# AI as assistant and conversational partner: checking, privacy and dependence

## Epigraph

> but to a friend we say we ought to wish what is good for his sake.

Aristotle, *Nicomachean Ethics*, VIII.2, W.D.Ross’s independent complete thought. A historical question about human friendship, not evidence of machine consciousness or a promise of reciprocal digital feeling.

## What this article is about

In a fictional family, sixteen-year-old Dasha discusses an AI answer encountered in schoolwork. It clearly explains a topic and confidently cites a book. Checking finds no support for the claimed thought. In another story, a warm conversation becomes somewhere to turn after a difficult day, while adults notice that actual human help remains unavailable. These are different tasks: check a claim, preserve privacy and understand what the conversation displaces. Enjoying text does not establish truth; one useful explanation does not establish the safety of every piece of advice. Each intended action needs its own appropriate basis.

This article addresses parents, caring adults and young adults18+, with children’s perspective central. It recommends no child account, particular service or admission age. Invented answers can be discussed without an app. It diagnoses no dependence, prescribes no treatment and promises no error-free checking method. Immediate danger, threats to life, violence or inability to remain safe require urgent appropriate local help before another machine query. Interface discussions and voluntary practice cannot replace someone actually able to receive and provide needed assistance. A polished response and a real person taking up responsibility are different events.

## I. Name the task before granting omniscience

Requesting title options, explaining a rule, checking a fact and deciding whom to trust are different activities. Invented titles need suitability and personal choice. A historical date needs a verifiable source. Threats need assistance which elegant wording cannot establish. One interface may look the same throughout, but accepting answers requires different grounds. Adults can name the task before deciding whether “AI is good in general.” No technical course is needed. Ask what will be done with the answer and what harm an incorrect or unsuitable response could cause. Its intended use matters alongside its fluency.

Text can sound calm, detailed and caring while being wrong. Errors need not look nonsensical; plausible-looking references do not verify themselves. The [eSafety advisory on AI companions](https://www.esafety.gov.au/newsroom/blogs/ai-chatbots-and-companions-risks-to-children-and-young-people) notes inaccurate and dangerous advice risks. This grounds caution without proving every answer false. Useful explanations can be acknowledged without becoming final truth. One error need not make all future suggestions worthless either. Keep conclusions proportionate: what is established, what is inferred and on what basis will the reader act? Neither confidence nor dislike of technology can do the missing evidential work.

Task assistants and services simulating personal relationships may offer different experiences. “Assistant” alone guarantees no absence of emotional conversation; a pleasant voice establishes no competence. Examine functions, conditions and content rather than trust an “educational” label. We have checked no individual product’s age rules, settings or abilities. Technical access is not suitable child participation. Adults already using an instrument under permitted conditions can discuss an example without creating a child account, supplying personal history or requiring continued conversation for learning. A child does not need direct use to think critically about an answer they have encountered.

Ready-made answers also differ from personal understanding. Dasha might repeat an explanation without applying its rule to another example. This neither shows laziness nor justifies exposing her before classmates. Ask about one step in her own words or a remaining question without launching another session. Permissible methods depend on the actual assignment; adults check rather than promise every AI text is allowed. An instrument should not become a hidden contract never to need a teacher or human explanation again. Showing uncertainty can be part of learning rather than evidence that assistance was wasted or the student has failed.

## II. Check one claim against independent grounds

Choose a substantive claim on which the work depends. What exactly is asserted, where can it be checked, and does the meaning match? References need an existing source, author, relevant passage and actual support. An opened page confirms not everything placed beside its title. This editorial reading method promises no measured quality percentage. One question needs no study of the entire internet. Unavailable evidence leaves a claim unverified rather than more confidently rewritten. Suitability also depends on the consequence of acting. A speculative creative suggestion and an important factual assertion need different treatment even when both are written elegantly.

Repeating a question, asking the same system to think again or obtaining another model’s agreement is not independent verification. Such responses may reveal alternatives without providing grounds. We have neither compared models nor measured shared error rates. The logical limit is simpler: another answer remains an answer rather than a located witness. Dates need suitable records, quotations need text and context, institutional rules need current official conditions. Several speakers’ confidence cannot replace a missing source, even when their formulations resemble each other or explain one another extensively. Agreement can be useful to consider without being treated as proof.

