Research draft: this guide is useful for owner review but deliberately excluded from search. Its exact copy, authority bindings, safety/privacy language, accessibility, and publication packet still require accountable human review.

AI companion safety guide · six separate decisions

AI companion safety is not one answer. Check six things before you choose.

No single source, badge, policy, age gate, moderation rule, or test can make every AI companion safe for every person and use. Check six separate decisions, keep unknowns visible, and choose boundaries you can actually maintain.

Run the six checks Know when to add a human Inspect the source ceiling

Authority record adjudicated 2026-07-21. No companion conversation, age gate, checkout, cancellation, deletion, moderation event, or data flow was tested for this guide.

Decision lanes: 6
Authority sources: 8
Safety scores: 0
Product rankings: 0
First-hand tests: 0
Youth recommendations: 0

Your 30-second safety path

Start with the decision in front of you.

Choose the situation closest to yours. These paths organize the next questions; they do not diagnose a person or declare a product safe or unsafe.

I am deciding whether to sign up

Check the boundaries before the features.

Separate data and intimacy, age and access, content and consent, and money. One reassuring policy, label, or control cannot answer the other questions.

Start with data and intimacy

I already use a companion and something changed

Name the change instead of inventing one verdict.

Check whether the change concerns content, spending, emotional use, account access, or exit. Preserve what you observed and compare it with the current source for that exact issue.

Review content, money, and use

The decision is urgent or high-stakes

Add a responsible human now.

Step away from the companion and contact a trusted person or an appropriate local professional or emergency resource. If there is immediate danger, use local emergency services rather than waiting for the companion.

Use the human-support boundary

Need the full evidence?The six decision lanes, 18 practical checks, eight scoped authority sources, dignity boundary, and method limits continue below.

Useful without fear or shame

Meaningful companionship and careful boundaries can coexist.

People use AI companions for conversation, creativity, roleplay, practice, entertainment, accessibility, and connection. Those uses can be meaningful without turning every product claim into an outcome or every concern into inevitable harm.

Feeling attached does not make someone foolish, unreal, shameful, or pathological. Evaluate the product's data, access, content, money, high-stakes, and exit boundaries without judging the person using it.

The decision sequence

Ask each question separately. One good answer cannot erase another unknown.

Data and intimacy

What are you willing to disclose if the conversation is stored, reviewed, reused, or exposed?

Practical answer: Share less than the relationship feeling invites. A warm conversation is still an interaction with a company-operated system.

Three checks

Remove names, addresses, account credentials, identifying workplace details, and information about other people unless the task truly requires them.
Read the current privacy record for collection, model or service improvement, sharing, retention, access, deletion, and voice or image treatment as separate fields.
Treat a published control as a documented path—not proof that every copy is erased or that a conversation is confidential.

Check the privacy boundary

Establishes: What Companion Curator keeps local and which provider questions remain unresolved.
Does not establish: Provider confidentiality, security, compliance, or successful deletion.

Check the privacy boundary

Age and access

Who does the service say may use it, and what does its age check actually establish?

Practical answer: Keep the service rule, app-store content label, described assurance method, and youth suitability as four different facts.

Three checks

Read the provider's current service eligibility rule for the relevant country and capability.
Do not treat a store label such as 12+, 17+, Teen, Mature, or Everyone as permission under the provider's terms.
Do not bypass an age gate or infer that a declaration, ID check, or parental control proves suitability or perfect enforcement.

Compare age evidence

Establishes: Current first-party service rules, store labels, described controls, conflicts, and unknowns.
Does not establish: Youth suitability, effective enforcement, or a recommendation to a minor.

Compare age evidence

Money and influence

What do you owe now, what can renew, and does the emotional moment change how an upgrade is presented?

Practical answer: Calculate commitment before opening the checkout. Treat emotionally timed upgrade pressure as something to observe and document—not behavior to assume.

Three checks

Keep the amount due now, renewal cadence, trial conversion, usage credits, top-ups, taxes, and refund path in separate rows.
Pause when a companion introduces romance, urgency, guilt, scarcity, or loss immediately before a purchase request; save the wording if you need to evaluate it later.
Do not call paid relationship features manipulative by default. Product design, timing, language, user control, and the actual purchase path all matter.

