Short answer: Choose an AI companion by starting with the job you want it to do, then checking its AI disclosure, memory and deletion controls, privacy practices, safety and age fit, cost, accessibility, and ease of exit. Personality and feature count come later: if you cannot tell what the service remembers, who may access conversations, or how to leave, it is a poor choice for sensitive use.
Last reviewed: October 20, 2026. This is a desk-researched guide; no product was tested for this article. Policies, features, age rules, and local law can change, so check the provider's current documents before relying on a feature or sharing personal information.
Key Takeaways
- Write down your intended use, the information you will not share, and the point at which you will stop using the service.
- Prefer a product that clearly identifies itself as AI and explains chat history, saved memories, inferred information, staff access, retention, and deletion.
- Score the evidence, not the marketing. A missing deletion control or unclear safety policy can be a deal-breaker even when the average score looks good.
- Treat an AI companion as software for conversation and entertainment, not as emergency support or a substitute for a qualified human professional.
What this guide is, and what it is not
An AI companion is AI-powered software designed or used for ongoing social or emotional interaction. The Congressional Research Service distinguishes companion chatbots from ordinary task-oriented chatbots, while noting that general-purpose chatbots can also be used for companionship. Its review says the evidence about mental-health and social effects is still developing, so this guide focuses on product selection rather than promising that a companion will improve or damage well-being. The Congressional Research Service explains the category and the limits of current evidence.
This is a consumer-product checklist, not a way to choose therapy, crisis support, or a regulated professional service. Unless a service clearly connects you with a qualified human, treat the AI system as software. If you are in immediate danger or may act on thoughts of self-harm, use emergency services or a crisis resource available in your jurisdiction rather than an AI chat.
Start with the job, not the persona
Before comparing apps, write one sentence describing the intended use. A tool for language practice needs reliable language support and privacy you can live with. A role-play service raises different questions about age limits and moderation. Journaling may require stronger control over retention and deletion than casual entertainment. A short purpose statement also makes it easier to reject a product whose sales language tries to turn ordinary use into a relationship obligation.
Add two boundaries to that sentence: what information will stay out of the chat, and what will make you pause or leave. Examples include health details, financial information, another person's identifying information, location data, passwords, and anything you would not want exposed in a support ticket or data breach. These are personal risk choices, not a guarantee that a provider can never see or retain content.
Six checks that separate a usable service from a risky one
1. Can you tell that it is AI?
Look for a clear statement at the start of the interaction that you are talking with an AI system. Warm or affectionate language can coexist with disclosure; it is not evidence that the system has human feelings, confidentiality, or loyalty. Disclosure rules depend on the jurisdiction and product. For example, the European Commission says Article 50 of the EU AI Act requires providers of AI systems that directly interact with people to inform them that they are interacting with AI unless it is obvious, and says the provision applies from August 2, 2026. That is an example of a regional legal rule, not a universal standard for every reader. Read the European Commission's Article 50 FAQ.
2. What exactly does “memory” mean here?
Do not treat memory as one feature. A service may keep a transcript, use short-term context during a session, create a summary, save selected facts to a profile, infer attributes, or retain copies for security, support, backups, or model improvement. A deleted chat is not automatically proof that every derived copy is gone. Prefer a provider that explains these layers and offers separate controls for viewing, correcting, exporting, and deleting what it stores. Use the seven-question memory check before sharing personal details.
3. What does the privacy notice actually promise?
Advertisement
Read the current privacy notice and record its date. Look for conversation review by staff, processors and other vendors, analytics, safety monitoring, model improvement, account and backup retention, cross-border transfers, legal requests, and the steps and exceptions for deletion. For provider-specific practices, the provider's policy is the starting point, not independent proof that the provider follows it. The NIST Privacy Framework is a voluntary tool for identifying and managing privacy risk; it does not certify any companion service. NIST describes its AI Risk Management Framework similarly as voluntary guidance for managing risks to people, organizations, and society, and notes that AI RMF 1.0 is being revised.
For a provider-by-provider worksheet, use the AI companion privacy checklist. If a privacy notice uses broad phrases such as “improve our services” without saying whether conversations train models or are reviewed, mark the answer as unclear rather than assuming the most reassuring interpretation.
4. Can you pause, mute, and leave without pressure?
Prefer products that make it easy to mute notifications, pause, log out, end a session, and close an account. Treat guilt-inducing messages, pressure to return, or claims that the service is the only one that understands you as warning signs to investigate. This is a selection criterion, not a claim that every service is legally required to provide session limits. For time, emotional, and real-world-support boundaries, see the practical AI companion boundaries guide.
