At HarperCollins, Simon & Schuster and Hachette, staff use ChatGPT and Claude for everyday writing: emails to literary agents, publicity copy, back-cover text and marketing material. Authors have not been told. Hachette's own policy rules out AI for communicating with authors and partners, and its employees say the emails go out anyway. What the agents did about it is the useful part. The reporting from inside the publishers shows how it started.
The reader sees what the writer cannot
When you use AI on one email, you check that one email. It reads fine, so you send it. The recipient is in a different position. An agent who gets rejections from the same editor on five books sees five emails built the same way, and agents talk to each other. Put the letters side by side and the habits show: the "not this, but that" sentence, the list of three.
The same thing holds wherever people receive many messages from one sender and share them. A supplier gets your quote replies, a candidate gets your recruiting emails, a tenant gets your notices, and each of them forwards things to a friend. Half of American workers already use AI to write emails, so the habit is not the problem. The problem is that a rejection letter is the message where a person most wants to feel that someone read their work, and a machine-sounding one tells them nobody did.
The clause that applies to both sides
The agents' second move matters more for most readers. Agents fear that publishers paste manuscripts into public chatbots to produce rejection letters and blurbs, so one has added a line to contracts saying the material may not be run through an LLM. It was written with tech companies in mind. But the clause says nothing about who is doing the pasting, so it covers the publisher too. The Authors Guild has proposed model language for the same worry: no uploading of an author's work to a consumer chatbot without written permission.
At HarperCollins, editors who asked whether they could paste a full manuscript into a public chatbot were told no by the legal team. So the rule exists inside the company. The authors and agents on the other side have not been told what it is.
Consider a bookkeeper who has signed a client's agreement with a line like this, and who then pastes the client's bank statements into ChatGPT to speed up a reconciliation. Nobody is angry at the time. The first anyone hears of it is when a client compares notes with another, or reads the contract again. Clauses like this will turn up in far more contracts than publishing's, because every client with confidential files has the same reason to ask for one.
Why the shortcut spreads where nobody checks
The staff describe a pattern that most workplaces will recognize. Teams were cut to five or six people while the number of books doubled or tripled. At HarperCollins, managers bought Claude licenses before they knew what to do with them, and staff were assigned to find uses. At Simon & Schuster, employees wrote an open letter against workflow-analysis software that they feared would be used to find work to automate; the company's chief executive replied that no decision has been made to use it.
When AI arrives as an instruction with no rule attached, it shows up first in outgoing messages, because that is where a tired person saves the most time and where a manager is least likely to look. It is also where outsiders are most likely to see it. A company that wants the time savings without the lost trust needs a short rule more than another license: which messages a person writes in their own voice, what is always disclosed, and which client files never go into a public chatbot. Disclosure rules are starting to appear elsewhere: one workplace tribunal now requires AI disclosure in filings.
Before the next client file goes into a chatbot, read what you signed about AI. Then tell your own clients which of your messages a person wrote.