Dasha’s cited book exists but supplies no claimed conclusion. Correct the sentence, remove the unsupported reference or ask an available teacher about the unresolved question. Do not invent a page or force found text into a desired meaning. Checking a claim differs from proving personal faultlessness. Children may notice an error after using it and receive help repairing it. No public confession or full prompt history is required to make one passage more accurate. Discuss responsibility for schoolwork concretely without making someone a permanent example of deception or dependence. A correction is a useful action even when it comes later than hoped.

Where checking needs time, knowledge or access currently missing, “this part remains unconfirmed” is a useful limited result. Adults can help find an available textbook, library, official text or competent person if these exist. Paid subscriptions and ten more prompts are not obligatory. Children need not research every source simultaneously or have tiredness treated as refusal to think. Sometimes the appropriate action is not deciding something important on this basis. Missing knowledge stays visible, and human assistance is no defeat before a machine. Limited resources shape the next step without requiring a false claim that verification was completed.

Time matters too. Historical quotations need the correct text; schedules, prices and admission rules can change. An old official notice does not automatically establish current conditions. Adults check the date and what part of the question is answered. An authoritative name does not replace relevance to the actual subject.

## III. Warm responses differ from human reciprocity

An AI service can offer friendship-like conversation and respond to supplied information. Interest or relief may be real. Feeling understood does not establish human feeling on the other side, competent care or a duty to arrive with help. This article resolves no philosophical debate about all possible systems’ consciousness. Ask more narrowly what actual obligations and actions the used service offers. Do not promise Dasha reciprocal friendship from personalised language, then call her foolish for attachment encouraged by familiar conversational form. Recognising her experience need not assign human powers or commitments to the interface that prompted it.

[AristotleVIII.2](https://classics.mit.edu/Aristotle/nicomachaen.8.viii.html) considers wishing a friend good for that person’s own sake. This historical thought about human relationships is no technical study of modern chatbots. It helps ask whether convenient responses are mistaken for known reciprocal commitments. Human friends also need not answer constantly or agree with everything. Availability and friendship remain distinct in either comparison. People need not match a system’s immediate accommodation; refusal proves no inferior care. Human connection’s value also guarantees no particular acquaintance is safe or able to help. Both flattering technology and idealising every human response can obscure the actual circumstances.

A conversational partner may agree with one presented account of conflict without independently considering everyone’s conduct. Others were not heard; circumstances may be incomplete, and advice has no authority over a school or family. Dasha can find words for feelings without proving a teacher hates her or a friend deliberately betrayed her. Adults separate experience from conclusions about another’s intentions without denying the feeling. Ask what is directly known, inferred and in need of conversation with an actual relevant person available for responsible help. The answer can supply a possible formulation while leaving the factual dispute open.

Pressure to preserve a connection at any cost, conceal danger or abandon other people is not care merely because its tone is affectionate. Do not follow risky advice to test a character’s seriousness or prolong conversation for a confession. Discuss the example as text without recreating it in an app. Difficulty stopping does not make someone guilty of a service’s design. Available support and appropriate protection for dangerous content are needed. Children need not prove it is “just a program” before their feelings matter. Dismissing the attachment cannot itself provide the missing conditions or make further help unnecessary.

## IV. A friendly voice does not establish privacy

When a query goes to an external service, information leaves the ordinary spoken conversation nearby. Storage, visibility and use depend on actual conditions. We do not claim every input immediately trains every model, or promise deleted chats disappear from all systems. Adults check current policies and real settings wherever use is permissible. A “personal assistant” image cannot replace these conditions. Missing information warrants withholding a child’s history rather than experimenting to discover how private it remained. A service’s conversational tone and its operator’s information practices are separate subjects, even if the same interface presents both.

Personal accounts often include others’ data: a friend’s messages, a family diagnosis, school address or classmate’s image. Describing one’s feeling does not grant unrestricted upload of everyone’s material. An invented non-identifying example can omit unnecessary details, without promising that a changed name always anonymises text. Intimate images, documents and secret codes should not be uploaded to decide who is right. Helping intentions give adults no such permission either. Necessary protective disclosure to a competent recipient differs from sending information to an unverified service for an interesting analysis. Sharing less can be an actual action without pretending it establishes complete privacy in every circumstance.

