Walk into any behavioral researcher’s office that studies how people communicate under emotional stress, and you will find some version of the same artifact waiting in their data: the unsent draft. Not the message that was sent and regretted. The one that was started, abandoned, and then replaced — sometimes with silence, sometimes with something blander, and increasingly, with a prompt typed into a chat window at 11 p.m. Help me write a message to my brother about why I haven’t called. Or: How do
I tell my manager this isn’t working without burning everything down? The prompt itself is often more honest than anything the person would have sent. It names the fear directly. It admits the stakes. And then it asks a language model to take it from there. What researchers who study digital communication have started to notice isn’t that people are using AI to write their messages. It’s what they’re using it to avoid. And the thing they’re avoiding isn’t the writing. It’s the thirty seconds
before the writing — the part where you sit with something you haven’t fully figured out yet, where the sentence keeps starting and collapsing, where you realize you don’t actually know how you feel until you’ve tried and failed to say it twice. That interval. That’s the one people are outsourcing. And it turns out that interval was doing most of the work. What the shortcut looks like from the outside From the outside, reaching for an AI tool before drafting a hard message looks
like a productivity decision. Efficient. Reasonable. You have a lot going on. The message needs to be sent. Why sit there staring at a blinking cursor when something can generate a serviceable draft in four seconds? A colleague, watching over your shoulder, might call it pragmatic. A productivity blogger would probably celebrate it. And to be fair, there are messages where that framing holds — the logistics email, the meeting reschedule, the routine follow-up that requires no particular emotional excavation. For those, a generated draft
is genuinely just a time-saver, and there’s nothing lost in using one. But the difficult text is different. The difficult text — the one to the friend you’ve been avoiding, the one to the parent who said something that landed badly, the one to the person you’re not sure you’re still in love with — that one carries a specific kind of weight. And the instinct to immediately offload that weight to something external is worth sitting with, because what behavioral researchers in this field
have observed for decades is that emotional clarity and verbal expression are not separate processes. They happen together, in the attempt. You don’t know what you think until you’ve tried to write it and found that the first version was wrong. What happens in those fifteen seconds nobody accounts for? There’s a particular texture to the moment before a hard message gets written. I’ve noticed it described in almost identical terms by people who have nothing else in common: a kind of low-grade dread, a
sense of the thought being slippery, an awareness that whatever comes out first will be imprecise and possibly embarrassing. The cursor blinks. The phone screen stays lit. Nothing comes. That feeling is not a malfunction. What researchers who study cognition and emotional processing have long observed is that this specific discomfort — sitting with a half-formed thought, tolerating its incompleteness — is the condition under which the thought finishes forming. The friction isn’t a sign that you don’t know what to say. It’s the mechanism
by which you find out. You write a bad first sentence. You feel that it’s wrong. That wrongness tells you something. You try again. The second attempt is closer. By the third, you’ve learned something about what you actually meant that you could not have accessed any other way. Fifteen seconds. Sometimes less. Sometimes longer, for the harder ones. But that interval — uncomfortable, unproductive-feeling, slightly humiliating in its blankness — is where the thinking lives. When you open a chat window instead, you skip
it. The draft arrives fully formed. It is usually competent. It often sounds more composed than you feel. And it says something — but it may not say what you meant, because you hadn’t finished finding out what you meant yet. You paste it in. You send it. And sometimes, an hour later, you feel a faint unease you can’t quite locate, like a coat left somewhere you can’t remember. Why this matters more than it sounds The argument here is not that AI writing
tools are bad, or that using them makes you intellectually lazy. That’s the easy version of this conversation, and it’s not a very interesting one. The more precise observation is about what category of task the difficult text actually belongs to. There is a class of writing — and the difficult personal message sits squarely inside it — where the output is not the point. The process is the point. Writing to a friend you’ve hurt isn’t primarily a communication task. It’s a thinking task
that produces a communication artifact. The same is true of the email you’ve been putting off to a colleague whose behavior has been affecting you, or the message to a sibling about something that happened at Christmas three years ago that nobody has mentioned since. These are not messages you write because you know what you want to say. They are messages through which you discover it. What researchers who study self-disclosure and emotional processing have found, consistently, is that the act of translating an
emotional state into language changes the emotional state. Not just describes it — changes it. People who write about difficult experiences report better clarity, lower arousal, sometimes a shift in how they feel about the person they’re writing to. The writing does something to the writer. But only if the writer actually does the writing — including the failed attempts, the crossed-out lines, the version that was too angry and the version that was too soft. A generated draft skips all of that. It arrives
at the artifact without the process. Which means you get the message but not the thinking, the email but not the understanding, the text but not the thirty seconds where you would have learned something about yourself you didn’t know before you opened the app. This connects to what researchers have observed about mental friction — that empty feeling after productive AI sessions isn’t burnout, it’s the absence of the cognitive work your brain was expecting to do. Why the most caring people avoid the
most Here is the part that makes this genuinely complicated, and why the lazy interpretation doesn’t hold: the people who reach for AI tools before difficult messages are often not avoiding because they’re careless. They’re avoiding because they’re sensitive. The difficult text feels dangerous to them — not because they don’t care, but because they care considerably, and the prospect of saying the wrong thing, of the sentence landing badly, of the half-formed thought being exposed mid-formation, feels like a real risk. In that sense,
the outsourcing is a kind of self-protection. And self-protection is not the same as laziness. It has its own logic, its own history, often rooted in experiences where saying the imprecise thing did cause damage, where the unfinished thought was met with impatience or ridicule, where the blinking cursor felt less like possibility and more like exposure. What behavioral researchers in this space have observed is that the people most likely to seek assistance with emotionally loaded communication tasks are often the ones who feel
the stakes most acutely. The avoidance is proportional to the care. Which is its own kind of painful irony: the more the message matters, the more likely you are to hand it to something that doesn’t know what it means to you. What gets preserved when you stay in the discomfort There’s a kind of competence that develops in people who have learned to tolerate the blinking cursor. It’s not eloquence exactly — some of them write messily, in run-on sentences, with the emotional logic
slightly out of order. But they tend to know, by the end, what they actually meant. And the person on the receiving end tends to feel it. Not because the message was well-crafted, but because it was arrived at. Because it carries the texture of someone who sat with something uncomfortable long enough to find out what it was. That texture is hard to generate. It is, in fact, the one thing that cannot be prompted. This quality — the sense that someone has genuinely
wrestled with their thoughts — shows up in ways that go far beyond digital communication. It’s visible in how people carry themselves, in their body language during difficult conversations, in their willingness to sit with uncertainty rather than rushing toward resolution. The difficult message you’ve been avoiding isn’t waiting for better words. It’s waiting for you to sit with it long enough to find out what it’s actually about. That part takes fifteen seconds, sometimes. Sometimes it takes longer. It never takes a chat window.
Somewhere tonight, someone is typing a prompt that begins help me tell my dad. The cursor in the chat window blinks back, ready. And in the gap between that prompt and the generated response, in the two or three seconds of waiting, there is a thought that almost formed — something true, something specific, something that belonged entirely to the person typing. It doesn’t make it into the message. But it was there. It was always going to be there, if they’d stayed in the
room with it a little longer.
ChatGPT, AI writing, difficult messages, emotional stress, digital communication, self-disclosure, behavioral research, cognitive friction