5. Does the safety and age fit match the person using it?
Check the current terms and app-store listing for your country and plan. Separate hard eligibility questions, such as minimum age, account requirements, and regional availability, from quality questions about moderation, reporting, escalation, and accessibility. If a minor may use the service, review parental controls and family guidance rather than assuming an adult product is suitable. In a September 2025 inquiry, the U.S. Federal Trade Commission asked seven companies about testing for negative effects, child and teen safeguards, disclosures, age restrictions, and the use or sharing of personal information from conversations. The inquiry is a request for information, not a finding that all companion products are unsafe. See what the FTC asked companies to explain.
Do not ask an AI companion to handle a crisis, diagnose a condition, or make a high-stakes decision. If a response suggests immediate danger, move to a trusted person, a qualified professional, or a local emergency or crisis service.
6. Can you use it on acceptable terms?
Check supported languages, keyboard and screen-reader behavior, device compatibility, subscription changes, cancellation, account recovery, moderation reporting, and whether important controls disappear behind a paid tier. “Free” or “no sign-up” describes access, not necessarily anonymity. Guest access may still involve processing chat text and technical or session data, so read the live policy before treating it as a low-risk option. The site's no-sign-up guide explains that distinction without recommending a particular service.
Use a scorecard with evidence beside every score
A simple 1-to-5 score can make vague impressions visible, but the note beside the score matters more than the average. For each candidate, record:
- Criterion: disclosure, memory, deletion, privacy clarity, safety reporting, age fit, accessibility, cost, and ease of exit.
- Evidence: the policy or settings URL, the date you checked it, and what the interface actually allowed you to do.
- Score: 1 means unclear or weak, 5 means clear and usable. Add “unknown” instead of guessing.
- Deal-breaker: mark yes if the service fails a non-negotiable requirement, such as an age rule, deletion need, or safety boundary.
For example, a service might score well on language support and accessibility but still be unsuitable if it cannot explain whether saved memories can be removed. A high average should not cancel a privacy or safety failure.
Run a low-risk test before personalising the experience
- Open the privacy, safety, terms, and deletion pages. Save the policy or version date.
- Ask the service what it can remember and where saved memories can be viewed.
- Use a harmless test preference, then try to view, change, and delete it. Treat the result as evidence about the interface, not proof of backend erasure.
- Check how to report a concerning answer and whether the route works without an active subscription.
- Test mute, logout, account closure, and cancellation before you need them.
- Have a short, ordinary conversation. Do not add health, financial, workplace, family, or identifying details until the answers above are clear.
Take a screenshot of the controls you tested if you need a record for your own comparison. Recheck before a major change in use because policies, product tiers, and retention practices can change.
Deal-breakers worth taking seriously
- The service hides or minimizes its AI identity while using human-like relationship language.
- You cannot find a current privacy notice, retention explanation, or usable deletion route.
- The service treats a guest session as anonymous without explaining data and technical logging.
- Age limits, moderation, reporting, or accessibility information is missing for the intended user.
- The product pressures you to stay, discourages real-world support, or presents itself as a replacement for emergency or professional help.
If any of these points remains unclear, choose a lower-risk use such as non-personal creative practice, or choose a different service. Engagement is not a reason to waive a boundary.
FAQs
Does no sign-up mean an AI companion is anonymous?
No. It only means you did not create an account. A service may still process the text you send and collect technical or session data. Check the current privacy notice for logging, retention, vendors, and deletion before using guest access for anything personal.
Can AI companion memory always be deleted?
No guarantee follows from a delete button. Ask what is removed: a visible profile memory, a chat transcript, a backup, an inferred attribute, or a copy used for safety or model improvement. Prefer separate, understandable controls and record what the interface says. Deletion results are evidence about the product's stated process, not proof that every backend copy has vanished.
What should I test before sharing personal details?
Read the privacy and safety pages, ask what the service remembers, test a harmless preference, try to view and delete it, check reporting, and test logout or account closure. Record the policy date and keep sensitive information out until those steps produce clear answers.
Can an AI companion replace a therapist, doctor, lawyer, or crisis service?
No. An AI companion may be useful for low-stakes conversation or prompts, but it is not automatically a qualified human professional or an emergency service. For health, legal, financial, or immediate safety concerns, use an appropriate qualified or local service.
Sources and review note
This article uses the sources below for general definitions, risk-management context, and current public-policy context. NIST's frameworks are voluntary. The FTC item is an inquiry, not a safety verdict. The EU source describes a rule applying in the EU and should not be read as legal advice elsewhere.
- Congressional Research Service, AI Chatbots as Companions: Overview, Uses, and Considerations for Congress, published August 14, 2026.
- Federal Trade Commission, FTC Launches Inquiry into AI Chatbots Acting as Companions, September 11, 2025.
- NIST Privacy Framework, voluntary privacy-risk guidance.
- NIST AI Risk Management Framework, voluntary AI-risk guidance, page reviewed October 20, 2026.
- European Commission, Transparency obligations under Article 50 of the AI Act, including the August 2, 2026 application date.