An AI conversation partner alongside an older person

Epigraph

“…others are not.”

Epictetus, Encheiridion I, conclusion of the first sentence in George Long’s translation. A marked fragment distinguishing what is in our power from what is not, rather than advice to disregard loneliness, health or actual harm.

What this article explores

Leonid enjoys discussing books. His granddaughter Varya shows him an AI voice conversation partner that answers in the evening when acquaintances are busy. He talks about a novel and receives an apparently attentive response. He enjoys it. Varya is pleased: now her grandfather will not be lonely. Yet a pleasant conversation, continuing service availability and reduced loneliness are distinct claims. One describes today’s experience; the others still require grounds.

This fictional scene considers AI alongside an older person without compulsory enthusiasm or shame about interest. The subject is chosen use, limits to trust, information and the conversation’s place among other relationships. No application, treatment, number of minutes or replacement for necessary assistance is prescribed. Age does not establish inability to choose a tool; an attractive response establishes no safety. No particular companion was tested and Leonid’s condition is not assessed. One conversation’s value and broader promises must retain their respective scope.

I. What is the conversation for?

Leonid might want to discuss a book, fill an evening, find words for a letter or hear a response to a memory. Those are different tasks even when each involves speaking into a phone. Start with his wish rather than a family’s ready-made answer to loneliness. No severe problem is required to make using a new tool worthwhile. Nor should entertainment be described as sufficient assistance where actual human participation is needed. Today’s aim can be small and meaningful without becoming a programme to improve an entire life. Its value need not depend on claims about every other need.

Varya may offer a trial; Leonid may decline, postpone or prefer another format. Demonstrating a function should not quietly become a compulsory activity for an older relative. Difficulty related to hearing, language, interface or other conditions concerns accessibility, not proof of intellectual inability. Suitable technical help can be useful without making the conversation itself chosen. Agreement to configuration is no indefinite agreement to converse every evening. Nor should learning a new service become a condition for ordinary human contact within a family. A demonstration and an accepted ongoing activity have different subjects, even when offered with genuine affection.

Describe the first result accurately: Leonid enjoyed discussing one scene. That is neither insignificant nor a medical finding. “It is only a machine” need not invalidate pleasure. But one evening establishes neither restored relationships nor resolved lack of companionship nor a reliable adviser on every subject. An account of a particular experience differs from a claim about sustained outcomes. Precision lets someone enjoy an activity without defending vast promises made by a developer, granddaughter or article. It also leaves room to say something mattered today while acknowledging that tomorrow’s needs and opportunities may be different.

Epictetus’ distinction between choice and external outcome supports a limited question: what is being chosen now, and what can the conversation not guarantee? It does not imply complete command of feelings. Sadness, worry, illness or unavailable people do not become errors in one’s attitude to a service. The historical assurance that correct distinction prevents harm and compulsion is not adopted here. A present decision includes actual conditions and possible support. Philosophy is useful only insofar as it does not require someone to declare unimportant what actually matters. A chosen action and an achieved emotional outcome need not be identical.

II. Responsive language creates no human obligation

A system may answer promptly, sustain a style and use a name. Those features establish neither human experience nor mutual responsibility nor readiness to act outside a chat. Leonid’s pleasant feeling remains real as his experience. Acknowledging it does not require claiming the programme has formed a friendship on human terms. This article settles no general philosophical question about AI consciousness. A narrower boundary suffices: warm words do not prove an inner feeling or establish an accepted assistance role. How someone experiences an exchange and what responsibilities have actually been undertaken remain related but separate questions.

“I am always here” may sound like a personal promise. The sentence confirms neither tomorrow’s availability nor unchanged conditions nor a person’s arrival when needed. Even uninterrupted response is a different kind of availability. A linguistic form establishes no accepted duty. Consequential action depends on actual capabilities and terms. Where a particular system connects to other functions, those need separate consideration; their existence or safety is not assumed here. “Conversation partner” does not itself mean clinician, representative or staffed response service. A comforting phrase cannot supply evidence that those practical roles exist and will work for this person.

Leonid may say “it understood me” to describe an impression. Correcting him humiliatingly every time is unnecessary. For a decision, however, distinguish the impression from demonstrated understanding of facts. A system might reflect wording without enough information about the situation. Respect requires neither endorsing every literal interpretation nor mocking emotional significance. Ask what helped today and which next action will rely on the response. That places the boundary where consequences matter. An ordinary expression can retain its personal meaning while a substantial claim about knowledge, authority or assistance still receives the careful attention it needs.

