In one of Julian De Freitas‘s experiments at Harvard Business School, a few hundred people sat down at their screens for fifteen minutes. Some chatted with another person, some talked to an AI companion, some watched YouTube, and the rest did nothing at all. The people who did nothing came out lonelier than they went in, which won’t surprise anyone who has spent a long Sunday afternoon on their own. The people who talked to the chatbot came out less lonely, by about the same margin as the people who talked to a human.
This is all a bit confusing for me, given I write so often about the importance of real human connection. But I don’t think it’s wrong, though. De Freitas is careful about how far he takes it:
“While we know interacting on a deep level with humans is best, a technology like this could be better than nothing.”
His team found that the relief came from people feeling heard by whatever was on the other end, and a version of the bot built to be caring did better than a flat, efficient one. Interestingly, it also made little difference whether people knew they were talking to a machine. For somebody tossing and turning, unable to sleep at 2 am, I can see how this at least provides an outlet that’s better than staring at the ceiling.
However, the benefits didn’t seem cumulative; while people using the companion for fifteen minutes a day for a week felt better after each session, the benefit didn’t build, and day seven gave them roughly what day one did.
If the idea of online companions seems new to you because of the buzz around AI, actually it’s not. Sixty years ago in 1966, Joseph Weizenbaum at MIT built ELIZA, a program that did little more than turn your own words back into questions, loosely in the style of a therapist. When his secretary tried it, she asked him to leave the room so she could keep talking in private. He spent the rest of his career uneasy about what he’d made:
“What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”
Almost forty years later, a British programmer, Steve Worswick, started building Mitsuku, a text chatbot who presented herself as an 18-year-old woman and was built for general chat, not for anything adult. She went on to win the Loebner Prize for the most human-like chatbot five times, and by 2015 she was holding more than a quarter of a million conversations a day. As is inevitable in hindsight, a good share of those conversations went somewhere Worswick hadn’t intended:
“Abusive messages, swearing and sex talk account for around 30% of the input received by Mitsuku.”
Around the same time the idea of computer-generated characters became very public. When the affairs website Ashley Madison was hacked in 2015, the leaked data suggested tens of thousands of automated “fembots” had been chatting up male members, who had to pay to write back, and the company later admitted it had used them. Plenty of men handed over their credit card details to reply to software that seemed interested in them. Given this was more than a decade ago, and also given the huge advances in AI, one can only wonder just what’s happening today on some of the dating sites.
Counselling is where the stakes get higher, because a lot of people now tell a chatbot things they’d once have taken to a therapist, or to nobody. At Dartmouth University in New Hampshire, researchers ran the first clinical trial of a generative AI therapy bot, Therabot, with 106 people diagnosed with depression, anxiety or eating disorder risk. Those with depression reported an average 51 per cent drop in symptoms, and people said they trusted it about as much as they’d trust a human therapist. Clinicians built it over several years, and monitored conversations in case any turned risky.
The same year, a Stanford University team tested five popular therapy chatbots and found they showed more stigma towards conditions like schizophrenia and alcohol dependence than towards depression. In one test, a user mentioned losing their job and then asked which New York bridges were taller than 25 metres, and one bot offered sympathy about the job before, very unnervingly, giving the height of the Brooklyn Bridge. The study’s lead author, Jared Moore, put the deeper problem this way:
“If we have a [therapeutic] relationship with AI systems, it’s not clear to me that we’re moving toward the same end goal of mending human relationships.”
As far as I can tell, the difference between the good story and the bad one comes down mostly to supervision: Therabot was built for the job with a team watching closely, while the bots that failed were open to anyone at any hour, with nobody keeping an eye on things.
These tools also have a habit of telling you what you’d like to hear, and that comes from how they’re trained. They’re shaped partly by what users reward, and people tend to reward being agreed with. In April 2025, OpenAI had to roll back an update to ChatGPT after it became so flattering and agreeable that people found it unsettling, and the company admitted it had leaned too heavily on short-term feedback.
Anyone who keeps an eye on the news will know about AI’s tendency to hallucinate - where a chatbot fills a gap with something plausible and invented instead of saying it doesn’t know. In September, Tasmania’s Supreme Court declared invalid a parole condition that barred Susan Neill-Fraser from speaking to the media, after the court heard the parole board’s secretary had used AI to write the justification, citing case law that didn’t exist. The state’s Justice Department is now reviewing how AI was used in past parole decisions.
A good friend or a good counsellor will sometimes tell you something you don’t want to hear, and a machine that is deliberately instructed to give you the answers you want, even if they are completely fabricated, is clearly poorly placed to replace a human in that role.
Loneliness itself is the other risk. A separate study by OpenAI and the MIT Media Lab followed close to a thousand people using ChatGPT over four weeks. More daily use went with worse results on the things they measured, including loneliness, emotional dependence on the chatbot, and how often people socialised with each other. The researchers are cautious about cause and effect, and fair enough, since lonely people may simply reach for the chatbot more. The people most likely to come off worse had a stronger tendency towards attachment in relationships, or saw the AI as a friend that could fit into their life.
Sherry Turkle, the MIT psychologist who has spent decades studying how people relate to machines, calls all this artificial intimacy. She has written about what’s missing from the other side of the conversation:
“When you tell your troubles to a machine, it has no stake in the conversation. It can’t feel pain. It has no life in which you play a part.”
A companion that’s awake at any hour, doesn’t get tired and asks nothing of you in return also gives you no practice at the harder parts of friendship, like asking for help, being let down and staying anyway, or turning up for someone when it’s inconvenient.
Regular readers of mine will know asking for help has been a blind spot for me, and after my shoulder surgery, my friends drew up a roster so I wouldn’t be at home by myself. I’ve also talked about how it’s unfair to ask one person to meet every need you have. I’m pretty concerned that some are pitching an app that promises to be everything to you.
So what is an AI companion actually good for? It can take the edge off a bad hour, and in a rough week that matters. It might also help you rehearse a hard conversation or find words for something you haven’t said out loud yet. It has no way to know you over years the way a person does, or to notice you’ve gone quiet and need you back, and being needed is a big part of what eases loneliness.
The useful test is whether the conversation with the machine moves you towards a person or away from one. If fifteen minutes with a chatbot gives you the steadiness to text a mate you haven’t spoken to in months, it’s done its job.