AI can write the sympathy note. It cannot spend the evening writing it.

Years ago, long before we used any AI tools, I sat at my desk past eleven at night writing to the widow of a colleague who had died. I remember the wastebasket. I remember starting four times. I remember that the words I wanted were not available to me, and that I had to go somewhere uncomfortable and measured to find something close enough to send. Over the years that I have led organizations, there have been more tragic events, more carefully crafted messages, and more statements I rehearsed and then abandoned the moment I saw the person’s face. 

Today I could produce those letters in twenty seconds, and they would be better than the ones I sent. 

Something in that sentence is wrong. It took me a while to work out what. 

Let’s start with what the research shows. In a 2025 systematic review in the British Medical Bulletin, Howcroft and colleagues pooled thirteen of fifteen studies comparing written responses from AI chatbots with those of human healthcare professionals, and found a standardized mean difference of 0.87 in favor of the machine. In Communications Psychology the same year, Ovsyannikova and colleagues found evaluators rated AI responses as more compassionate than human ones, including expert crisis responders, and the preference remained when participants were told which was which. 

The most useful study came later. In the Journal of General Internal Medicine in 2026, Ruben, Blanch-Hartigan and Hall went back to the materials behind a 2023 study by Ayers and colleagues comparing physician and chatbot answers to patient questions on a public forum, and coded what the machine was doing differently. AI responses contained more validation, more reassurance and more non-judgmental language. They were less rushed and more structured. In the original Ayers data, physician replies averaged 52 words and the chatbot’s averaged 211. 

My reading is that what the ratings reward is the appearance of unhurried attention. A system with nothing but time will always look more caring than a person who has none. That is not a competition worth entering. Efficiency is now abundant and it belongs to the machine. What is left to us is the second half of Peter Drucker's well-known distinction: “Efficiency is doing things right; effectiveness is doing the right things.” 

Doing the right things – and the part he left implicit, answering for them. A condolence letter is not an efficiency problem. It is a question of what is right, and of who will stand behind it. 

Professor Debbie Bayntun-Lees has written that empathy fatigue comes not from feeling too much but from acting too little on what we feel. The research points somewhere adjacent. Empathy is not one capacity. Perspective-taking, empathic concern and emotional contagion recruit partly distinct neural systems.  In a European social-neuroscience study, Klimecki and colleagues trained participants first in empathic resonance (feeling what the other person felt), then in compassion (being moved to help), and scanned their brains throughout. Resonance left people feeling worse. Compassion reversed that and activated a separate network. The implication is that care may require more than sharing another’s distress. 

AI operates in between. It produces the language of concern quickly and almost for free, and it produces no real concern. Here, in the specific case of empathic communication, three mechanisms are operating in terms of how empathetic messages are perceived. 

First, convention collapse. Delegating empathic language is conventional in C-suites. Assistants draft condolences. Speechwriters write what presidents say to bereaved families. The sympathy-card industry has commoditized sentiment for a century without anyone feeling deceived. What made this acceptable was a shared understanding of which occasions permitted delegation and which did not. The card was known to be bought. The speech was known to be written. 

That knowing is gone. In 2023, Jakesch and colleagues found that readers can identify AI-written text at between 50 and 52 percent, the accuracy of a coin. Three years later, Zhu and Molnar showed more than 1,300 people a selection of personal messages and told some of them who had written them – others knew nothing about the author. Those readers who did not know who had written the messages formed impressions as warm as readers who had been told a human wrote it. Suspicion did not arise on its own. A convention governs conduct only while breaches of it can be seen, and with AI, we cannot see the breaches.  

Anat Perry argues that empathy works as a predictive signal, one that weakens when empathy becomes cheap or outsourced. Cheap empathy was always available. What dissolved were the social rules about where it was allowed. 

Second, attribution drift. Rubin and colleagues recently ran nine studies involving 6,282 participants in Nature Human Behaviour. Identical words of care were rated less empathic and less supportive when attributed to an AI tool, across different message lengths, different delays before replying, and different models. The value of a message has moved out of the message itself and into the inferred author. 

It would be easy to put that down to unfamiliarity with a new technology. There are deeper reasons. Two decades ago, Gray and colleagues showed that we judge any mind on two things: whether it can act and intend, and whether it can feel and suffer. In 2025, Colombatto and colleagues tested ChatGPT on both. Participants rated its capacity to act at 59 out of 100, and its capacity to feel at 12. A message of sympathy is worth something because someone who can be hurt chose to attend to your hurt. Competence to write it was never what mattered most. 

Third, the suspicion tax. In the same Rubin studies, participants' own uninstructed belief that AI had aided a human-attributed response reduced perceived empathy and support. Nobody told them. They deduced it, and the suspicion alone was enough. That is what the loss of convention costs. Recipients never could observe the attention we spent, and they never needed to, because the social rules told them what to assume. Now many leaders delegate their writing to AI tools, and I suspect the more perfected the message, the more suspicious it will look. 

Nobody has shown that using these tools reduces a leader’s capacity for empathy. Research on what Gerlich calls cognitive offloading, which is what happens when we let tools do our thinking, covers critical thinking and executive function. Ong and colleagues’ 2026 review treats the question as open, but it seems AI tools are more capable but less credible. 

Leaders keep asking whether they may use AI for this kind of writing. The research gives a good answer. In 2026, Sedefoglu-Ulucak and colleagues published two vignette experiments in Computers in Human Behavior, with the second varying how much AI went into a leader’s message. Generating the whole message lowered ratings of empathy, credibility and quality. Using AI to fix wording drew no negative reaction. The reason to draft an important message yourself is not that anyone can tell. It is that this is the only condition under which you can say you did. 

In February 2023, the Office for Equity, Diversity and Inclusion at Vanderbilt’s Peabody College sent students a message of condolence after the shooting at Michigan State University. It was generated by ChatGPT, and a line at the bottom said so. One student wrote of the “sick and twisted irony in making a computer write your message about community and togetherness”. The university apologized the next day and two deans temporarily stepped back. 

Nobody would have known without the footnote. That is the uncomfortable part, and I told this story in my book Human Magic as a failure of presence. The message read like every other message of its kind, which is precisely what should have told them it was not worth sending. Declaring AI assistance protects a leader who used it for something it is good for. It destroys a leader who used it for something it is not. The test is whether you would have been willing to disclose before you hit send. 

When Sharma and colleagues gave 300 peer supporters an AI system that critiqued their attempts rather than writing for them, conversational empathy rose 19.6% overall, and 38.9% among those struggling most. The message is to use AI as a coach, not as a substitute.  

The letter I sent that night many years ago was worse than the one I could generate now in seconds. It took me an evening, and was really hard to write , and that offering was invisible on the page. It always was. What has changed is that the family could once assume I cared and had tried to show it. Now nobody can assume anything, which makes the evening spent writing it the only measurement left to show that the empathy was really there.