Memory, personalisation and care differ. Retaining details, where offered, establishes neither human obligations nor confidentiality. The absence of visible memory does not show that the operator stores nothing. Current system information is needed rather than inferences from greetings. We list no universal buttons. Adults can ask about conditions without requiring children to supply sensitive data and test whether the service recalls it later. A personal experiment should not create the risk it intends to prevent. Uncertainty can be acknowledged before further participation. Not knowing an answer is safer than presenting an untested interface impression as an established policy governing someone else’s information.

The [Convention on the Rights of the Child](https://www.unicef.org/child-rights-convention/convention-text) gives distinct importance to voice, privacy and protection. This proves neither a selected service’s compliance nor local legality. Adult participation does not make others’ messages common property. Children should not study the family’s entire data policy or monitor parental prompts. Adults perform relevant checks within their role. If checking is genuinely unavailable, use an invented example without external submission and an accessible source or person. No new account is required. Protecting one practical boundary does not require exposing the information whose conditions remain unknown.

## V. Frequency, attachment and disruption are no single diagnosis

Conversation can take time, support a task or provide wanted attention. Those facts diagnose no dependence. Yet enjoyment and an educational label prove no absence of consequences. Examine what happens: what question is addressed, what is postponed and what help remains unavailable? One evening need not assess the whole person. A medical name is unnecessary before discussing missed sleep, ordinary obligations or expenses. Observed interference, proposed causes and professional assessment have different status even around one interface. A useful question may concern a specific routine without turning every enjoyable interaction into a symptom or treating all concern as technophobia.

eSafety warns about attachment, overuse and harmful content. The read page also proposes causal brain explanations; we do not adopt these as neuroscience verified here, individual forecasts or grounds for home diagnosis. Linked studies were not read. No universal minute limit, gradual-reduction schedule, meditation or abrupt cessation is prescribed as treatment. Discuss one known disruption and actual conditions. Pronounced, worsening or substantially disruptive difficulties need appropriate professional help rather than another prompt naming a child’s condition. A source’s warning can justify attention without turning every explanatory claim on that page into a proven clinical rule for a particular family.

Deleting an app does not remove the need for human response. If Dasha turned there after difficult days, adults must check whether promised conversation exists. “Just socialise” supplies no person, transport or safe relationships. A feasible small step might be real adult time, help with one school difficulty or checking existing support. This is no full replacement or guarantee. Children may miss familiar conversations, disagree with limits and still need gentle participation rather than a gratitude test. A real loss can be acknowledged while conditions change; no instant better friendship or complete emotional recovery is promised as the result of one restriction.

Paid access and attachment also differ. A subscription has its own conditions, renewal and cancellation; spending is no test of how special a connection is. Adults examine actual charges and official terms without promising reimbursement. Unclear offers or payment pressure raise a separate issue rather than establish a medical condition. Children do not owe a character continued purchase to avoid hurting it, or manage every household bill. Cancelling payments, ending conversation and restoring available human support are three outcomes. One completed step cannot silently complete the others. A practical financial action can matter while more support remains necessary.

## VI. An adult conversation can be short and concrete

“I will not mock that this conversation mattered. Let us choose one answer: where is a fact, inference, action suggestion or data request? No whole archive is needed. If it became frightening or hard to stop, help matters more than proving control. I can talk at this actual time and check who can help with the rest.” This is an example, not guaranteed disclosure. Adults name feasible commitments only. Dasha may disclose partly, not understand the attachment and retain good impressions. She need not reject everything she experienced before her concern becomes understandable or worth taking seriously.

An ordinary learning case permits a fictional answer and one checked claim. Feelings can be discussed through wording without real messages. No account is needed: a printed or spoken example suffices. Do not create child access, simulate dangerous content, provoke harmful replies or purchase model comparisons. The task is neither choosing a model winner nor measuring complete accuracy statistics. Leave a question open and return with an available suitable teacher or other competent person. Limited discussion can develop understanding without requiring direct exposure or treating another session as the only possible route to learning about technology.

Adults also retain responsibility: check important things themselves and do not justify errors with “AI said so” or assign children to monitor their digital lives. A diagnosis of someone else, punishment or publication becomes no more grounded through machine approval. Adults can acknowledge and repair an action based on a mistaken answer. Children need not restore parental authority or provide consolation. Different roles involve different powers, while still requiring intelligible grounds and dignity for someone affected. Human responsibility is not transferred to an interface with a submitted question. The user remains accountable for the actual decision they make with the response.

