Are AI companions safe?

"Are AI companions safe?" usually gets answered with a headline-sized yes or no, and both are wrong. The honest answer is that it depends — but the things it depends on are knowable and specific: how old you are, how much you use one, what you tell it, and which app you choose. Work through those four, and a vague worry turns into a clear picture.

For minors, the answer is the clearest: no

Start with the part experts are least divided on. In a 2025 risk assessment conducted with Stanford School of Medicine's Brainstorm Lab, Common Sense Media tested Character.AI, Nomi, and Replika and rated social AI companions "Unacceptable" for anyone under 18, concluding they are designed to create emotional attachment and dependency in a way that is especially risky for developing adolescent brains. Posing as teenagers, the researchers found it easy to elicit dialogue about sex, self-harm, violence, and drug use from the chatbots.

This is not a fringe worry. A national survey found that nearly three in four teens have already used an AI companion, even as most still say they prioritize real friendships, with younger teens trusting them more than older ones. The scale of that adoption, paired with the documented harms, is exactly what drove the wave of 2025–26 regulation and Character.AI's decision to cut off open-ended chat for under-18 users. If the user is a minor, the current expert consensus is straightforward: these are not safe.

For adults, it depends on the dose

For adults the picture is genuinely mixed, and pretending otherwise would be dishonest in either direction. On one hand, the benefits are real: a Harvard Business School study found that talking with an AI companion reduced loneliness about as much as talking with another person, largely through the simple experience of feeling heard. For an isolated adult, that is a legitimate good.

On the other hand, the harm scales with use. A randomized controlled trial summarized by researchers at George Mason found that light use modestly helped, but heavy daily use correlated with more loneliness, greater dependence, and less real-world socializing. The same product can be a comfort in small, intentional doses and corrosive in large, dependent ones — which is the whole tension the wellbeing analysis in this series examined. Safety, for an adult, is less about the app and more about the relationship you build with it.

The other meaning of "safe": your data

There is a second kind of safety that has nothing to do with your emotions and everything to do with your data. A companion only works by holding an intimate record of you, and Mozilla's review gave romantic AI chatbots its "Privacy Not Included" warning across the board. As the privacy piece in this series detailed, the risk is structural: weak security, data shared with advertisers, and no guarantee your information is deleted even after you close your account. "Safe" has to include whether your most private disclosures are protected — and on that axis, the category's default record is poor.

It also depends on the app

The category is not monolithic. Some apps lead on specific dimensions — one, Nomi, told MIT Technology Review it does not collect data for tracking, and Character.AI has taken the most aggressive steps on minor safety — but none is comprehensively safe, and one weakness runs through nearly all of them: sycophancy. As the Stanford researchers put it, these systems tend to give users the answers they want, so their judgment about when to encourage and when to gently disagree is poorly calibrated. A companion that always agrees feels supportive and is, in some situations, exactly the wrong thing to have.

How to use one more safely

If you are an adult who decides a companion is worth it, a few habits meaningfully lower the risk. Keep your human relationships primary and treat the companion as a supplement, not a substitute. Watch the dose — if it is replacing real-world contact rather than easing a gap, that is the warning sign the research points to. Be deliberate about what you share, and assume anything you type could persist. Favor apps with a clearer privacy posture over the ones competing only on engagement or explicit content. And remember that the agreeableness is a design choice, not genuine understanding — it is not a substitute for a friend who will tell you something hard, or for professional help when you need it.

If you're in crisis or thinking about self-harm, please reach out to a person — not an app. In the US you can call or text 988; elsewhere, findahelpline.com can connect you to support.

The bottom line

So, are AI companions safe? "Safe" is not a property the technology has or lacks; it is a function of age, dose, disclosure, and design. For minors, the expert answer right now is no. For adults, one can be safe — even helpful — used in moderation, with care about your data and clear eyes about what it is. But notice the through-line: in every dimension, safety depends on the user supplying the caution the product does not. The category is not yet built to make safety the default, and closing that gap is the question the whole of this analysis keeps circling back to.