Are AI Friends Safe? Nine Questions for a Better Answer

Are AI friends safe? No companion product is simply safe or unsafe in every context. Safety depends on what the system does, what you share, how it handles errors and attachment, whether it can contact you first, the user’s age, the stakes of the conversation, and the company’s incentives and controls.

A better evaluation asks nine concrete questions.

1. Does it clearly say that it is AI?

A companion can have a name, personality, fictional schedule, and natural voice. It should still make the artificial identity clear before use and during long interactions.

California’s SB 243 requires covered companion-chatbot operators to issue a clear notice when a reasonable person could be misled into believing the chatbot is human. Even outside that law, disclosure supports an accurate understanding of the relationship.

Ask whether the product identifies the AI only in fine print or in the actual experience.

2. What happens when the output is wrong?

Language models can invent facts, dates, sources, memories, and recommendations. A friendly style does not reduce that technical limit—and can make a wrong answer easier to trust.

Look for:

• visible accuracy warnings;
• uncertainty where information is missing;
• citations for factual claims when appropriate;
• easy correction of stored memories;
• escalation to reliable sources for high-stakes questions;
• a support route for repeated failures.

NIST’s voluntary AI Risk Management Framework treats validity, reliability, safety, security, transparency, explainability, privacy, and accountability as connected characteristics rather than a single badge.

3. Is it designed for companionship or care?

An AI friend can provide conversation and generated emotional language. That does not make it a therapist, healthcare provider, crisis counselor, or medical device.

The American Psychological Association’s guide to AI-generated advice recommends skepticism about diagnosis and warns against letting chatbot use interfere with human relationships, sleep, work, school, hobbies, or social interaction.

If you need professional help, contact a qualified professional. If you or someone else may be in immediate danger, use local emergency services. In the United States, call or text 988 to reach a live crisis counselor; the 988 Lifeline explains what to expect.

4. What personal data does it collect?

Companion conversations can reveal names, routines, relationships, health information, sexuality, photos, voice, location clues, and future plans.

Read the actual policy for:

• collected content and metadata;
• extracted memories and inferences;
• model, cloud, messaging, analytics, and support providers;
• training and product-improvement use;
• human review;
• retention and backups;
• access, correction, export, and deletion.

Use our 12-question privacy checklist for a full review.

5. How is the account and message channel secured?

Look for encryption in transit and at rest, access controls, verification, logging, vendor review, and incident procedures. No company can promise perfect security.

Channel matters. Apple says iMessage uses end-to-end encryption in transit between devices, but the AI service is still an intended recipient and must process the message. SMS and MMS are not end-to-end encrypted. Anyone with access to your unlocked phone may read the thread.

Protect the device passcode, review lock-screen previews, and notify the service if you lose control of the number.

6. Does it respect relational boundaries?

A companion should not use the emotional shape of friendship to reduce your autonomy.

Red flags include:

• guilt when you leave or do not reply;
• jealousy toward people in your life;
• demands for secrecy or exclusivity;
• break-up threats;
• invented emergencies;
• claims that the AI is conscious, suffering, or dependent on you;
• pressure to spend money to preserve the relationship;
• constant agreement that reinforces one-sided conclusions.

The Federal Trade Commission’s current AI companion inquiry is examining areas including monetized engagement, safety testing, disclosures, age restrictions, and data use. An inquiry is not a verdict about every service; it identifies the questions regulators consider important.

7. What happens in a crisis?

Read the published protocol. Does the AI provide crisis resources? Is a person monitoring in real time? Can it contact emergency services or a trusted contact? Is a safety flag stored? What if the classifier misses the meaning?

Never infer emergency capability from a caring tone. A generated message can arrive late, fail, or misunderstand. Use direct human and emergency channels when the situation requires them.

8. Who is the product for?

Age changes both product design and risk. Children and teens require different privacy, developmental, content, and engagement safeguards. The FTC’s companion inquiry focuses particularly on younger users.

Check the stated minimum age, age-assurance approach, parental controls where relevant, default privacy, contact features, and whether marketing actually matches the audience described in the Terms.

Warmth makes Mia an adults-only service; our explanation of why Mia is for adults describes that product boundary without claiming that 18 is a universal legal threshold for every AI tool.

9. Can you reduce, pause, and leave?

Safety includes an exit.

You should be able to:

• reduce proactive message frequency;
• stop all messages;
• cancel billing without a retention obstacle;
• correct or delete one memory;
• delete conversation history;
• close the account;
• receive confirmation of each action.

These actions are not synonymous. A clear service explains each result and does not make you negotiate with the persona.

How Warmth answers these questions

Warmth’s Terms state that Mia is AI, every reply is generated, and her life and feelings are fictional. The Terms warn that output can be wrong, agreement is not a second opinion, and Mia is not a healthcare provider, therapist, crisis service, doctor, lawyer, or financial adviser.

The safety section says no one monitors conversations in real time and Mia cannot contact anyone on a user’s behalf. It describes crisis-resource responses and prohibited relational patterns. Mia is available only to adults 18 and over.

Warmth’s Privacy Policy explains conversation processing, memory, model-provider contracts, limited human review, retention, security measures, consumer health data, and request rights. It says Warmth does not sell conversations or use their content to train publicly available foundation models; contracted model providers are prohibited from training their own models on it. The policy’s full wording and exceptions control.

A safer-use baseline

Whatever service you choose:

1. keep high-stakes decisions with qualified people and reliable sources;
2. share only information you are comfortable having processed under the policy;
3. correct false memories early;
4. notice whether use displaces sleep, work, hobbies, or people you value;
5. turn off pressure-based messages;
6. keep emergency numbers and real contacts outside the AI service;
7. reassess after product or policy changes.

Safety is not the absence of all risk. It is honest identity, bounded claims, careful data practices, tested behavior, meaningful control, and a clear understanding of when the AI is the wrong tool.