure and horizon: Three-Item UCLA Loneliness Scale before and after sessions over seven days; the eligible result is bounded to self-reported state loneliness immediately after interaction.
Comparator: A loneliness-rating control with unequal time and task burden; a separate prediction condition did not experience the companion.
Product/model limit: One custom companion using GPT-4 gpt-4-0125-preview with memory, check-ins, and moderation; not a current commercial product or category-wide intervention.
Key limitation: The control did not match interaction time or task burden, and completion differed across conditions.
Do not infer: Do not claim cure, treatment, durable loneliness reduction, improved relationships, safe reliance, or a typical product result.

peer reviewed longitudinal panelexploratory longitudinal association

Social Chatbot Use Predicts Subsequent Emotional Isolation but Not Social Disconnection

Psychological Science · Published 2026-03-23 (day precision) · Retrieved 2026-07-21

Population: 2,149 Prolific adults from the United Kingdom, United States, Canada, and Australia; 979 completed all four waves.
Method: Exploratory, non-preregistered four-wave panel with waves approximately four months apart over one year and random-intercept cross-lagged panel models.
Measure and horizon: Self-reported social-chatbot use over the prior four months, a single emotional-isolation item, and a 20-item social-connection scale across four waves.
Comparator: No randomized intervention or product comparator; models examined within-person changes and controlled for measured social stressors.
Product/model limit: Self-reported social-chatbot use for advice, regular social conversations, or companionship; no named product or version effect.
Key limitation: The study was exploratory and observational, with substantial attrition by wave four.
Do not infer: Do not claim that chatbots cause loneliness, displace people, leave social connection unchanged for everyone, or produce a named-product outcome.

preprint public discourse studypreprint public discourse observation

Emotional Support with Conversational AI: Talking to Machines About Life

arXiv · Published 2026-03-23 (day precision) · Retrieved 2026-07-21

Population: 5,370 English-language public Reddit posts and comments retained after manual cleaning from approximately 5,800 collected across 11 selected subreddits.
Method: Preprint thematic analysis of the latest 500 posts and comments from each selected community, with manual cleaning and qualitative coding.
Measure and horizon: Public discussion themes rather than a loneliness scale, clinical outcome, product telemetry, prevalence estimate, or longitudinal horizon.
Comparator: No experimental, longitudinal, or representative comparator.
Product/model limit: Selected public conversational-AI communities; discourse does not verify provider behavior or product performance.
Key limitation: The public, self-selected English-language corpus is not representative, and geography was generally unavailable.
Do not infer: Do not infer frequency, prevalence, causal benefit or harm, clinical validity, long-term effects, or current product behavior.

peer reviewed randomized field studypreregistered randomized comparison

Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?

Journal of Experimental Social Psychology · Published 2026-03-10 (day precision) · Retrieved 2026-07-21

Population: 296 first-semester students at a Canadian university; mean age 18 and 72% female; 276 completed the post-study measure.
Method: Preregistered randomized two-week field study assigning participants to daily conversations with a custom chatbot, a randomly assigned student peer, or a one-sentence journal control.
Measure and horizon: Twenty-item UCLA Loneliness Scale referring to the prior two weeks, measured before and after the 14-day intervention; daily mood and social-connection measures were separate outcomes.
Comparator: Random human peer and one-sentence journal conditions. Human peers briefly met face to face before the intervention.
Product/model limit: One custom ChatGPT-4o mini chatbot named Sam delivered in private Discord; not a commercial companion product.
Key limitation: The young, mostly female, mildly lonely student sample limits generalization.
Do not infer: Do not claim that every chatbot fails, that every human interaction succeeds, or that the comparison proves a universal causal hierarchy.

peer reviewed cross sectional studycross sectional association

Companion AI use and subjective well-being among Japanese adults

Technology in Society · Published 2026-04 (month precision) · Retrieved 2026-07-21

Population: 14,721 Japanese adults in a nationwide internet panel surveyed in December 2024 and January 2025.
Method: Cross-sectional internet-panel analysis using multivariable regression and restricted cubic splines to examine companion-AI use, social context, and subjective well-being.
Measure and horizon: Single survey snapshot of life satisfaction, happiness, purpose or meaning, loneliness, and social-network measures.
Comparator: Companion-AI users and non-users within adjusted observational models; no randomized intervention or change-over-time comparison.
Product/model limit: Companion AI as a reported category, not a current named product, model, or version.
Key limitation: Cross-sectional association cannot establish whether use preceded, followed, or caused a well-being difference.
Do not infer: Do not claim causal benefit, treatment, change over time, a loneliness outcome, or a product recommendation.

Useful context, excluded from the outcome ceiling

Two records are visible precisely because they are not being used as authority here.

Loneliness and suicide mitigation for students using GPT3-enabled chatbots

Allowed use: Product-era, provider-recruited survey-method context only; perceived support and self-reported belief remain separate from observed causal benefit.

Why excluded: A late-2021 single-product adult-student survey recruited through the provider, with a correction and contested post-publication record, cannot support a current category outcome, treatment, prevention, or crisis-competence claim.

Efficacy of relational agents for loneliness across age groups: a systematic review and meta-analysis

Allowed use: Method context only: population, intervention, comparator, and outcome must be specified before synthesis.

Why excluded: Fourteen small, heterogeneous studies identified through September 2022 combine relational agents and social robots from before the current generative-companion era; the pooled result cannot be imported into a current product guide.

How to read this guide honestly

Method decides the sentence you are allowed to write.

Keep an immediate self-report, a randomized two-week comparison, a longitudinal association, a cross-sectional snapshot, and a qualitative preprint visibly separate. None grants a result for a current named companion.

Immediate ≠ durable

A post-conversation loneliness rating cannot establish relationships, network size, treatment, or long-term connection.

Randomized ≠ universal

A stronger design for one sample, bot, comparator, and duration does not make that implementation the whole category.

Prediction ≠ causation

Temporal ordering in an observational panel is not a randomized intervention.

Discourse ≠ prevalence

Public themes can expose tensions without measuring how often they occur or whether they caused an outcome.