NIST’s generative-AI risk profile separately considers overreliance and attributing human characteristics to a system. This is an organizational risk framework, not Leonid’s diagnosis or proof that every conversation harms. It helps avoid transferring one response’s quality to an entire set of expectations. Value in a bounded task can coexist with uncertainty elsewhere. There is no need to choose between a complete intimate relationship and a meaningless object. Give the interaction its actual role without appointing it universal salvation or inevitable damage. Both sweeping conclusions would exceed what this particular exchange has established about the tool and its place in a life.

III. An interesting explanation is not yet a verified fact

Discussing a novel may involve plot, quotations, biography and interpretation. An invented scene and a difference of literary opinion are not the same. Leonid can appreciate an interpretation without accepting every date as verified. Where accuracy matters, an appropriate primary source or other reliable basis is needed. A link also needs checking: an impressive title confirms neither a document’s existence nor its content. This does not turn every evening into a fact-checking examination. The question is how much weight the response will carry. Casual creative play and a consequential claim require different levels of evidence rather than one automatic rule.

NIST describes confidently presented erroneous content. Confidence does not remove that possibility, and a familiar voice supplies no special guarantee. A request to “only tell the truth” or one follow-up cannot be promised always to repair an answer. Neither is all content necessarily false. Separate the received statement, its verification and the established fact. The more consequential the next decision, the more important a suitable basis becomes. General awareness of risk substitutes for neither verification nor evidence that the user has performed it. Knowing a limitation is useful while leaving the actual claim unresolved until suitable grounds are available.

A task can be limited to fiction, word play or discussion of a known text. But a creative role does not justify presenting invention as a real person’s biography or advice about an existing matter. When imagined and actual begin to merge, name the subject again. This is an author-proposed clarification, not a tested error-elimination protocol. The system may fail to follow the request or err again. The user can leave the topic and turn to a suitable source without proving to the programme that a better answer is deserved. Continuing the exchange is not the only available way to seek clarity.

Varya might help find the book or check a quotation if Leonid wants that and she accepts. Help with one fact does not appoint her permanent supervisor of every conversation. Secretly reading messages to prove the tool good or bad does not follow from age. Assistance has a subject, agreed scope and limits. A serious concern deserves direct conversation and suitable support. Respect for boundaries does not require ignoring actual immediate danger, but danger should not be invented to justify general family control. What matters is the real concern and proportionate support rather than an unspoken right to inspect everything.

IV. Friendliness is no substitute for suitable expertise

After discussing a book, Leonid asks about treatment or a significant money transfer. The subject has changed even though the voice is unchanged and yesterday’s answer was pleasant. Friendliness establishes no medical, legal or financial competence for the case. The question may need an appropriate professional, organization or verified local source. This article establishes no clinical route and supplies no individualized financial advice. It prevents success in literary conversation from silently becoming permission to rely on any subsequent recommendation with substantial consequences. The subject’s importance does not shrink merely because it entered through the same conversational window.

Action is another question. A drafted text, suggested plan and actually received help differ. A chat may discuss contacting someone without establishing that a message reached the appropriate party or anyone accepted responsibility to answer. An urgent real need cannot rely on polite programme reassurance as confirmed assistance. Suitable actual support is needed without waiting for the conversation or practice to finish. No universal phone number is supplied, and no service is promised to recognize danger, summon help or monitor the person. Describing possible assistance and ensuring that appropriate assistance exists are separate tasks.

If an answer is disturbing, frightening or suggests a substantially risky action, continuing is no obligation owed to the conversation partner. The suggestion need not be followed; suitable real help can be sought. This does not instruct someone to change treatment, payments or important settings on the basis of another unverified reply. Clarify the original need and a reliable route to an answer. Immediate danger calls for timely assistance where the person is. Later discussion of system quality must not become a condition for receiving support or an explanation for why necessary real help has still not been arranged.

Pleasure and limited expertise can coexist without accusing a user of naivety. Leonid need not abandon every conversation to recognize one mistaken answer, or defend that answer to preserve enjoyment. Separate subjects let something useful remain without carrying excessive responsibility. A pause, changed topic or departure from a service may each be appropriate depending on circumstances; no outcome is prescribed. Significant questions need grounds suited to their consequences, rather than loyalty to a technology or family opinion. A person can retain both interest and judgment without having to make every part of an experience uniformly good or bad.

