AINEWS
An engineer coins "meat proxy" for the failure mode of forwarding AI replies without reading them, and the simple standard that should replace it.
When a teammate pastes a 600-word Claude reply into a Slack thread under a one-line question, the room does not get faster. It gets slower, because everyone downstream now has to read what the supposed expert did not. The pasted answer is verbose, jargon-dense, and full of plausible nonsense; the next person to reply is doing the relay person's job, and the one after that does the same. The cost is the relay person's work, offloaded onto everyone who reads it.
The pattern has a name. In an essay on gruhn.me, the author calls it a meat proxy: a person who forwards an AI's reply as if it were their own answer, without reading, validating, or rewriting it. The term is blunt on purpose. A meat proxy is a human relay that adds no judgment between the model and the reader. The author's argument is pro-literacy, not anti-AI. The model wrote the response; the proxy never read it, and that gap is the problem.
The failure shows up most clearly in code review. The author describes a loop that is now familiar to anyone shipping software: a developer pastes a ticket description into Claude Code, ignores the generated diff, and forwards the AI's review comments back into the model until something merges. The reviewer on the other end is approving a model's output wrapped in a peer's name, not a peer's code, and the merge button is signing off on a stack of plausible nonsense. The author admits he has done this himself. The critique targets frequency, not a single bad-faith actor.
A manager forwards a ChatGPT summary of a meeting to a customer and signs it as their own update. A student pastes an essay-length answer into a class chat, hoping the group will edit what the model wrote first. A support agent pastes a model-generated reply into a ticket and closes it; the customer, who already tried the obvious fix, writes back asking why the answer ignores what they sent. In every case, the relay person is the one who should have read the output, understood it, and decided whether to send it. Under this pattern, the human does not author the response — the model does, and the person merely delivers it.
The Hacker News thread on the author's post collects the same pattern across workplaces. Commenters describe being asked to read 300-line Claude responses they did not ask for, and one notes an AI-leadership colleague who generated thousands of lines of documentation that product managers and business analysts are now expected to fact-check. The thread is anecdotal, which is the right read of any comment section. The pattern is consistent: the volume of AI output has grown faster than the literacy to handle it, and the load has slid from the people who prompted the model onto the people downstream of them.
Teams, classes, and conversations stop building it, because the relay person never understood the answer in the first place, and the reader on the other end has to do the work the proxy skipped. The author puts the standard the relay person is supposed to meet in plain terms: prompt the AI, read the output, understand it, validate it, then write the response in your own words as a "decent certificate" that the prior steps happened. That last sentence is the load-bearing one. A response written in your own words is evidence that you actually read the model. A pasted reply is not.
Before forwarding any AI-generated text, ask whether you can summarize it in your own words to the next reader. If you cannot, the reply is not yours to send. If you can, the rewrite is yours, and the model is a tool you used, not a stand-in for the part of the job only you can do. The line between relay and synthesis is the line between a meat proxy and a colleague.