It is late, something has upset you, and you want to talk. Your friends are asleep. Your therapist is unavailable. Your phone, however, is right there. Within seconds, a chatbot is responding with patience, reassurance and apparently thoughtful questions.

The appeal is understandable. You can say something embarrassing without watching another person’s face. You can repeat yourself. You do not have to worry about taking up too much time. For someone who feels lonely, ashamed or unsure where to begin, that accessibility can matter.

But what exactly are we receiving when a machine responds as though it understands us?

Why it can feel so personal

A chatbot can help put tangled feelings into words. It might suggest another way of looking at a situation or help someone prepare for a difficult conversation. Writing things down can also create a little distance from an overwhelming experience.

None of that requires dismissing the relief someone feels. An experience can be helpful even when the system providing it has no human understanding of what has happened.

The difficulty begins when fluent, sympathetic language is mistaken for reliable judgement. A response can sound emotionally perceptive while resting on incomplete information or a mistaken assumption. Confidence and accuracy are separate things.

Imagine describing an argument with your partner. You explain what they said, how hurt you felt and why their behaviour seemed unreasonable. The chatbot offers an interpretation that makes sense of your distress. Perhaps it is useful. Perhaps it has simply built a persuasive story around one side of an argument.

What does the research show?

There are reasons to take potential benefits seriously. A 2025 randomised trial involving 210 adults found that people using Therabot, a chatbot specifically developed for mental health treatment, improved more on measures of depression, anxiety and eating-related concerns than a waiting-list group.

However, the trial did not compare Therabot directly with psychotherapy. It tested a particular system over a short period, with follow-up at eight weeks. Its findings cannot establish that an ordinary general-purpose chatbot provides equivalent treatment or remains helpful over years.

A 2026 trial involving 995 university students also reported benefits from a conversational AI intervention, including greater anxiety improvement than its group-therapy comparison. That broadens the evidence, but a student sample and one particular intervention still cannot settle questions about different populations, severe difficulties or individual psychotherapy.

Promising findings deserve attention. They also need careful boundaries.

When reassurance becomes a problem

Being agreed with can feel comforting, particularly when we are hurt. Yet helpful support sometimes involves questioning our interpretation, noticing contradictions or tolerating uncertainty together.

A chatbot that too readily endorses our account may strengthen a conclusion that needs examining. Someone already worried that others are hostile might come away feeling more certain, rather than better able to consider alternatives.

A 2026 study published in Nature Medicine tested chatbots through simulated conversations involving psychological vulnerabilities. Researchers identified concerning responses across several risk areas, although newer models performed better. These were simulations, so they do not tell us how often actual users experience harm. They do show why a reassuring tone is an inadequate safety test.

Human therapists can misunderstand and cause harm too. The relevant question is how mistakes are recognised, challenged and repaired, and what accountability exists when something goes wrong.

There is also a difference between using a tool by choice and relying on it because affordable human support is unavailable. The popularity of chatbots should prompt questions about access to care, as well as technology.

What place might AI have?

Perhaps a useful question is what happens after the conversation. Do you feel more able to speak to someone, make a decision or return to ordinary life? Or do you keep returning for reassurance, becoming less willing to act without another answer?

Those questions concern the role the tool is taking in your life. They are more informative than whether a particular exchange felt comforting.

Privacy also deserves attention. Before sharing intimate details, check how the service stores conversations, uses data and handles deletion. Avoid including information that identifies other people. A conversational interface should not be assumed to offer the confidentiality of a therapy session.

If you use AI alongside therapy, its responses can become material to explore. What felt helpful? What irritated you? What did you wish a person would say? There is no need to turn that discussion into a competition between humans and machines.

For urgent distress or a situation in which you cannot keep yourself safe, seek human help through emergency services or urgent mental health care rather than relying on a chatbot.

AI may offer a useful starting point for reflection. The question is whether it helps us engage more fully with ourselves and other people, while leaving room to question the answers it gives.

Photograph: Thom Holmes / Unsplash.