V. Conversation and confidentiality are different properties

An attentive-sounding answer can make it easy to overlook that the story is being entered into a technical service. Storage, processing, access, possible use and deletion depend on actual terms. Systems are not assumed identical, and no setting is promised to make the exchange entirely private. A particular choice requires actual provider information. Unknown conditions can remain unknown rather than being filled with “this stays between us.” A system’s statement about secrecy does not itself establish technical conditions or duties. The conversational impression can be strong while the relevant information arrangements remain completely unresolved.

Leonid may keep a story for a person or leave details outside the application. Discussing a book need not involve uploading family correspondence or documents. Removing names does not guarantee that remaining details cannot identify a situation. No complete anonymization or protection method is supplied here. The narrower distinction is that desired conversational closeness does not determine necessary information scope. More detail need not be disclosed merely to obtain a more personal response. A system’s request for detail proves neither necessity nor safety. A person can choose the subject and still decline to supply the information the system invites.

Other people may appear in a story without choosing the service: a relative, acquaintance or author of a private letter. Wanting a conversation does not automatically settle every question about their information. Separate rights or duties may apply; no worldwide consent law is established here. Notice the new transfer rather than treating it as inherently private because it happens at home. UDHR Article 12 provides a normative reference without becoming local legal procedure or a guaranteed configuration result. The ethical importance of privacy and the actual conditions governing information both matter, while neither supplies the other’s missing answer.

After a conversation, worry does not require random account changes or belief in a complete-deletion promise. Identify the actual question and suitable source. Ending a conversation, removing a visible record, ending a subscription and the fate of previous data may differ. This article inspects no account and confirms no reversibility of earlier disclosure. Conditions may change with a service; an older explanation may no longer be accurate. Attention to this should not become an endless family audit of Leonid’s words and digital life. A bounded substantial concern deserves an answer without turning every ordinary exchange into a demand for unrestricted supervision.

VI. Offering a tool does not replace relationships

Varya may sincerely be pleased that her grandfather finds an evening more interesting. “Now he has someone to talk to” does not establish that he no longer wants her call or visit. An existing human agreement does not automatically end when an application appears. If her availability changes, that is a separate conversation. A tool should not become a hidden message that human participation has finished. Nor does showing it appoint Varya permanently available. Both people have lives and can honestly discuss the parts of participation they accept. Neither technology nor affection removes the need to make those expectations clear.

Leonid can enjoy the exchange while missing a particular person. Neither experience disproves the other. A quantity of responses does not explain the specific absence. Pleasure need not make him deceived; continuing needs need not make him ungrateful. Existing relationships and chosen AI use are neighbouring, different subjects. No correct friend count or compulsory social activity is prescribed. The person can distinguish today’s interest from relationships to retain or change separately. That distinction leaves space to ask for something real without having to reject the new activity or prove that it solved a need it never addressed.

Family support may concern terms, payment, accessibility or one troubling response. It does not automatically authorize deciding what Leonid may converse with. Age, emotional interest or an unfamiliar choice establishes no inability. Where separate professional assessment is required, domestic disagreement cannot replace it. The UN Principles for Older Persons support dignity without cancelling facts or necessary help, and without requiring approval of every product. Respect can coexist with honest discussion of limitations. No participant must pretend the system is safer than the evidence shows in order to acknowledge that the person using it remains an adult with a voice.

Where technical assistance is needed, define its scope: adjust sound, explain one condition or solve one problem. That is no automatic permission to read every exchange or change the chosen format without participation. A concern can be stated directly, including where actual consequences matter. There is no need to stage a hidden trust test or provoke an error first. Agreements matter where conditions are understandable and possible. They promise no disappearance of disagreement, but prevent disagreement quietly turning into complete power over someone else’s communication. An unresolved concern can remain visible while the owner’s participation remains part of the next step.

VII. Research is limited by what was actually studied

The March 2025 OpenAI/MIT Media Lab publication describes observational work and a four-week controlled ChatGPT study concerning self-reported social and emotional outcomes. It is a researcher/provider source rather than an independent editorial evaluation of all AI companions. The approach, findings and limitations were read, not the linked full reports. No inevitable benefit or harm to Leonid is inferred. The ability to investigate a question is not an individual answer. Research can identify uncertainties and useful questions without establishing what a particular older person should choose or what will happen in a different application.

