The Difficulty Was The Message cover

The Difficulty Was the Message

Leaders are quietly letting machines write the hardest words they have to say. The research says the words get better and the meaning disappears, and it explains exactly why

There is a moment that arrives in every leadership career, usually on a Tuesday, usually without warning. You have to tell someone their role is gone. Or that the promotion went to the person sitting next to them. Or that their father died while they were in your meeting. And you sit in front of a blank message, and the words will not come, because there are no good words, and the absence of good words is the entire problem.

Except now there is a box on your screen that will produce them. Fluent, warm, structured, appropriately sorrowful, with a closing line that lands better than anything you would have managed at eleven at night with your jaw tight. Thirty seconds. And here is the thing that should trouble you far more than it does: the message the machine writes is probably better than yours. That is not the fear. That is the finding.

This is already normal, and nobody discussed it

A survey of 1,342 American managers conducted by Resume Builder in June 2025 found that 65% were using AI at work, and that among those, 94% were using it to make decisions about the people reporting to them. Broken down, that meant 78% on raises, 77% on promotions, 66% on layoffs and 64% on terminations. One in five said they let the tool make the decision with no human input at all. Just 32% had received any formal training in the ethics of using it this way.

Treat those numbers as a weather report rather than a measurement. It is a vendor survey run through an online panel among relatively senior, well-paid American managers, not peer-reviewed research, and people answering questions about their own AI use are not famously accurate. But you do not need the number to be exact. You need only to walk through your own organisation and count the people who have quietly stopped writing their own difficult messages, and then count how many of their teams know.

The uncomfortable part: the machine is better at this

In 2023, John Ayers and colleagues published a study in JAMA Internal Medicine that should have started a much larger argument than it did. They took 195 real patient questions from a public forum, alongside the answers given by verified physicians, and generated a second set of answers with a chatbot. A blinded panel of healthcare professionals then compared them. They preferred the chatbot’s answers 79% of the time. Responses rated good or very good in quality were three and a half times more common from the machine. And responses rated empathetic or very empathetic were nearly ten times more common: 45.1% for the chatbot against 4.6% for the doctors.

Read that last figure slowly, because it is the one that undoes the comfortable version of this argument. Doctors, dealing with frightened people, managed something an evaluator would call empathetic in fewer than five cases in a hundred. The machine managed it in almost half. If your defence of human communication rests on the claim that we are warmer, the data does not support you. We are not warmer. We are tired, rushed, defensive, and working at the end of a day that already took everything.

The study has a fair objection against it, which is worth stating rather than hiding. Those physicians were volunteers typing free answers to strangers on a public forum, with no relationship, no duty of care and no time, and a machine with unlimited patience was always going to write longer and kinder. That objection is correct. It is also, for our purposes, the whole point, because it describes the exact conditions under which most managers write most of their difficult messages.

So the argument cannot be that the machine writes worse. It writes better. The argument has to be about something else entirely, and fortunately the research on that is now unusually clear.

The label is the message

In 2024, Yidan Yin, Nan Jia and Cheryl Wakslak published a paper in the Proceedings of the National Academy of Sciences with a title that contains the entire problem: AI can help people feel heard, but an AI label diminishes this impact. People who received AI-generated responses to their emotional disclosures felt more heard than people who received responses from untrained humans. They reported more hope and less distress. And when they were told the response came from an AI, that gain vanished. The size of the penalty from the label was roughly equal to the size of the benefit from the quality. The two cancelled each other out, almost exactly.

Sit with the arithmetic of that. The machine gives you a better message and takes back precisely as much as it gave, the moment the recipient knows where it came from. You have not gained anything. You have created a position that only pays while the other person remains ignorant.

It gets sharper. A 2026 study by Tobias Kleinert and colleagues in Communications Psychology, running 492 participants through emotionally engaging conversations, found that AI outperformed humans at building interpersonal closeness, but only when people believed they were talking to a human. When participants were told their partner was an AI, closeness dropped, and something else happened that matters enormously for anyone who leads people: the participants themselves wrote shorter replies. They invested less. Believing the other side was a machine did not just change how they felt about the message. It changed how much of themselves they put into the exchange.

And in the same journal, Joshua Wenger, Daryl Cameron and Michael Inzlicht reported what they call the empathy choice paradox. Across four studies, participants consistently rated AI-generated empathetic responses as higher in quality and more likely to make them feel understood, and then, when given the choice, chose to receive empathy from a human somewhere around six times in ten. They knew the machine’s words were better. They wanted the person anyway.

What they actually wanted

There is a reading of these results that flatters us and is probably wrong, which is that people are sentimental about humanity and will get over it. There is a second reading that I think is correct and considerably less comfortable.

When someone receives a difficult message from you, they are not primarily processing the language. They are reading evidence. The evidence they are reading is about cost. They are working out how much this exchange took from you, because that is the only reliable information available about where they stand with you. A message that cost you a sleepless night and three deleted drafts tells them something no sentence construction can fake: that they were worth your discomfort. This is why the clumsy, halting, obviously difficult conversation so often lands better than the polished one, and why every leader has had the disorienting experience of being thanked for a conversation they thought they handled badly.