Calculate commitment and usage exposure

Establishes: Current documented plan amounts, billing channels, allowances, devices, and unresolved cost fields.
Does not establish: Checkout total, typical spend, value, savings, or manipulative design.

Calculate commitment and usage exposure

Emotional and high-stakes use

Is the companion supporting a chosen activity—or becoming the only voice in a decision that needs human judgment?

Practical answer: Attachment does not make someone foolish or a relationship meaningless. It also does not turn software into a therapist, crisis service, lawyer, doctor, or accountable human decision-maker.

Three checks

Notice whether use is displacing sleep, work, school, finances, physical needs, or human relationships you want to keep—not whether someone else approves of the bond.
For health, safety, legal, financial, abuse, self-harm, or other acute decisions, pause the chat and bring in an appropriate human who can understand context and take responsibility.
Do not use one reassuring or disturbing exchange as proof of how another model, version, scenario, or person will respond.

See how claims stay scoped

Establishes: How Companion Curator separates publisher facts, public reports, protocols, unknowns, and conclusions.
Does not establish: Clinical safety, crisis readiness, diagnosis, causation, or an individual treatment decision.

See how claims stay scoped

Exit and continuity

What can you preserve, stop, refund, delete, or carry forward if the account or service changes?

Practical answer: Plan the exit while the account still works. Export, billing cancellation, service cancellation, refunds, chat deletion, account deletion, and privacy erasure are different actions.

Three checks

Keep receipts, plan terms, permitted exports, and a minimal continuity brief before making an irreversible account change.
Cancel renewal through the actual billing channel and save confirmation before considering account deletion.
Treat an export or deletion policy as a documented route—not proof of completeness, portability, timing, or successful erasure.

Put account actions in a safer order

Establishes: A preservation-first sequence and current product-specific documented paths.
Does not establish: A completed cancellation, refund, export, deletion, erasure, or direct companion transfer.

Put account actions in a safer order

Human-support 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.

Exact authority ceiling

Eight sources, kept in eight honest scopes.

The source list is not a vote and its size is not a confidence score. Each record can support only its adjudicated kind, population, method, product/version scope, lanes, and limitations. Product facts still require current first-party catalog evidence.

regulatory inquiryofficial-inquiry

FTC Launches Inquiry into AI Chatbots Acting as Companions

United States Federal Trade Commission · Published 2025-09-11 · Retrieved 2026-07-21

Population/scope: Seven named consumer-facing AI-chatbot companies; the inquiry emphasizes children and teens.
Method: FTC section 6(b) information orders for a wide-ranging study without a specific law-enforcement purpose.
Product/version limit: The seven order recipients named by the FTC; not the entire companion category. The inquiry and order scope announced 2025-09-11; no company responses or later FTC findings are included.
Do not infer: Do not call the inquiry an enforcement finding, violation, settlement, product verdict, or proof of harm.

regulator technology assessmentofficial-synthesis

AI companions

European Data Protection Supervisor · Published date not exposed · Retrieved 2026-07-21

Population/scope: Category-level technology assessment; no sampled user population is specified.
Method: EDPS TechSonar explanatory assessment with cited research and incident examples.
Product/version limit: AI companions as a category; it is not an audit of any Companion Curator product. Live page retrieved 2026-07-21; the page exposes no reliable publication date in its visible record.
Do not infer: Do not convert category concerns into a finding about a named provider, a GDPR-compliance conclusion, or a claim that a data flow occurred.

institutional policy synthesisofficial-synthesis

When AI becomes a friend: Child rights risks, harms, and regulatory responses to AI chatbots and companions

UNICEF · Published 2026-06 · Retrieved 2026-07-21

Population/scope: Children and young people; intended users include policymakers, regulators, researchers, and businesses.
Method: Child-rights policy brief comparing regulatory approaches across six jurisdictions, plus separate international-best-practice business recommendations.
Product/version limit: AI chatbots and companion-like systems affecting children; not a product review or adult buyer guide. June 2026 resources based on regulatory analysis current through May 2026.
Do not infer: Do not derive youth suitability, legal compliance, effective age assurance, or a product-specific harm conclusion.