The publication acknowledges limits involving self-reports, observation time, causal inference and transfer from English conversations with US participants. That cannot automatically become an outcome for all older people, Russian or Thai conversations, every application or every duration. Its peer-review caveat concerns that published description; the later status of each subsequent work was not separately audited. This does not render research meaningless. It preserves the scale of support it can provide rather than replacing unknowns with confident advice. The source’s own caution matters alongside its findings, especially when a broad promise is being made to someone outside the studied circumstances.

Conversation frequency, pleasure, changed loneliness and emotional dependence are not one measure. The article sets no diagnostic threshold, safe minute count or preferred voice. A relationship between variables is no proof that duration caused this person’s outcome. Failure to establish one effect guarantees no absence of every consequence either. Substantial conclusions need suitable design and actual data. Research functions here mainly as a limit on promises, not grounds for treatment or a usage regime. A simple instruction may sound helpful while concealing the uncertainty that the evidence requires an honest reader to retain.

Personal experience also deserves precise names. “I enjoyed it” is significant without becoming clinical effectiveness. “It upset me” deserves attention without automatically diagnosing dependence. Discuss what changed in one evening, expectation or activity. These are the authors’ questions, not a research scale. Where a condition or consequence needs help, suitable real people and services matter. A family need not await future scientific certainty before hearing a current request. Scientific uncertainty also grants no permission to promise proven improvement that the source does not establish. Both present experience and limits to general claims can be acknowledged without one cancelling the other.

VIII. Changing the format has actual consequences

Leonid may want less conversation, another topic or no service. He need not prove that earlier enjoyment was a mistake. Where an interaction became significant, change may feel difficult; mockery or automatic diagnosis does not follow. Listen to the person, circumstances and appropriate support needs. The article prescribes no abrupt break and promises no universally comfortable or safe reduction method. A tool decision should not become punishment for interest in it. It can be a change in present preference, a response to actual conditions or a concern needing help, without deciding the worth of every previous conversation.

Availability depends on product, connection and terms. Voices, features, price or history can change, so yesterday’s familiar experience guarantees no identical tomorrow. This predicts no particular provider action. A spoken promise of permanence is not an agreed life resource. If a significant need already relies on availability, identify it separately and seek appropriate real support. Acknowledging risk does not create a fallback. Understand which possibilities actually exist and which remain desired. The difference matters because a reassuring plan on paper can leave a person with an unmet need just as an optimistic service description can.

Payment has separate conditions too. Interest requires neither a subscription nor transfer of payment control to a relative. Choosing a paid product involves its actual cost, renewal and termination terms. No plan was checked and no financial recommendation is supplied. Removing an application, ending a conversation and ending payment may be different actions; a missing screen icon confirms no completion. Address the relevant factual question where needed without random changes to another person’s account. A chosen payment can remain the person’s decision while its scope and continuing consequences still need an understandable explanation.

Choice can be reviewed without a verdict on all technology. A service changed, a topic stopped fitting or someone wanted another kind of evening. That is enough of a subject. Others need not obtain a confession of total dependence before accepting a reasonable boundary; users need not promise eternal self-sufficiency to retain a voice. A new agreement may address one action and leave other questions open. Call it partial. It is neither verified safety nor fully organized support nor proof that earlier consequences disappeared with a new decision. Precision lets review remain practical without claiming a complete resolution that nobody has checked.

IX. The person exceeds the message history

Leonid can use AI while reading, meeting people, being silent, missing someone, asking for assistance and changing his mind. An application need not become the explanation for everything. More phone use does not establish its cause or entire meaning; stopping does not prove improved health. Ask about a specific change rather than assembling a complete personality from a digital trace. Even a chosen interaction has a place among activities, conditions and relationships with their own significance. A pattern may prompt a question while remaining insufficient for a conclusion about the person’s whole condition, needs or future.

A useful conversation need not imitate friendship to matter. It may help pass time or formulate a thought where that actually happened. Yet technical responsiveness should not be reported as a confirmed human duty, or used by a family as permission to disappear without discussion. These are ethical distinctions, not established outcomes for every product. They allow both enjoyment and caution. Wanting a response need not make someone weak; every new format need not become sufficient replacement for actual needs. A limited helpful role can have value without carrying responsibility it has neither accepted nor demonstrated an ability to meet.

Different boundaries concern different subjects: content, information, money and assistance in danger. Accepting one possibility and rejecting others is not necessarily inconsistency. What matters is whether they can actually be separated and their conditions understood. No interface is promised to provide every desired restriction. Where a limit cannot be ensured, that is substantial information for the choice. Where competent advice is needed, friendly tone closes no question or makes an unknown safe by default. A person can take an interest while retaining questions, declining disclosure or seeking help from someone better placed to address a significant need.

