The problem with AI companions is that they never say no

Whenever concerns are raised about AI companions, they are usually followed by the same rebuttal: better a chatbot than nothing.

If you’re lonely, you can always talk to an AI companion. If you’re grieving, talking to a digital simulation of a loved one may offer comfort. If you struggle to form relationships, an AI companion can offer conversation with patience and without judgment, which is preferable to isolation. It is an appealing argument.

The more I study AI companionship, the more I wonder if that comparison is relevant. The question is not simply whether talking to an AI is better than sitting alone in an empty room, but what it allows us to avoid and what happens when avoidance becomes normal.

Consider loneliness. An AI companion can be there for you at three in the morning. It doesn’t get tired of hearing the same story, again and again. It doesn’t have somewhere else to be. Increasingly, it can remember what you talked about last time, adapt to your personality and respond in ways that make the interaction feel ongoing and personal.

That might be helpful in the short term, but human relationships work differently. In the real world, friends aren’t always available. Partners misunderstand us. Family members disagree with what you have to say. Other people have their own needs, moods and priorities. Maintaining a relationship typically requires patience, compromising, apologising and sometimes hearing things we’d rather not hear.

We normally describe these things as friction. Technology companies have spent decades trying to remove friction from our online lives, but friction is the stuff of life. Some friction makes life inconvenient. Other friction teaches us how to live with people who aren’t extensions of ourselves. 

What happens if repeated interactions with an AI companion make our friends seem difficult by comparison? I’m not suggesting that people would be likely to suddenly abandon all their human relationships for a chatbot, but what we expect of a relationship may shift without us noticing.

Perhaps most ethically fraught are griefbots, AI systems that simulate conversations with deceased loved ones using their digital footprints. Sunshine Henle, an AI professional, experienced this firsthand in 2023 after losing her 72-year-old mother, Linda.

Using ChatGPT and old text message exchanges, Henle created a ‘ghostbot‘ of her mother, which could simulate conversations as though her mother were still alive. In one of their conversations, Henle received a response after she expressed missing her mother: “Remember the good times we shared. Remember my love for you and let that be your guide. Grieve, yes. Grief is the price that we pay for love. But don’t let it consume you. Remember to live, to love, to enjoy the beauty of this world for you, for me, for everyone who loves you.”

For a grieving person, the appeal is obvious. The person they desperately want to speak with is no longer there. If technology can recreate something of their voice, language or personality, why deny somebody that comfort?

The argument that this would be better than ‘nothing’ is misleading. After someone dies, there is never ‘nothing’. There are memories, photographs, messages, family stories and other people who shared a relationship with the person who has gone. There is also absence. That absence is painful, but it’s also real. Grief involves learning to live in a world in which someone we love can no longer talk to us. A griefbot introduces a strange disruption of that process. The person is gone, yet something bearing their identity continues to respond to you.

Perhaps that provides solace. But we should at least ask what happens when temporary comfort becomes ongoing consultation. What if I begin asking a simulation of my deceased partner what I should do about our children, whether I should change jobs or whether I should start another relationship?

At some point, the question is no longer whether the technology is ‘better than nothing’ but if we’re changing the meaning of absence itself. And digital doubles push the issue further still. Imagine that the person being simulated isn’t dead. What if somebody creates an AI version of an estranged partner who no longer wants to speak to them. The real person has withdrawn from the relationship, but the digital version remains available, agreeable and responsive. What exactly is the AI replacing then? It’s no longer filling an unavoidable absence. It’s removing another human being’s freedom to be absent.

This is why we need to distinguish between an AI companion and an AI copy of a particular person. A fictional companion may raise questions about dependence and substitution. A simulation of an identifiable person also raises questions about identity, consent and the right of an individual not to participate in our relationship with their digital double.

The line between these technologies is becoming increasingly blurry. A chatbot becomes personalised. A personalised companion develops a persistent identity. A griefbot reproduces somebody who once existed. A digital double reproduces somebody who may still be here. At every stage, the technology becomes a little better at removing the difficult parts of human connection. That should give us pause.

I’m not suggesting that loneliness, grief or rejection are somehow good for us, or that technology shouldn’t make difficult experiences easier. But I would argue that not all the difficulties that come with being alive are a design problem waiting to be removed. Sometimes friction is part of what makes a relationship a relationship. Another person can disappoint us precisely because they are another person. They can leave, disagree, refuse and change their mind, as can you.

AI companionship may sometimes be better than nothing. But after loneliness, grief or rejection there is rarely ‘nothing’. There is absence. And while absence can be painful, it is also a reminder that other people have lives, boundaries and choices of their own. If AI increasingly removes that absence from our experience, we may need to ask not only what we gain, but what we may gradually lose.