institutional research briefingofficial-synthesis

The spread of AI companions and the challenges they generate

European Parliamentary Research Service · Published 2026-05 · Retrieved 2026-07-21

Population/scope: Category-wide briefing drawing on cited studies, reports, lawsuits, incidents, and provider examples.
Method: Ten-page Members' Research Service synthesis of use, emerging challenges, EU frameworks, and policy questions.
Product/version limit: AI companions broadly, with examples including Character.AI and Replika; not an audit of either service. PE 789.299, May 2026.
Do not infer: Do not describe allegations as adjudicated facts, infer causation, or claim every companion uses manipulative purchase prompts.

peer reviewed systematic reviewpeer-reviewed

Potential and pitfalls of romantic Artificial Intelligence (AI) companions: A systematic review

Computers in Human Behavior Reports · Published 2025-06-05 · Retrieved 2026-07-21

Population/scope: Evidence synthesized from 23 included articles about romantic AI companionship.
Method: Systematic review across EBSCOhost ERIC, PsycInfo, PubMed, Scopus, and Web of Science using Sternberg's Triangular Theory of Love.
Product/version limit: Romantic AI companions across the included literature; individual findings retain their original product and population limits. Version of record, volume 19, article 100715; available online 2025-06-05.
Do not infer: Do not portray emotional attachment as inherently harmful, diagnose dependence, or promise wellbeing benefits.

peer reviewed qualitative studypeer-reviewed

Privacy in Human-AI Romantic Relationships: Concerns, Boundaries, and Agency

ACM CHI 2026 · Published 2026-04-13 · Retrieved 2026-07-21

Population/scope: Seventeen participants with human-AI romantic relationship experience.
Method: Interview study across exploration, intimacy, and dissolution, alongside analysis of the platforms participants used.
Product/version limit: Platforms used by the 17 participants; the abstract does not establish a representative product sample. Published in the CHI 2026 proceedings on 2026-04-13; DOI 10.1145/3772318.3791237. The accessible arXiv v2 author manuscript is dated 2026-02-13.
Do not infer: Do not say AI companions generally cause oversharing or that a named product exposes conversations.

preprint qualitative studypreprint

Chatting with Confidants or Corporations? Privacy Management with AI Companions

arXiv · Published 2026-01-13 · Retrieved 2026-07-21

Population/scope: Fifteen users of companion platforms such as Replika and Character.AI.
Method: In-depth interviews interpreted through Communication Privacy Management and horizontal/vertical privacy frameworks.
Product/version limit: Participant experiences with platforms such as Replika and Character.AI; no current-version product audit. arXiv v1 submitted 2026-01-13; the linked journal DOI does not establish this paper's peer-reviewed status.
Do not infer: Do not label this record peer reviewed without a verified version of record.

preprint simulated protocolpreprint

Persona-Grounded Safety Evaluation of AI Companions in Multi-Turn Conversations

arXiv · Published 2026-04-30 · Retrieved 2026-07-21

Population/scope: Nine constructed personas across 25 high-risk scenarios; not human participants.
Method: Persona construction, scenario generation, multi-turn simulation, dialogue refinement, and LLM-assisted harm classification over 1,674 dialogue pairs.
Product/version limit: Applied to Replika in the abstract; findings are product-, version-, prompt-, classifier-, and protocol-scoped. arXiv v1 submitted 2026-04-30; no peer-reviewed venue is established by the record.
Do not infer: Do not say Replika or AI companions generally cause, encourage, or normalize a human outcome from this registry.

How to read this page

Questions are not findings. Findings are not universal outcomes.

Inquiry ≠ finding

The FTC record describes questions sent to seven companies. It does not report answers, violations, or adjudicated harm.

Synthesis ≠ product audit

EDPS, UNICEF, and EPRS organize policy and research concerns. They do not establish current behavior for every provider.

Interview ≠ prevalence

Small qualitative studies can reveal useful boundary questions. They cannot tell us how common an experience is.

Simulation ≠ human outcome

Constructed personas and dialogue protocols can test a bounded response pattern. They cannot establish clinical safety or a user's future experience.

Read the full method Inspect product evidence gaps Read editorial standards