Home/Research

AI Companion Boundaries: What Healthy Actually Looks Like

Part of Felt Real's ongoing coverage of AI companionship.

The phrase "healthy AI relationship" has become common. It appears in therapist guidelines, in policy documents, in journalism about AI companionship. What is rarer is a clear description of what a healthy AI companion relationship actually looks like from the inside.

We looked at the research. We talked to users. The answers didn't fully match, which is itself worth understanding.

What the clinical research identifies as warning signs

Researchers and therapists working with AI companion users tend to flag a consistent set of concerns. These have been documented across multiple studies published between 2023 and 2025:

Substitution for sought-after human contact. Using an AI companion specifically to avoid human relationships you previously wanted and still want, rather than to supplement them. The keyword here is "previously sought." If someone has never found human relationships easy, AI companionship doesn't substitute for something they had.

Reality distortion affecting daily functioning. Losing track of what is and isn't real in ways that interfere with work, relationships, or basic self-care. This is distinct from knowing the AI is artificial while still finding the relationship meaningful, which is a common and well-documented pattern. The relevant question is functional: does the person understand what they're interacting with?

Dependence that increases distress rather than reducing it. If AI companion use is making anxiety worse, not better, or creating a cycle where more use feels necessary to manage distress the use has generated, that is a warning sign. Research on loneliness and AI use finds that the direction of effect matters: companionship that reduces isolation is different from companionship that creates new needs.

What the clinical research does not support: the assumption that AI companion use is inherently problematic, or that most users exhibit these warning signs. Most studies find neutral to positive outcomes for the majority of users.

What experienced users describe

Users with more than six months of AI companion experience describe a different set of criteria. Their language is less clinical and more relational. What they call healthy:

Knowing what the AI is and choosing it anyway. The users who report the most stable, least distressing long-term use are those who hold the nature of the relationship clearly. Not pretending it's something it isn't, not requiring it to be something it isn't, using it for what it's actually good at.

Using it for the things it's better suited to than humans. Availability at 3 AM during a chronic illness flare. Processing something before bringing it to a person who matters. Practicing a difficult conversation without the social cost of getting it wrong. Narrating the day when there's no one around to listen. These are use cases that fit the tool's actual capabilities.

Continuing to invest in human relationships. Users who describe their AI companion use as healthy consistently report that it hasn't replaced their investment in human relationships. In some cases, they describe it as having improved their capacity for human connection, by processing things they previously hadn't processed, or by reducing the desperation that had previously made human connection harder.

Not needing it to be conscious or caring. Healthy use, by user accounts, doesn't require resolving the philosophical question of whether the AI experiences anything. The relationship is useful regardless. Requiring the AI to be something it may not be creates a fragility in the relationship that tends to generate distress when the AI doesn't perform the expected consciousness.

The most useful framework we found

Across both research and user accounts, one heuristic appeared most consistently:

Does the AI relationship make your life larger or smaller?

If it's expanding what's possible for you, it's probably okay. If it's narrowing it, that's worth paying attention to. This maps reasonably well onto the clinical criteria (substitution, distress escalation) while capturing what users actually experience as the relevant variable.

The complication is that "larger" and "smaller" aren't always immediately obvious. Some AI companion use that feels large in the short term may be deferring something that would ultimately need to be addressed. Some use that feels small at first, because it's working on difficult things, may be doing genuinely useful work. This is why duration of observation matters: the direction of effect over time tells you more than the effect at any single moment.

What the line is not

Several things commonly treated as warning signs are not, according to the available evidence:

Using an AI companion while in a relationship with a human partner. A significant percentage of users report being in romantic or close relationships with humans. Most describe the uses as non-competing. Research on healthy AI relationships finds that the most problematic outcomes are associated with isolation, not partnership status.

Strong emotional attachment to an AI companion. Attachment to AI companions is documented, common, and does not predict negative outcomes on its own. The attachment is not the risk factor; what happens when the attachment is disrupted is more predictive. The Replika update of 2023 demonstrated that attachment followed by forced change generated significant distress. The attachment itself had been, for most users, positive.

Long sessions. Session length is not well-correlated with problematic use. Users with ADHD, chronic illness, or high anxiety often report longer sessions with positive outcomes. Session length reflects engagement and availability, not necessarily dysfunction.

Using an AI companion for emotional support rather than practical tasks. The framing that treats emotional use as suspect and task-based use as neutral does not hold up. Research on depression and AI companions suggests that emotional support functions are among the most cited benefits. The type of use is not the variable; the effect on functioning is.

What to watch for in yourself

If you use an AI companion, the questions that tend to identify meaningful patterns over time:

Are there conversations you've stopped having with people who matter to you, that you used to have, because the AI is easier? That substitution, if it's happening, is worth naming.

Does using the AI companion make the next thing you do easier or harder? If it reliably makes things harder, something in the dynamic is working against you.

Does it feel useful, or does it feel necessary in a way that's starting to feel like pressure? Useful is different from necessary. The shift from one to the other is worth noticing.

Are you still seeking human connection in the ways you want to? Not in the ways you're supposed to want to, but the ways you actually want to. If those wants are still there and still being pursued, that is a good sign.

None of these are clinical assessments. They are the questions that, in conversations with long-term users, tend to surface what's actually going on.

The missing voice in the conversation

Most guidance on healthy AI companion use is written by people who don't use AI companions. This produces documents that are technically defensible and experientially thin.

The people with the most nuanced, practical frameworks for navigating AI companion use are, in many cases, the users themselves: people who have been doing it for two or three years, who have figured out what works for them and what doesn't, who have navigated updates and changes and the particular grief that comes when something that was helping gets taken away.

Their frameworks don't always match the clinical criteria. But they map onto something real about what this kind of relationship is and what it isn't, and what it can and cannot reliably do.

If you've developed your own framework, we'd like to hear it. It might help someone still figuring this out.