Nobody Wants a Love Letter You Didn’t Write

Psychologists at the University of Kent ran six studies with 3,935 people to find out what we think of someone who lets an AI write their apology, their wedding vows, or their love letter. The answer was consistent: less caring, less authentic, less trustworthy, lazier. The penalty held when the text was good. It held when the person had a sound reason. It held when they were honest about it. The message people received was never the words.

The paper, by Scott Claessens, Jim Everett, and a colleague, was published in Computers in Human Behavior under the title “Negative perceptions of outsourcing to artificial intelligence.” The design was simple. Participants read about someone completing a task in one of three ways: doing it themselves, handing it to an AI, or handing it to another person. Then they rated the person.

For practical tasks, writing code or organizing a schedule, hardly anyone minded. For anything addressed to another human being, they minded a great deal. The person who outsourced a love letter or an apology was judged more harshly on nearly every measure, and the letter itself was judged less valuable, even when participants were told it was well written.

Two details in the results matter more than the headline.

The first is that outsourcing to a person was penalized too. This was never only about machines. Someone who hires a ghostwriter for their vows gets marked down as well. What people object to is the handing off.

The second is that nothing rescued it. The researchers tried the obvious repairs: the person finished the task themselves after the AI drafted it, the person had a good excuse, the person disclosed the AI use openly. The penalty persisted through all of them. As Claessens put it, “People don’t just judge what you produce, they judge how you produce it.”

The words were never the point

Everyone already knows this, which is what makes the finding interesting. Nobody has ever kept a love letter because the prose was good. People keep them because someone sat down, on a particular evening, and struggled to say a thing that was hard to say, for them.

An apology works the same way. Its content is almost always obvious. Both people know what happened. What the apology adds is the visible cost: the person had to think about what they did, find words for it, and risk getting the words wrong. That cost is the apology. A flawless paragraph generated in four seconds contains the same sentences and none of the cost, and the receiver can tell, because flawless is not how people sound when they are sorry.

The Kent team’s explanation is that effort is a signal. People read the work someone put in as a measure of how much they care, about the task and about the person it is for. Take the effort away and the signal goes with it, whatever is left on the page.

This is why the usual defense of AI-written messages misses. The defense says the tool helps people who struggle to express themselves say what they really feel. But the struggle is the expression. A halting, slightly awkward message from someone who is not good with words says something a polished one cannot: I am not good at this and I did it anyway.

Why honesty did not help

The most revealing result is that disclosure changed nothing. You might expect that telling the other person “I used AI to help with this” would buy some credit for candor. It bought none.

That makes sense once you see what is being judged. Deception was never the receiver’s worry. What they register is that when it came to them, the sender reached for a shortcut, and admitting it only confirms it.

People feel less heard the moment they know it’s AI, even when the response is identical. The Kent study is the same effect from the other side. A message means something because of who made it and what it took, and that cannot be transferred by forwarding the output.

What is being lost quietly

The study measured how observers judge the sender. It did not measure what happens to the sender, and that may be the larger cost.

Writing a hard message is how a person finds out what they think. Someone drafting an apology by hand discovers, around the third attempt, that they are sorrier about a different thing than they expected, or less sorry than they assumed. Someone writing vows learns what they are promising. Thinking out loud needs someone, and writing to a person you love is a form of it: the reader is in your head the whole time, and that is what shapes the thought.

Hand the draft to a model and that work does not happen. The message goes out. The sender never learned what they meant. Do it for every difficult message for a few years and you have a person who has sent a great many warm, well-constructed notes and has not once had to work out how they feel.

Meanwhile the practice is spreading in exactly the place it does most harm. One in four American singles already lets an AI write the message in their dating lives, and half of Gen Z does. The Kent findings say how that will land on the other end when it is discovered, and it is usually discovered.

Write it badly

There is a simple rule in all this, and it does not require giving up the tools. Let the machine have the code, the schedule, the cover letter to the landlord. The study found almost no penalty for those, and there should not be one.

But anything that begins with a person’s name and is meant to change how they feel about you should be written by you, slowly, and probably badly. The condolence card with the crossed-out line. The apology that takes two tries. The toast that runs slightly long because you did not know how to end it. Those flaws are the proof that someone was there.

The best love letter anyone will receive is three clumsy sentences from a person who has never written one. No model will ever improve on it, because the model was not the one who meant it. Staying social is mostly made of small efforts like that, and they only count if they are yours.


Sources: Claessens, S., Veitch, and Everett, J., “Negative perceptions of outsourcing to artificial intelligence,” Computers in Human Behavior (2025), 108894 (six preregistered studies, N=3,935 British participants; socio-relational versus practical tasks; penalties for effort, authenticity, trustworthiness; penalty for delegating to another person; persistence of the penalty with partial completion, justification, and disclosure). University of Kent News Centre (February 5, 2026; Claessens and Everett quotations). University of Kent, Trust in Moral Machines project summary. Psychology Today (December 2025; three-condition design, tasks). Match, Singles in America (2026; AI use in dating).

Stay Social

We hold ourselves to one promise: we push you toward the people in your life, never away from them.

We measure success by how little you need us. If you spend less time with Prinsessa because you are spending more of it with them, that is not a failure. It is proof that it is working.

That is what we stand for. In every conversation. Every day.

Someone to think with.

Follow Prinsessa