464,687 Messages

For most of human history, companionship required another human being. Inconvenient, really.

They had to be awake. Available. Interested. And occasionally willing to listen to the same problem for the seventh time.

AI has removed several of those requirements.

It does not get tired. It does not have somewhere else to be. It does not suddenly remember that it promised to call its mother. You can arrive at 2:13 in the morning with a paragraph about your ex, your career, your attachment style or the meaning of your life and receive an answer almost immediately.

We have spent years talking about artificial intelligence as a productivity tool. Increasingly, people are using it as somewhere to go.

“AI has removed friction from intimacy.”

That makes one recent dataset particularly interesting. Researchers analysed 464,687 messages across 4,664 AI-companion chat sessions contributed by 237 participants, alongside wider survey data. Nearly half a million messages between humans and machines—not weather forecasts, not customer-service questions, but conversations.

One finding stood out: people with smaller offline social networks were more likely to report using AI primarily for companionship. Companionship-focused use was also associated with lower wellbeing, particularly when conversations were more intensive and highly disclosive.

That does not mean AI caused the lower wellbeing. The direction may run the other way. People who already feel lonely, isolated or unable to access satisfying human relationships may simply be more likely to seek connection somewhere that is always available.

And that might be the more important observation. AI has removed friction from intimacy.

You do not have to wonder whether you are oversharing. You do not have to be interesting. You do not have to wait until tomorrow. The machine can remember what you told it, ask a follow-up question and remain, at least structurally, astonishingly patient.

For someone struggling to access human connection, that can be genuinely appealing. Sometimes it may be useful. A conversation can help someone organise a thought, rehearse what they want to say to another person or feel less alone in a difficult hour.

But it also creates something humans have never really had before: a relationship-like interaction that can be almost infinitely accommodating.

Real relationships are not. Friendship requires compromise. Partners misunderstand us. People get distracted. Someone eventually has to leave for dinner. Even the person who loves you most will occasionally be in a bad mood when you would prefer them not to be.

We usually describe those inconveniences as problems. They may also be part of what makes a relationship real. Another person has an interior life that does not belong to you. That is the friction. It is also the point.

AI companionship does not necessarily replace human relationships. That is too simple, and the current evidence does not justify it. But as systems become better at remembering us, mirroring us and responding in emotionally fluent language, they introduce a new comparison into ordinary life.

Human connection versus connection on demand. One is messy, reciprocal and occasionally disappointing. The other can be available whenever you open the screen.

The biggest question, then, is not whether humans are capable of feeling attached to machines. We already know people can form attachments to voices, fictional characters, objects, celebrities they have never met and versions of people that exist mostly in their own imagination. Human beings are extremely talented at attachment.

The more interesting question is what happens when the easiest relationship in your life is the one that asks almost nothing from you.

Does it become a bridge back to other people? A rehearsal space? A diary that answers? Or, for some people, does the absence of friction make ordinary human relationships feel even harder by comparison? We do not know yet.

But 464,687 messages suggest we should stop treating the question as science fiction. The machine is already listening. The next question is what we want that listening to be for.