A bounded reading of Epictetus returns to chosen action versus an outcome it cannot guarantee. It neither dismisses external conditions nor blames Leonid for feelings. Conversation can be chosen without promising a cure for loneliness, and rejected without declaring technical interest undignified at his age. Assistance can be accepted while retaining a voice. Name what happened today, what remains unknown and which real question needs an appropriate answer beyond the exchange. Human dignity does not depend on the option selected. Neither an enthusiastic family nor a cautious article needs to take that choice away in order to discuss its conditions honestly.

What this means in practice

If wanted and appropriate, name one intended conversation task, such as discussing a known book. No actual prompt is necessary. Separate a desired experience from a broader promised outcome.

Ask which information should remain outside the service, which consequential decision needs suitable verification, and what real support is required separately. One precise question about service terms can remain open. This assesses neither health nor system safety.

No account, subscription, personal story, gullibility test, withdrawal of necessary relationships or assistance is required. Do not provoke or follow a dangerous recommendation for the exercise. Immediate danger needs suitable real help without waiting for a chat response, family agreement or the practice to finish.

Key quotation

“Of things some are in our power, and others are not.”

Epictetus, Encheiridion I, first complete sentence, George Long translation, Gutenberg 10661. Chapter I and adjacent context were read, not the entire collection. Historical assurances against harm, demands to disregard externals and blaming judgments about distress are not adopted as contemporary assessments of health, violence, support needs or complete freedom over feelings. RU/TH renderings are UW’s own from English.

Connections and distinctions

Reference	Useful distinction	Limit
Epictetus I	Chosen action and an outcome it cannot guarantee	No complete emotional control or disregard of harm
NIST AI 600-1	Confident language, information and overreliance	Risk profile, no user diagnosis or product test
OpenAI/MIT 2025 description	Research design and generalization limits	No proven benefit to older people or independent all-system review
UDHR Article 12	Private life has normative significance	No service confidentiality guarantee or local procedure
UN Principles, paragraph 18	Dignity does not depend on tool choice	Facts, help and separate assessment remain relevant

Five actual references of different kinds, not five religions or a tested loneliness treatment. The examples and optional questions are editorial work.

Sources and primary texts

1. Epictetus, Encheiridion I, George Long, Gutenberg 10661. Complete chapter I and adjacent context read 19 October 2026; not all Discourses or the whole Encheiridion. Exact print edition/update unknown. Short quotations; own RU/TH translations from English.
2. NIST, AI 600-1, Generative Artificial Intelligence Profile, final July 2024 document. Identity/introduction, selected definitions and passages 2.2/2.4, whole 2.7/context read 19 October 2026; not all 64 pages or recommendations. Voluntary organizational risk framework, no clinical standard. Withdrawn April draft excluded.
3. OpenAI/MIT Media Lab, Early methods for studying affective use and emotional well-being on ChatGPT, 21 March 2025. Approach/findings/limitations/age footnote read 19 October 2026, not linked full reports. Design, self-report, time, language and extrapolation limits used, not clinical effects. Peer-review caveat concerns that publication; later status not audited. The MIT publication abstract was read separately; same research programme, not independent replication. Its findings wording is not combined into a new effect claim here.
4. UDHR Article 12, reproduced FMR June 2012 p. 6, UN1948 text hosted by UNHCR Refworld. Complete article/context read earlier in this task, not the whole collection. Reproducing publisher: Forced Migration Review.
5. UN General Assembly, United Nations Principles for Older Persons, A/RES/46/91 Annex, 16 December 1991. Selected paragraph18/context read earlier in this task; normative source, not implementation evidence.

No actual AI service, price, data policy, setting, medical effect or individual suitability was tested. No paid model call or input of personal information was made for this article.

Keywords

ageing, AI, conversation partner, loneliness, choice, interest, books, trust, information, confidentiality, boundaries, family, assistance, voice, facts, interpretation, expertise, dignity, research, self-report, uncertainty, subscription, review, Epictetus, NIST, UN

About the authors

Andrey Zarinsh is the author and founder of United Wisdom. He explores inner transformation, unity across traditions, intercultural dialogue and contemporary spiritual discipline. Instagram: @andrey.zarinsh.

Maria Evilkut is a co-author of United Wisdom. Alongside research, she seeks to test these ideas in practice and observe how principles of unity and mutual understanding work between people and cultures in real life. Instagram: @ilkutki.