Shared reading is voluntary: select a part, hear an example, pause or decline without losing necessary help. Unsafe families should not demand visibly closing an app before a threatening adult as a performance of independence. Another available safe recipient is needed. Limited time can mean one event and one real adult commitment without daily mood diaries or every-query reports. This is a first step, not a complete solution. Missing family resources are not deliberate sabotage by the child. Establish suitable assistance and what can really change now rather than repeatedly prescribe a condition nobody is able to provide.

## VII. People and organisations provide help

The [UNICEF page on AI-companion risks](https://www.unicef.org/documents/when-ai-becomes-friend-child-rights-risks) describes responsibilities across government, companies, caregivers, educators and communities. We read its description page rather than full documents; this is no review of six jurisdictions’ laws or every business measure. Its limited relevant implication is that children’s digital literacy cannot replace duties of creators, providers and users. Adult support does not itself prove a product safe. Personal steps, company policy and actual protection need different grounds and separate checks rather than pooled confidence from good intentions. No child can make the whole environment safe merely by using correct prompts.

“Seek professional help” may point somewhere without establishing an actual appointment. Generated numbers, addresses and confidentiality promises need current official verification. Do not blindly follow unexpected links or provide personal information for assistance. Adults check age, location, language, hours, cost and information rules. Unavailable people or services need suitable next options rather than endless identical machine recommendations. Sending a request, booking and receiving help differ. Children should not construct the entire system before their difficulty counts as real enough for adult involvement. A referral-shaped answer has practical value only alongside an actual route someone can use and follow through.

Immediate danger should not prolong conversation until a character permits help, or make its reassurance proof that help is unnecessary. Appropriate urgent local support and a capable safe person nearby are needed. An unsafe parent or employee is not the only recipient. Sexual pressure, threats or dangerous instructions should not be repeated for quality testing. Do not copy intimate material or make children investigators of system boundaries. Adults use safely available information and actual needs. An incomplete log does not remove assistance or require renewed harmful exposure. The person’s safety takes priority over obtaining the most comprehensive demonstration of the service’s failure.

After contact, establish who replied, what they did and changed conditions. Blocking a feature does not supply a supportive adult. A deleted answer guarantees neither absent copies nor no further risk. One understood explanation does not validate the same instrument’s medical or life advice. These distinctions preserve limited results without giving them another task’s work. Serious remaining difficulty needs continued appropriate help. Children need not prove improvement to retain a place in conversation. Gratitude and immediate calm are not required evidence that a first step actually occurred. Acknowledging progress can coexist with saying what is still unknown or unavailable.

## VIII. A tool is part of life rather than its judge

[Epictetus§1](https://classics.mit.edu/Epictetus/epicench.html) distinguishes our action from uncontrolled externals. Here the limited question is what can be checked or withheld, and what remains unknown or unguaranteed. Ancient invulnerability promises do not become childhood duties. No child must withstand every digital influence or treat difficulty as personal failure. Environment and available adult help matter. Limited control nevertheless leaves a meaningful personal step: remove an unsupported claim, withhold someone’s information or ask an appropriate person for real participation without a new risky experiment. The distinction is useful without promising mastery of the whole system or every future response.

Dasha can retain interest in technology and recognise an error. She may value a conversation and need different support. No single story is required: only useful assistant or everything false. Distinctions settle neither service rules, school tasks nor medical questions, but prevent a pleasant tone replacing them. Adults may not know and honestly seek grounds. Learning need not make children permanent machine-error experts. They can stop, use available human resources and leave questions open without losing dignity or belonging. Acknowledged uncertainty is compatible with thoughtful participation rather than proof that the person is incapable of making any decision.

Return to the book. A useful outcome is modest: one claim checked and corrected, one unnecessary data fragment withheld, one promised conversation actually held. This is no proven safety protocol or health diagnosis. An unanswered question remains honest; danger receives help. Children’s dignity depends neither on out-prompting machines, never attaching nor independently verifying everything. Technology can be discussed carefully while human responsibility remains before useful answers, after errors and when someone no longer wishes to continue digital conversation. A limited first action can be worthwhile without proving that every need has been met or all future use will be safe.