The difficulty was never an obstacle to the message. The difficulty was the message. When you remove the cost, you do not deliver the same content more efficiently. You deliver different content, which reads, accurately, as: this mattered to me for thirty seconds.

This is why the quality of the writing cannot save you, and why leaders are going to walk into this trap at speed. Every instinct we have trained into executives says that if the output is better and the input is cheaper, you take the trade. Here, the input was the output. You are optimising away the payload.

The second leak: nobody wrote it, so nobody owns it

There is a further cost that shows up later and is harder to see.

In 2020, Jess Hohenstein and Malte Jung published work in Computers in Human Behavior on what they called AI as a moral crumple zone. When conversations went wrong, participants whose exchange had been mediated by AI assigned less responsibility to the human they were talking to, and some responsibility to the system itself. The machine in the channel absorbed blame that would otherwise have landed on a person. It was a small study, ninety-eight participants, using a messaging assistant that now looks primitive, so hold it loosely. The same experiment also found that the AI suggestions made people trust their human partner slightly more, which is not a result I can fold neatly into my argument and which you should know about before you use the study in one of your own. But the attribution mechanism it describes is one you can already watch operating in organisations.

Because the diffusion runs in both directions. The leader who did not quite write the termination letter is not quite the author of it, and knows it, and carries less. That feels like relief. What it actually is, is the slow separation of authority from accountability, which is the precondition for every institutional failure your audience has ever studied. An organisation where every hard message has a co-author who cannot be fired, promoted, or asked to explain themselves is an organisation that has quietly stopped being able to say who decided anything.

Where the line actually sits

I am not going to tell you to stop using these tools, which would be both useless and wrong. Most of what crosses a leader’s desk is administrative, and the machine should take all of it. Use it to summarise, to structure, to draft the things where nobody was ever going to read effort into the prose, to rehearse a conversation you are dreading, to tell you what you have failed to consider. Use it, particularly, to think better before you speak.

The line is narrow and it is not about importance. It is about whether the cost is the content. A restructuring announcement to two thousand people is important and the machine can help you write it, because nobody reads a company-wide email as a personal gift of your attention. The message to the one person in that restructuring whose work you have watched for four years is a different object entirely, and if a model writes it, you have sent them a beautifully worded notification that they were not worth the hour.

Terminations. Condolences. Apologies where you were wrong. Feedback that will hurt. Recognition that is meant to land as personal. In those five, the awkwardness is not a bug in your delivery, it is the proof of the claim you are making, and the machine cannot manufacture proof of a cost it did not pay.

And if you do use it, say so. The disclosure penalty in the research is real, but it is a one-time cost paid in exchange for the thing you cannot rebuild once it goes: the recipient’s ability to believe that when you did write something yourself, you meant it. The risk was never that you used AI. The risk is discovery, and in 2026 discovery is a matter of time.

Why this is an HX problem, not an IT problem

HX = CX + EX. Human experience is what customers experience plus what employees experience, and it cannot be separated, because the same people produce both on the same day with the same nervous system.

Watch what is happening on the customer side and you will see the identical move: the apology written by a model, the personalised note that was personalised by a system, the warmth that scales. It works right up until the customer recognises the pattern, and then it does not merely stop working, it becomes evidence against you, because they now have to reconsider every previous message they believed. The internal version is worse only because the relationships are longer and the memory is better.

What you do to people on the inside, they do to customers on the outside. If your managers learn that hard conversations are a text-generation problem, your service recovery will be a text-generation problem within the year, and your customers will work it out faster than your employees did.

So here is the experiment for this week, and it takes ten seconds. Before you send the next message that a model drafted for you, ask whether you would be comfortable telling the recipient that a model drafted it.

If the answer is no, you have just learned something important, and notice that no machine could have told you.

References

Ayers, J. W., et al. (2023). Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Internal Medicine, 183(6), 589–596.

Hohenstein, J., & Jung, M. (2020). AI as a moral crumple zone: The effects of AI-mediated communication on attribution and trust. Computers in Human Behavior, 106, 106190.

Kleinert, T., Waldschütz, M., Blau, J., Heinrichs, M., & Schiller, B. (2026). AI outperforms humans in establishing interpersonal closeness in emotionally engaging interactions, but only when labelled as human. Communications Psychology, 4, 23.

Resume Builder (2025, June 24). Half of managers use AI to determine who gets promoted and fired. Online survey of 1,342 U.S. full-time managers, fielded via Pollfish.

Wenger, J. D., Cameron, C. D., & Inzlicht, M. (2026). People choose to receive human empathy despite rating AI empathy higher. Communications Psychology, 4, 19.

Yin, Y., Jia, N., & 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.

#TheH2HExperiment , #Leadership , #ArtificialIntelligence , #EmployeeExperience , #FutureOfWork

Link to original article: The Difficulty Was the Message