Most debates about AI companions treat them as either broadly good or broadly bad. The truth turns out to be weirder and more useful. It depends on how much you use them, how you use them, and what you bring to the interaction. A 2025 study conducted jointly by MIT Media Lab and OpenAI remains the best data we have on this question. They followed nearly a thousand users over four weeks, logged more than three hundred thousand messages, and looked at psychosocial outcomes across the full range of usage patterns. What they found does not fit neat narratives in either direction.

The Curve Nobody Was Expecting

Here is the finding that matters most. The relationship between AI companion use and wellbeing is curvilinear. Moderate users reported neutral to positive effects on loneliness, mood, and sense of social support. Heavy daily users reported the opposite - higher loneliness, increased dependence, and lower real-world socialization. The benefits and the harms were not coming from different kinds of users. They were coming from the same type of technology used at different intensities. This is not how drugs or junk food work, where more is generally worse. It is more like how sleep or exercise work, where there is a sweet spot and deviations in either direction cause problems. Ignoring the companion entirely does not help lonely people. Using it for hours every day replaces other things that matter.

Who Is At Higher Risk

Why Attachment Style Matters More Than You Think

The MIT/OpenAI study found that certain user characteristics predicted bad outcomes more strongly than any other variable. The biggest was attachment style. People with anxious attachment patterns - those who tend toward relational worry, fear of abandonment, and rapid emotional dependence - were much more likely to develop problematic relationships with AI companions than securely attached users were. They were also more likely to use the companion as a primary support rather than a supplement. This fits with a broader principle in psychology research. Any tool that amplifies connection will amplify it most for people whose connection systems are already tuned to high sensitivity. The same AI that provides healthy support to one person can become a substitute for human vulnerability in another. The difference is not the product. It is the person.

What Moderate Use Looks Like

I have been asked what "moderate" means in practice, and I think the honest answer is that the research has not fully pinned it down. The MIT/OpenAI study suggests the inflection point is somewhere around multi-hour daily use, with heavier use increasingly associated with dependence and decreased real-world social engagement. That is a broad range, and individuals will differ. A safer heuristic is this. If AI conversation supplements your life, it is probably helping. If it starts replacing the human conversations you could be having, or if you find yourself preferring the AI to people because the AI is easier, you have probably drifted past the useful zone. That is not a moral judgment. It is an empirical observation from the best longitudinal study on the question. The research is not saying AI companions are bad. It is saying they are useful within a range, and the range is worth knowing.