A familiar situation: you ask a colleague a specific question, and the answer comes back as a wall of confident text filled with dense terminology. You start to dissect it — and realize that the person didn’t even read it: they forwarded your question to a chat with a neural network and returned the answer as is. In the West, such things have recently gained a concise name — meat proxy, a “meat proxy”: a middleman who routes traffic between people and the model, adding nothing of their own. The term spread instantly because it hit the mark. Let’s unpack the phenomenon itself — it concerns everyone who works with agents, and on our forum, most of us do.
Why is this even a problem
It may seem that the person “helps quickly.” In reality, the scheme breaks in three places at once.
The value of the intermediary is zero. If you didn’t read the model’s answer — the interlocutor could have asked it themselves, faster and without the broken telephone. You were asked not because you have access to the neural network, but because you (presumably) understand.
The economics are inverted. Generating text now costs seconds, while verification — still minutes and hours. By forwarding an unread answer, you transfer the most expensive part of the work — verification — to the recipient, keeping for yourself the cheapest: Ctrl+C, Ctrl+V. That’s why such behavior annoys people more than a frank “I don’t know.”
Accountability is vanishing. The code might be “yours,” the answer might be “yours” — but you can’t explain them. When something breaks, there’s no one to pin blame on: you’ll shrug at the model, the model will stay silent.
How this looks in our reality
Examples that everyone has seen (and some have even practiced):
- Neural network configs without testing. A person asks the model to assemble a config Xray or sing-box, doesn’t run it themselves — and posts in the thread as if it’s a working answer. The next three hours the thread’s author debugs hallucinated fields that don’t exist in this core version. With JSON configs, model proxies err readily and convincingly.
- Code review through a middleman. The reviewer leaves a remark, the developer feeds it to the agent, commits the result without looking and waits for the next remark. Formally it’s review, but in practice the reviewer talks to the model through a human pipe.
- “Help” in chats and forums. The answer to a beginner’s question is a wall of ChatGPT, where half the commands are outdated. It looks like a community contribution, but works as cluttering.
How to use agents without becoming a proxy
Irony is that active use of AI here isn’t the problem — it’s whether you stay a filter or become a pipe. Simple rules:
- Read before you send. Being lazy to read your own answer isn’t an excuse to send it: the recipient will have to spend more effort on it than you.
- Summarize in your own words. This is the cheapest test of understanding: if you can’t paraphrase it, you didn’t grasp it, and you shouldn’t send it.
- Run it before you advise. Config, command, script from the model — first on your own stack, then in the thread. If you didn’t check it — honestly mark: “not tested, but direction is this.”
- You’re responsible for fixes to the agent. Making changes through the model is fine. Not understanding what it changed — no.
- Compress, don’t inflate. A model’s answer can almost always be shortened by a factor of three without loss of meaning. Do it yourself — don’t burden the reader.
There’s also a downside: forwarding the model’s output is possible — if you’ve checked it, filtered it, and honestly said where it came from. The value is created precisely in this layer — selection, verification, adaptation to context. Remove it and you’ll be left with only the delay of transfer, like a bad proxy.
In essence, this is the border between conscious vibe coding and its caricature: the model is an amplifier of your thinking, not a replacement for it. An amplifier without input signal just outputs noise.
Sources
- Essay “Don’t be a meat proxy,” which gave the phenomenon its name — one of the year’s most discussed Hacker News pieces (~1700 points)
- Thread discussing on Hacker News (~700 comments)
Have you run into “meat proxies” in your team or in threads — or caught yourself doing it? And how should the forum respond to unverified AI answers: the rule to “note that it’s generated,” or is common sense enough?