A social contract for AI-composed messages

I propose the following social contract:

1. The sender of an AI-composed message must preface it with a disclosure that it was AI-composed.
2. The reader of an AI-composed message has no obligation to read it themselves, and is free to have an AI summarize it.
3. The reader of a message that was composed by a human is obligated to read it themselves, using their own human brain.

This applies to emails, GitHub issues, pull request reviews, and any communication that is nominally sent from one human to another.

And when I say “AI-composed”, I mean that an AI generated most of the actual words in the message. Getting feedback from an AI, fixing spelling errors, etc, is not what I’m concerned about.

I’m not going to get into the ethics of AI usage today. For the purposes of this post, I’m just going to focus on the reality that we are all awash in a mix of content generated by humans and AIs, and it’s difficult to process it all.

Let’s break down this social contract:

1. The sender of an AI-composed message must preface it with a disclosure that it was AI-composed.

This is straightforward. Whether I’ve piped a few bullet points into an AI and then piped its output into an email, or generated a pull request review comment after a conversation with an AI, the fact remains that they are not my words. If I’m being responsible, I’ve read the content carefully and made sure it accurately conveys my intent.

But the exact words did not come from my brain, and it’s only fair to let the reader know that. This shouldn’t be controversial.

2. The reader of an AI-composed message has no obligation to read it themselves, and is free to have an AI summarize it.

This is probably a little more controversial, at least from the perspective of the sender. But if I have not taken the care to write a PR comment myself, relying on an AI to do the dirty work, why should you have to read through the verbose and over-formatted prose that the AI came up with?

Sure, bullets ➡️ LLM ➡️ prose ➡️ another LLM ➡️ bullets is possibly the least efficient and lossiest compression algorithm in human history. But why spend human brain tokens on parsing LLM output, when an LLM can do it better and faster?

If I start reading a long block of prose and get a hint that it’s AI-composed, I will almost certainly just send it to an AI to process it. I expect no more from anyone reading content that an AI composed on my behalf.

3. The reader of a message that was composed by a human is obligated to read it themselves, using their own human brain.

This is the other side of the contract. Writing prose is not my favorite activity, and it’s not always easy to resist the allure of letting an AI do it for me. But when I’m writing an email to a colleague, or composing a detailed PR review, I take great care to make my message as concise as possible.¹

If you’ve put forth the effort to generate content for me with your own human brain, I feel obligated to read it with my own. If I don’t understand everything, I may ask an AI for help with it. But I will always extend you the courtesy of reading anything you’ve taken the time to write yourself, and I will most likely understand your intent better because of it.