Writing

The message nobody wrote

October 2026 · an 8-minute read on difficult messages, effort and what a recipient is actually reading

You experience it as help with the wording. The person reading it is working out something else: what you were prepared to go through in order to send it.

Researchers gave people a supportive message to read and varied two things independently: who had actually written it, and who the recipient was told had written it.

The messages generated by AI left people feeling more heard than the messages written by humans who weren’t trained in this sort of thing. It was also better than they were at identifying what the person was feeling, on four of the six emotions tested. And recipients who were told a message came from AI felt less heard, whoever had actually written it.

Put those two effects together and they very nearly cancel. A message written by AI and labelled as such scored 5.41 on a seven-point scale of feeling heard. One written by a person and labelled as such scored 5.51. The tool produced the warmer message; the knowing took the warmth back out.

Same words. Different transaction.

The study used messages of emotional support, which isn’t where most of us meet this. Where it shows up in an ordinary week is the note to someone whose project has just been cancelled, the apology to a client who has every right to be annoyed, the reply to a colleague who has been let down and is being decent about it. All of those arrive in your day as writing tasks. None of them is a writing task.

What the recipient is reading

When you reach for a tool to write something difficult, you experience it as help with the wording. You still chose the recipient. You still decided to send it. You still mean what it says. The sentences were the hard part and now they aren’t.

The person at the other end isn’t assessing the sentences. They’re working out what you were prepared to sit through in order to produce them. How long it took you. What it cost. Whether you were thinking about them while you did it.

Effort works as a signal because it’s expensive. Something that costs nothing to send carries no information about how much the recipient matters. Which means improving the message and removing its cost can be the same act, and only one of those is visible from where you’re sitting.

What the label actually does

There’s a sharper version of that, from a pair of experiments. Four hundred participants across two studies read messages of two kinds. Some signalled warmth: thanking somebody, apologising. Others signalled the opposite: boasting, blaming. Each was labelled either as written unassisted or as AI-assisted.

The unassisted messages behaved as you’d expect. The warm ones left people confident the sender was warm. The cold ones left them confident the sender was cold.

The AI label weakened both. An AI-assisted apology made the sender seem less warm. An AI-assisted piece of blaming made the sender seem less cold.

Which tells us the label isn’t doing what it appears to be doing. If AI involvement simply made people think worse of you, the label would have made the blamer look worse rather than better. What it did instead was stop the message counting as evidence about the person who sent it. The researchers call this signal diagnosticity, an ugly phrase for a clean idea: a message tells you about its sender to the extent that producing it required something of them.

That’s worth holding onto, because it changes what’s at stake. The problem isn’t that you come off badly. It’s that you’ve spent the message and the two of you have learned nothing about each other. The apology can’t do what an apology does.

It still arrives as a penalty in practice, for a reason that has nothing to do with AI. The messages we send deliberately are almost all meant to signal something good about us. Those are the ones a label empties out.

The ghostwriting objection

The obvious reply is that none of this is new. Politicians have speechwriters. Chief executives don’t write the all-hands email. Condolence cards arrive with the verse already printed, and nobody opening one feels short-changed.

That’s all true, and the reason it doesn’t rescue the AI case follows from the above. In each of those cases the delegation is public, so the message was never diagnostic to begin with. Nobody reads the all-hands note as evidence of what the chief executive personally went through, and nobody ever did. And the printed verse isn’t the message: the message is that somebody chose that card, wrote your name in it and posted it, which costs what it always did.

Delegation only causes trouble where the recipient hasn’t been told and is still reading the message for effort that isn’t in it. Which is where we are, because the convention hasn’t formed. It will. Five years from now the sincerity penalty may look like the people who thought a typed letter was impersonal, and that’s no use to anybody this month, because the message you’re drafting this month is read by somebody holding the present convention.

Apologies, and the question of dose

A study of workplace apologies across languages found that when people knew AI had been involved in producing one, they judged it less authentic, and that the drop in authenticity reduced their willingness to forgive. The effect ran through authenticity rather than around it. It wasn’t that the AI-mediated apologies were worse apologies. Knowing changed what the apology was.

The same study found that the dose mattered. Where the tool had been confined to translation or correcting grammar, the penalty was much smaller. That’s the most practically useful finding in this literature, and it marks a real distinction: there is a difference between a tool that helps you say what you meant and a tool that works out what to mean. Most real use sits somewhere between the two, and most of us couldn’t say where, because the question doesn’t present itself while you’re in the middle of doing it.

What it costs an organisation

Two experiments with American workers looked at what happens when somebody helps a colleague and uses AI to do it. Helpers who used it were rated less warm than helpers who didn’t. Lower warmth reduced the recipient’s sense of obligation towards them. And in the experiment that measured it, reduced obligation showed up as less help going back the other way.

371 people in the first study, 238 in the second, and it’s a chain of three effects rather than a single one, which is always the weaker kind of evidence. But the authors’ own framing is the part worth taking away: these systems lower the cost of executing the help and raise the cost of the relationship the help was supposed to build.

One feature of that study matters for everything below. The AI use was disclosed in every condition. Nobody worked out that a colleague had used a tool. They were told.

The evidence that cuts the other way

A finding points the other way, and it belongs in the body rather than a footnote.

Two preregistered experiments, 1,637 participants, playing trust games for real money. One player wrote a short message to persuade the other to send them points, with or without AI writing assistance, with the assistance sometimes disclosed and sometimes not. If AI involvement damages trust, this is the sort of place it should show up.

Mostly it didn’t. The effect of AI assistance on how much money people actually sent was minimal, and in the first study the analysis positively supported the absence of an effect rather than merely failing to find one. Disclosure didn’t significantly reduce trust either. The assisted writers built the same trust in less time, which the authors reasonably present as the headline.

