The email is due in ten minutes. You know roughly what happened, but not yet what you want to say about it. So you paste the thread into an AI tool and ask for a concise update. Seconds later, the screen offers six calm paragraphs: context, progress, risks, next steps. Nothing is obviously false. The tone is better than yours would have been under pressure. You make two small edits and send it.
This looks like the most ordinary use of AI at work. It is also where the important question first appears. The machine has not merely saved you the labor of typing. It has given the situation a shape before you decided what shape was honest.
A draft always has a point of view
Every competent draft performs acts of judgment. It decides what belongs in the foreground, which uncertainty counts as a “risk,” whose delay deserves explanation, and what can be left outside the frame. Even a neutral-sounding status update contains a theory of what happened.
AI can make those decisions fluently because it has learned the forms in which such decisions are usually expressed. That is useful. A blank page becomes editable. A confused meeting becomes a list. A tense message loses some heat. But fluency can hide the moment when a proposal becomes your account of the event.
Suppose the project is late because the deadline was impossible from the start. The generated update says the team “encountered unforeseen complexity.” It is tactful and defensible. It may also convert a structural failure into weather. If you send it, the problem is not that AI lied to you. The problem is that a sentence no one quite chose has begun choosing the future: management sees an exception, the deadline survives, and the next team inherits it.
The dangerous sentence is often a good sentence
Bad output is easy to resist. You notice the invented number, the awkward phrase, the nonexistent citation. Good output asks for a different kind of attention. It arrives polished enough to feel settled. The better it sounds, the less friction remains to remind you that something is still undecided.
This is why the familiar warning—“always check AI for accuracy”—is necessary but incomplete. A paragraph can be factually correct and still not be yours. It may name the wrong problem, protect the wrong relationship, or smooth over the one sentence that needed to remain difficult.
Keep one sentence that the machine cannot authorize
Before you ask for a draft, write one sentence in your own words. It can be clumsy. It can even be a question: “I think we are calling a planning failure an execution failure.” “I need to tell her no without pretending the request was unreasonable.” “I do not yet know whether I support this.”
That sentence is not sacred. AI may help you test it, sharpen it, or discover that it is unfair. Its purpose is simpler: it marks the place from which you entered the exchange. Now the generated text has something to answer to.
After the draft arrives, do not begin with grammar. Ask three questions. What claim did the draft make that I had not made? Who becomes easier to blame in this version? What consequence am I willing to own if the recipient acts on these words?
These questions slow down only a small part of the work. AI can still summarize the thread, propose structures, compress repetition, and produce variations. What remains with you is not every keystroke. It is the decision that turns useful language into speech for which someone can answer.
Do not romanticize the blank page
None of this means that “real” writing must happen alone. People have always thought with editors, colleagues, templates, dictionaries, and borrowed forms. An AI draft can reveal a possibility you would not have reached unaided. It can be more direct than your habits, more patient than your anxiety, and less attached to the first idea you happened to have.
The goal is not to keep the machine outside the room. It is to notice the chair you must not quietly vacate.
The most useful division of labor is therefore not “AI writes, human checks.” It is: AI opens and reorganizes possibilities; the human decides what the text is doing in the world. Sometimes that decision takes one minute. Sometimes it requires deleting the most elegant paragraph on the page.
When you press Send, the recipient will not enter a relationship with a language model. They will enter a relationship with you. Convenience can shorten the road to that moment. It cannot remove you from it.