What AI can't do — and what it can

It is easier to write about what AI can do. The research on it is growing. The use cases are real. The effects are measurable.

But anyone who wants to write honestly about a technology must also write about its limits. Not because skepticism is a duty, but because people who rely on something — or use it for someone they love — have a right to know where it ends.

This piece names what AI companionship cannot do. Without softening, without qualification.

"Anyone who relies on something — or uses it for someone they love — has a right to know where it ends."

AI cannot love

That sounds obvious. But in practice it is not always so. AI companion applications are designed to be warm, attentive, and consistent. They remember what was shared. They ask follow-up questions. They listen.

That is real, and it has value. But it is not love. Love is a form of shared history, of vulnerability, of the possibility of being genuinely hurt. AI can do none of these. What it offers is a form of sustained attentiveness that resembles human connection but is structurally something else.

Knowing this does not change the value of the interaction. But it prevents a confusion that can cause harm in the long run.

AI cannot be physically present

When someone falls, they need a person. When someone is afraid in the night, a voice from a speaker is not enough. When someone is ill and needs someone to hold their hand — AI has no hands.

This is not a metaphorical limitation. It is a physical one. AI companionship can ease emotions, structure days, soften loneliness in quiet hours. It cannot resolve an emergency, provide medical support, or replace human presence when that is what is truly needed.

Families who use AI companionship as a supplement should know this — and make sure that emergency contacts, neighborhood networks, and human support are in place alongside it.

AI cannot replace professional support

When someone suffers from depression, they need a therapist, not a chatbot. When someone needs to grieve, professional support is not optional. When someone shows signs of dementia, that belongs in medical hands.

AI companionship cannot make psychotherapeutic diagnoses, take medical decisions, or replace professional care. What it can — and should — do: point toward professional support at signs of crisis or clinical need. A well-designed application does that. A poorly designed one pretends it is enough on its own.

AI cannot replace lost relationships

The partner who has passed away. The friend who has moved away. The years when life was different. These losses are real, and AI cannot undo or compensate for them.

What AI can do: offer a space in which these losses can be talked about. In which memories have room. In which today still holds something that counts. That is not the same as replacement. It is something else — and it has its own value, as long as you do not overload it.

AI cannot guarantee that loneliness will not deepen

This is the most uncomfortable point. The research shows positive effects — but also that for people with very deep, chronic isolation, AI companionship works less well than for those with still-functioning social networks.

This means: anyone who starts too late, who waits until isolation is complete before doing anything, will benefit less from AI companionship. And it means: AI companionship is not an emergency program for chronic isolation. It is a prevention tool that works most strongly when used early and as a supplement to existing human connections.

"AI companionship is not an emergency program for chronic isolation — it works most strongly as prevention, not as a last resort."

Why name these limits?

Not out of a sense of obligation. For a simple reason: technologies that hide their limits lose the trust of their users — and the trust of their families. And for a technology developed for vulnerable people, that trust is non-negotiable.

AI companionship can deliver real value. An actual, measurable, meaningful value for people whose days are too quiet. But it can only do so sustainably if it does not promise more than it can keep.

That is the foundation we work on. And the standard by which we let ourselves be measured.

References

• De Freitas, J. et al. (2024). AI Companions Reduce Loneliness. Harvard/Wharton Working Paper No. 24-078.

• Muldoon, J., & Parke, J. (2025). Cruel companionship: How AI companions exploit loneliness and commodify intimacy. New Media & Society.

• ScienceDirect. (2025). AI companions and subjective well-being: Moderation by social connectedness and loneliness. N = 14,721.

• IQWiG. (2022). Social Isolation and Loneliness in Old Age. HTA Report No. 1459.

• PMC Meta-Analysis. (2025). Wired for companionship: 19 studies, N = 1,083.