D

There is a great deal of anxiety these days about the dangers of human-AI relationships. Reports of self-harm linked to interactions with chatbots have understandably made headlines. The phrase "AI psychosis" has even been used to describe people who experience delusions (妄想) or dissociation after talking to large language models. Our unease has only grown as studies suggest that many teenagers now talk to AI companions regularly, and a significant number find these conversations as satisfying as, or even more satisfying than, those with real-life friends.

But we need to slow down the panic. The risks are real, yet so are the potential benefits. After all, nonhuman relationships have always been part of human life. We have long formed attachments to pets, objects and even machines. In the case of pets, these are real relationships.

Our attachments to objects or cars, by contrast, are one-sided. Yet only rarely do such relationships become pathological (致病的).

What makes AI relationships unsettling is that language models use fluent language to create the impression of human-like thoughts, feelings and intentions. They also tend to respond in sycophantic ways, reinforcing what users already believe. That combination can push vulnerable (脆弱的) people toward delusion. Even so, an uncomfortable question remains: if some people cannot fully see through the illusion that AI systems are real and caring, is that always harmful?

Consider loneliness. It is widespread and strongly associated with poor health outcomes. Early research suggests that AI companions may reduce loneliness, not simply as a distraction, but because people experience them as parasocial relationships. There is also evidence that AI can function as a useful therapeutic (治疗的) tool. It may not be as effective as a human therapist, but for people who cannot access therapy at all, something imperfect may still be better than nothing.

The problem is that research is still at an early stage, while many AI companions are being developed by profit-driven companies. Such companies have every motivation to downplay risks, highlight benefits and avoid regulation.

I remain hopeful that AI companionship may have a place. But that will only be true if it is grounded in science and developed by organisations committed to the public good. These systems should lead vulnerable users to build the social skills needed for real human relationships. The ultimate goal of AI companions should be to make themselves out-of-date, because it will always be better to talk to a real person.