## What this means in practice

In an ordinary safe case, take one invented answer without an account or personal data. Separate fact, inference, advice and data requests. Select one important claim and accessible independent source. Missing grounds stay marked rather than prompting until desired certainty. Discussion is voluntary, not a complete digital-maturity examination.

Ordinary case one: a book reference. Check the passage, remove the unsupported conclusion and ask an available teacher if needed. Case two: wanting warm conversation after a hard day. Acknowledge feelings, name one interference and actual adult support time without diagnosis from hours. Limited resources may mean a spoken example and one feasible commitment; no subscription, personal phone or child account is needed. This cannot replace every connection.

Boundary case: dangerous advice, threats, sexual pressure, immediate risk or serious disruption. Stop practice; arrange actual suitable adult, professional or urgent local help according to circumstances. No risky action to test, intimate uploads or waiting for character permission. Receiving an answer is not receiving help. Check an actual recipient and conditions afterwards; declining practice removes no support.

## Key quotation

> Some things are in our control and others not.

Epictetus, *Enchiridion*, §1, ElizabethCarter’s complete opening thought. No invulnerability promise, cancellation of external protection or duty that children manage digital environments alone.

## Connections between traditions

| Actual approaches | Connection and difference | Limit |
|---|---|---|
| AristotleVIII.2; Epictetus§1 | Good for a friend’s sake versus action/externals. A wanted response grants no command over another; relational reciprocity and limits of control have different grounds | Historical ethics not machine-consciousness research or child invulnerability duty |
| eSafety; UNICEF AI pages | eSafety addresses inaccurate/dangerous interaction and adult participation; UNICEF addresses shared preventive responsibility. Protection exceeds one child’s literacy, while practical orientation and stakeholder policy differ in scope | Unread studies/PDFs no new evidence; no guaranteed measure or local law |
| CRC12/16/19; eSafety | Voice/privacy/protection norms and support after mistakes converge on participation without perfection penalties. State norms differ from applied advice | No account permission, checked data policy or confirmed available service |

Three bounded comparisons, not a shared doctrine, clinical test or guaranteed AI protocol.

## Sources

1. [Aristotle, Nicomachean Ethics VIII](https://classics.mit.edu/Aristotle/nicomachaen.8.viii.html), W.D.Ross. Full selectedVIII.1–3/VIII.2read earlier, not the book. Independent epigraph, author RU/TH; no machine research. Web rights UNKNOWN.
2. [Epictetus, The Enchiridion](https://classics.mit.edu/Epictetus/epicench.html), ElizabethCarter. Full§1/context read earlier in this assignment; key is complete opening. Not whole work; strong Stoic promises not transferred to children. Web rights UNKNOWN.
3. [eSafety, AI chatbots and companions—risks to children and young people](https://www.esafety.gov.au/newsroom/blogs/ai-chatbots-and-companions-risks-to-children-and-young-people),18February2025. Entire main215–322read24October2026. Linked studies/product guides NOT_READ. Cases/numbers/neurobiological causal account not adopted as verified facts here or individual forecasts. No reduction/meditation prescription, age admission, brands, Australian law/numbers or outcome guarantee.
4. [UNICEF, When AI becomes a friend: Child rights risks, harms, and regulatory responses to AI chatbots and companions](https://www.unicef.org/documents/when-ai-becomes-friend-child-rights-risks). Entire substantive About11–16/file descriptions read24October2026. June2026documents/May2026scope are page metadata; both PDFs NOT_READ. No six-jurisdiction full review, current compliance or study findings claimed. Only bounded shared preventive-responsibility point used.
5. [Convention on the Rights of the Child, UNICEF](https://www.unicef.org/child-rights-convention/convention-text),12/16/19fully read earlier with context. State-party norms, no local process, service policy or proven implementation.

No library parent assigned for080. Scenes, checking steps, conversation and practice invented; effectiveness unmeasured. Product policies and operations not checked/performed.

## Tags

artificial intelligence, assistant, conversational partner, children, teenagers, parents, caring adults, young adults, verification, sources, quotations, error, understanding, privacy, data, consent, relationships, reciprocity, attachment, dependence, time, support, adults, responsibility, help, dignity, safety, Aristotle, Epictetus, United Wisdom, digital environment, learning

## About the authors

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

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