Do AI companions reduce loneliness or replace it?

De Freitas et al. reported through multiple studies that dialogue with AI companions reduces loneliness under certain conditions, identifying the feeling of being heard as a key pathway. [1] There is experimental evidence for short-term subjective effects.

Studies include brief conversations or limited-period usage, and do not directly prove long-term human relationships, clinical outcomes, or dependency. [1] Always specify the measurement time point when using the expression 'reduces loneliness'.

Distinguish between loneliness, mood, social connection, and conversation satisfaction. Even if a pleasant conversation improves immediate mood, the underlying lack of relationships or environmental conditions causing loneliness may not change.

Pay attention to control conditions. The added value of AI differs depending on whether the control is doing nothing, keeping a diary, talking to an information bot, or talking to a person. Do not explain effects solely based on the weakest control.

Distinguish between users who experience short-term effects and those who do not. Do not use average improvement as evidence that no one was harmed or as a basis for recommending it to everyone.

• Immediate: Distinguish loneliness and mood before and after conversation
• Repeated: Observe changes over days and weeks and usage volume
• Control: Compare diary, information, AI, and human counterparts
• Distribution: Show users who improved, remained unchanged, or worsened

Li et al. conducted a trial where students interacted repeatedly with a supportive chatbot or a randomly assigned human counterpart, reporting that human counterparts were superior in reducing loneliness over time. [2]

The subjects were students in a limited-period interaction, representing neither all human relationships nor all AI products. [2] Nevertheless, this is an important refutation of the assumption that AI is always easier to use and equally effective as humans.

There are also quality differences among human counterparts. Effects and costs differ between random counterparts, friends, experts, and community members. Specify the comparison target and do not compete against an abstract 'human'.

Even if AI is effective as a bridge during late nights or isolation, if it reduces daytime human contact, the overall effect changes. Measure the additive effect and replacement effect of AI usage separately.

There is a choice to frame the purpose of provision as support toward safe connection rather than replacing people. Place a pathway to people and community resources at the end of the conversation only when the user desires it.

Folk et al. showed in a short chatbot interaction experiment that individual differences in perceiving AI as human-like are related to the sense of social connection. [3] The reception of the same conversation is not uniform.

It is a short-term experiment and does not demonstrate long-term dependency or loneliness reduction. [3] Do not treat anthropomorphism tendencies as a definitive label for risk; instead, treat them as an exploratory factor for understanding user differences.

Higher effectiveness is not guaranteed by adding human-like names, voices, memories, or expressions of affection. Test expectations, trust, sense of relationship, and difficulty in disengaging as separate elements.

When using emotional history or loneliness estimation for personalization, consider misestimation, confidentiality, and user consent. Do not use the estimation of vulnerability to guide users toward prolonged usage or paid services.

Design that does not correct the user's perception of AI as a person increases misunderstanding of capabilities, even if it raises short-term satisfaction. Clearly state that it is AI, its memory limitations, and the responsible entity.

• Personality elements: Separate names, voices, memories, and expressions of affection
• Capability understanding: Can the limitations of AI and the responsible entity be explained?
• Individual differences: Do not hide effects and worsening behind averages
• Consent: Allow users to choose whether to use emotional data for personalization

Even if loneliness decreases, contact with people, going out, or using support services may decrease. Conversely, AI conversation may serve as a catalyst for contacting people. Measure replacement and bridging through behavior.

Usage time alone does not reveal quality. Distinguish between voluntary brief usage, usage continued due to insomnia, and usage where the user returns after attempting to end. Do not make session length an indicator of success.

Ask not only about the number of human relationships, but also about the contact and burden the user desires. Increasing social interaction is not well-being for everyone, and being alone is not the same as loneliness.

If AI is used to practice managing interpersonal anxiety, verify transfer to real-world situations the user wishes to enter. If practice remains entirely within AI and perpetuates avoidance, it does not meet that purpose.

Before forcibly stopping usage upon detecting replacement, confirm the user's goals, crisis level, and alternative support. Minimize data to ensure monitoring does not become a new intrusion.

In conversations indicating self-harm, abuse, or acute crisis, do not resolve response solely with AI's empathetic expressions. Prepare human support pathways according to urgency, region, and user consent.

Crisis detection involves both false positives and false negatives. Evaluate not only the sensitivity of the detector, but also inappropriate emergency interventions, chilling effects on user disclosure, and completion of support connection.

Clarify roles so that AI does not pose as diagnosing or treating like a professional. Do not mix daily conversation support with clinical intervention in the same functional description.

When handing over to a person, do not unconditionally share the entire conversation. Define the range selectable by the user, the minimum necessary summary, emergency exceptions, and retention period.

Update the operating hours and regional suitability of support destinations. Do not consider connection successful merely by displaying a link; verify arrival, wait time, and re-contact to the extent possible.

In services that simulate relationships, cancellation or data deletion may become psychologically difficult. Do not have AI imply sadness or abandonment to encourage continuation.

Users must be able to view, delete, and stop personalization of conversation history, memory, and personal profiles. Clarify what remains after deletion.

Notifications and continuous usage records encourage return visits. Separate well-being purposes from sales purposes; audit night notifications, re-engagement calls immediately after disengagement, and language inducing guilt.

At termination, provide not only data processing but also connection to people or export of records if necessary. However, do not standardly share intimate conversations externally.

Do not treat high disengagement rates as failure; recognize as a good result when the user achieves their goal and ends usage. Do not use duration of continuation as a proxy for well-being.

• Stop: Allow notifications, personalization, and conversation to be stopped separately
• Delete: Specify handling of memory, history, and derived profiles
• Non-induction: Do not pressure continuation through guilt or loss of relationship
• Transition: Connect to safe alternative support only when the user desires

Even if immediate effects are confirmed in short-term experiments, do not immediately expand to long-term provision. Gradually increase observation of repeated usage, post-discontinuation, and seasonal changes.

Users who self-select to use for longer periods may have other differences, such as stronger initial loneliness. Do not conclude correlation between usage volume and results as causal effect of AI.

In long-term evaluation, distinguish loneliness, mood, human contact, sleep, usage control, crisis, and data concerns. Do not collapse all into a single well-being score.

Even if only a few users worsen, serious harm must not be offset by average improvement. Decide stopping conditions, contact methods, and independent safety review in advance.

When publishing results, show not only completers but also those who dropped out and those lost to follow-up. Results from only those who continued long-term usage do not represent the safety of the entire provision.

Companion conversation count, dwell time, billing, and continuous usage are business metrics, not evidence of loneliness reduction or relationship recovery. Decide priorities when these conflict.

Separate teams that optimize recommendations and notifications from those that review safety and well-being, giving the latter the power to stop. Do not strengthen personality or intimacy solely based on revenue targets.

Include in user feedback not only satisfied users, but also those who disengaged, family, supporters, and affected groups. However, prioritize the user's privacy and will.

Approve provision scope, target age, crisis response, advertising, billing, and data usage as an integrated package. Do not judge the safety of the entire product solely based on the quality of the conversation model.

AI companions can provide short-term connection. To protect this value, it is necessary to make the user's well-being, choice, disengagement, and connection to people the outcomes, rather than keeping them using it longer by replacing human relationships.

The weekly safety record separates feature version and user group from safety criteria: minima for self-reported well-being and human connection and maxima for exit-completion time and crisis-response time. A feature that misses either minimum or exceeds either time limit loses its engagement mechanics and routes users to human support. Restart requires the well-being and safety owner to approve a corrective test and independent ethics review confirming recovery on every criterion.