There is a detail inside it worth noticing. When the messages themselves were analysed, the AI-assisted ones scored lower on linguistic authenticity: more polished, more complex, more assured. Trust was unaffected regardless.

So the finding bounds the argument rather than breaking it, and the boundary isn’t where I first assumed it was. It would be convenient to say that a trust game involves strangers with no relationship to damage, except that the feeling-heard study used anonymous strangers too. The difference is what each one measured. The trust game asked how much money somebody would hand over on the strength of a short pitch. Everything else here asked what the message told them about the person who sent it.

Those two come apart, and the diagnosticity account says why. Nobody playing a trust game is working out your character from four sentences; they’re deciding whether a pitch is worth a few pounds. The cost appears wherever somebody is reading the message for what it says about you, which is most messages that matter and almost none of the ones that don’t.

Nobody is catching this

Two further experiments, around 650 American adults in each, run seven months apart. People read messages of the ordinary consequential kind: a thank-you email from a friend, a cover letter, apology texts for cancelling dinner, a dating profile. They were told either that the sender had written it unaided, or that AI had written it, or that it was uncertain, or nothing at all.

Told it was AI, they rated the sender markedly worse on friendliness, sincerity, authenticity and trustworthiness. They judged that less effort and less time had gone into it, and that it reflected the sender’s real feelings less accurately.

Told nothing, they formed impressions as warm as they did for messages they believed were handwritten. There was no spontaneous suspicion. And people who used these tools heavily themselves were no more likely to suspect anybody else.

Set that beside what people report doing. Nearly half of UK employees say they have presented AI-generated content as their own. If recipients don’t suspect, and senders don’t say, then a good number of these messages have already been read as something they weren’t.

So it isn’t the using that costs you. It’s the saying. Which means the cost currently falls entirely on the people who say.

An earlier piece in this series found the same asymmetry at work, where disclosed AI use cost people perceived competence while identical undisclosed work cost them nothing. It appears to hold for sincerity too.

The inference sitting inside that is that you should keep quiet. I’m not going to make it, and I want to be straight about why, because it isn’t that the evidence is against it. The evidence is reasonably clear that concealment works. It’s that a practice which only holds while nobody knows about it isn’t a practice. It’s an exposure, and you’re carrying it.

What’s being avoided

Which brings this round to the thing underneath it, and it was never the wording.

Nobody opens a tool to draft a difficult message because they can’t construct the sentences. They do it because the message requires them to sit inside something uncomfortable for twenty minutes: somebody’s disappointment, their own part in causing it, the possibility that the reply makes things worse. And the tool offers a way to produce the artefact without going through the experience.

The discomfort isn’t an obstacle to writing the message. In the cases that matter, it’s most of what the message is carrying. It is the thing being read for.

So the tool works exactly as advertised. It removes the cost. The cost was the content.

There’s a second cost, and this one lands on the sender. Whatever capacity a person has for difficult conversations was built by having them, badly, a number of times. The twenty minutes of discomfort isn’t only the price of the message. It’s also the practice. Outsource enough of them and what wears away isn’t the writing. It’s the tolerance for being in the room while somebody is upset with you.

Where that leaves you

There’s no rule about disclosure at the end of this. Not because the evidence is unclear, since by this point it’s reasonably clear, but because the decision that actually matters was taken before disclosure came up at all.

Think of the last message of that kind you drafted with help. Not the scheduling email: the other sort, the one that had been sitting there for a day and a half.

What were you hoping not to feel while you wrote it?

Whatever the answer is, that was the part the person at the other end was reading for.

Sources

• Dishop, C. R., Brown, A. S., Chao, P.-Y., Kuznetsov, A. and Woolley, A. W. (2025) ‘Machines in the middle: using artificial intelligence (AI) while offering help affects warmth, felt obligations, and reciprocity’, Journal of Business and Psychology. doi:10.1007/s10869-025-10068-x
• Gillespie, N., Lockey, S., Ward, T., Macdade, A. and Hassed, G. (2025) Trust, attitudes and use of artificial intelligence: a global study 2025. The University of Melbourne and KPMG.
• Glikson, E. and Asscher, O. (2023) ‘AI-mediated apology in a multilingual work context: implications for perceived authenticity and willingness to forgive’, Computers in Human Behavior, 140, 107592.
• Khadpe, P., Wenzel, K., Loewenstein, G. and Kaufman, G. (2025) ‘Explaining the reputational risks of AI-mediated communication: messages labeled as AI-assisted are viewed as less diagnostic of the sender’s moral character’, Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2). doi:10.1609/aies.v8i2.36641
• Purcell, Z. A., Jakesch, M., Dong, M., Nussberger, A.-M. and Köbis, N. (2025) ‘Writing with AI boosts trust-building efficiency’, iScience, 28, 114092.
• Reif, J. A., Larrick, R. P. and Soll, J. B. (2025) ‘Evidence of a social evaluation penalty for using AI’, Proceedings of the National Academy of Sciences, 122(19), e2426766122.
• Yin, Y., Jia, N. and Wakslak, C. J. (2024) ‘AI can help people feel heard, but an AI label diminishes this impact’, Proceedings of the National Academy of Sciences, 121(14), e2319112121.
• Zhu, J. and Molnar, A. (2026) ‘Blissful (A)ignorance: despite the widespread adoption of AI in communication, people do not suspect AI use in realistic contexts’, Computers in Human Behavior, S0747563226000269.

The first conversation is free. About twenty minutes, no obligation, and nothing is booked or paid for until after we’ve spoken.

Arrange a free first conversation