Don't be a meat proxy

Niklas Gruhn’s Don’t be a meat proxy went up yesterday and was all over my feeds by the evening, mostly, I suspect, because of the name. A meat proxy is someone who takes what a model said and relays it to other people, in Slack, in a PR comment, in a WhatsApp group, without reading it, understanding it or checking it. His complaint is simple: I can talk to the model myself, so you forwarding it to me adds nothing except a long, confident block of text I now have to read. His ask is just as simple: read it, understand it, validate it, then write the answer in your own words. As he puts it, “By all means, prompt AI. But don’t just relay the output.”

(Simon Willison, linking to it, pointed out that Frank Allenby used a similar phrase in March. The idea’s been around. Gruhn’s version is the one that stuck.)

The part that landed for me was the code review story. Someone gets a ticket, pastes it into a coding agent, ships whatever comes out without reading it, and then, when reviewers leave comments, pastes the comments back into the agent and ships that. The reviewers end up doing the actual engineering, through a person who’s only there to move text between two windows.

I’ve seen a mild version of that on my own team, and I’ve been the meat proxy myself, once, which is enough to make me not want to be smug about this. In the spring our CFO asked me how much of our agent spend was going to the long unattended upgrade jobs. I asked an agent to work it out from the billing export, it gave me a clear answer with a table, and I pasted the table into my reply. She came back an hour later asking why the percentages added up to 112. I hadn’t read the table. I’d read the sentence above it.

We already have the customer-facing version of this rule. Since 2023 nothing goes to a customer that a model wrote and a person didn’t read, and the support-bot episode last year was about what happens when the reading gets thin. The internal version is newer. In June I asked the team to put two lines of their own above anything they paste from an agent: what they think it found, and whether they checked. That’s helped, and I’ve just proposed going further, with a rule that’s shorter than Gruhn’s. If it goes out under your name, you wrote it. That includes PR descriptions, review comments, design doc sections, incident updates, and the reply in the thread where someone asked a question you didn’t know the answer to. You can have an agent draft every word. You’re still the author, and “the agent said” isn’t an answer to “why did you say that?”

June’s two lines stay as the exception. A pasted explanation of a stack trace, labeled as the model’s and marked unchecked, is honest, and the reader knows how much to trust it. What Gruhn’s describing is AI text dressed up as a person’s considered view, with nobody behind it.

I wrote last year that I’d rather read the prompt than the output. I still would, most of the time. A meat proxy is what you get when someone sends you neither, just the output with their name on it.