Recently I had a realization that seems obvious in retrospect.
LLMs should default to avoiding speaking in any author’s voice, whether that’s the robot’s or the human’s.
Unless explicitly asked to, it should avoid using “I” and “me” when writing documents, emails, and other prose artifacts. By default it should never put words in the user’s mouth.
I think this would be an improvement for users and for society.
- It would stop nudging users to become secret cyborgs.
- It would result in more professional documents.
- It would help slow the breakdown of trust in written communication.
I’ve never had a work secretary, but I imagine that’s how they work. When you ask your secretary to call someone for you they don’t pretend to be you making the call. They speak in a professional manner without drawing attention to the person making the call.
Why haven’t the big labs made this the default? Why is Claude so eager to write any outgoing communication as if I had written it? It’s kind of baffling.
To be clear: I am not without sin. I’ve been a secret cyborg. I’ve asked Claude for something, it’s given me the email in my voice, and I’ve taken the lazy way out: copy, paste, send. I have no excuse. But I wish Claude hadn’t made it the default, easy option.
In response to this realization, I worked with Claude to add this to my global CLAUDE.md file to try to do better moving forward:
Default to prose that makes no reference to the author. Describe the thing; don’t narrate it as me. No “I”, “me”, “my”, no thanks, apologies, feelings or motives offered on my behalf.
Write in my voice only when I explicitly ask for it. This holds even for something I am going to sign or send — a letter, an email, a statement. Err toward no reference to the author; if first person is wanted, I will ask for it or add it myself.