Should you let AI write your messages to friends?

The study of this has a name, AI-mediated communication, and a framework set out by Hancock, Naaman and Levy. The experimental work that follows it compares how the same kind of message is judged depending on how much AI involvement the reader believes is behind it. The result is consistent: messages known to involve AI are judged less trustworthy, less authentic and less useful, and the more the AI is thought to have done, the larger the penalty.

Two details matter more than the headline. The first is that the penalty attaches to belief, not to the text. The same words land differently depending on what the reader thinks produced them, so a well-written message is no protection once the suspicion is there. The second is a disclosure paradox: people say that being told about AI use matters to them, and then mark the message down when they are told. Honesty is penalised, which is an incentive to say nothing.

The limits are real. This is a fast-moving area with a short literature, and habits around AI-assisted writing are changing quickly enough that the size of the penalty may not hold. Most of the studies also judge a message on its own, not one arriving inside a friendship where trust is already established. Nobody has measured what a friend of ten years makes of it.

The reason people reach for help is usually the blank message. You want to say something kind, you are not sure how to put it, and a tool will put it well in four seconds. The assumption underneath is that the wording is what stands between you and a good message.

Research on gratitude letters says it is not. In a set of experiments, people wrote a letter of thanks and predicted how the recipient would react. They underestimated how positive the recipient would feel and overestimated how awkward it would be. The explanation is that the two sides attend to different things: the writer evaluates their own competence, whether it was well put, whether it sounded strange, while the recipient is not assessing the writing. They register the warmth.

So the thing AI improves is the thing only you were grading. Those were one-off exchanges in study conditions, measured straight away, and nothing in them says a clumsy message is better than a graceful one. What they say is that the clumsy sentence costs you far less than you think, and that is the cost the tool is being used to avoid.

The relationship literature has a working definition of intimacy: feeling understood, validated and cared for. The word researchers lean on is perceived. What moves a relationship is not the effort you made but what the other person registered. This is a framework resting on mostly correlational research, so it describes what goes together, not what causes what.

A specific detail is one of the few reliable ways to make attention visible. "How did your mum’s surgery go, it was this week, wasn’t it?" can only be written by someone who took in what they were told. It is difficult to fake and it cannot be produced without having listened.

That is the part a writing tool cannot supply. It can generate warmth in general, in any quantity. It cannot know that the surgery was in March unless you tell it, and once you have remembered that, the message is already written. As generated text becomes ordinary, the scarce thing in a message is the sign that one particular person noticed.

The urge to have something else write the message is often strongest after a long silence, when the message feels as though it has to be good enough to justify the gap. That fear has been tested. Across a series of experiments, people who reached out underestimated how much the recipient appreciated it, and the gap widened the more surprising the contact was.

Those experiments are about the moment of contact, not about what happens to a friendship over years. But they are enough to say that the fear of a short, plain message being unwelcome is a measured mis-prediction and not a forecast. The message does not need rescuing.

The research does not draw a line between fixing a typo and composing the whole thing. It says only that the penalty scales with involvement. A fair reading is that the nearer the tool gets to deciding what you say, the more of the message’s worth it takes with it. If you use tools at all, the safer place for them is on your side of the conversation, helping you keep hold of what people told you, and not speaking for you.

Write the short, plain version yourself. Two lines is a complete message. "Thinking of you today, hope the move went all right" will not win a prize, and your friend is not judging one.

Put one specific thing in it: the interview, the scan, the dog, the flat. If you can’t think of one, that is the real problem, and it is solved earlier, by writing the detail down in the minutes after you hear it and keeping it where you will see it again.

None of this needs an app. A note with their name and one line under it is enough. What does not work is handing over the part of the message that was the point of sending it.