The 'Meat Proxy' Test: Are You Using AI, or Is AI Using You?

There's a new term for the person who pastes AI output into an email and hits send without reading it twice: a "meat proxy." It's not a compliment, and the argument for why it matters landed twice on Hacker News in the same week, from two different angles that add up to one professional bar.

What happened

"Don't Be a Meat Proxy," a post by developer Niklas Gruhn, hit 1,835 points on Hacker News on August 3. His definition: a meat proxy is someone who prompts AI, gets an answer, and relays it to a colleague or client without reading it closely enough to stand behind it. His rule is blunt — prompt the AI, but don't just relay the output. Read it, understand it, validate it, then write the response in your own words.

The same day, a second post called "LLMs Reward Expertise" hit 1,409 points making a related but distinct case: people who already know a subject get dramatically better results from AI than people who don't, and it has nothing to do with prompting technique. The author's example was Terence Tao, the Fields Medal-winning mathematician, who has published his own ChatGPT sessions. Tao doesn't extract better math because he phrases questions cleverly — he extracts it because he can tell within a sentence when the model has gone wrong, push back on the specific error, and redirect. A novice reading the same output has no way to know which parts to distrust.

Why this matters

Put the two together and they describe the same failure from opposite ends. The meat proxy has nothing to push back with — they relay confidently-worded text because they have no independent basis to doubt it. The non-expert gets worse AI output for the identical reason: without domain knowledge, there's no internal signal that says "this sounds right but isn't." Fluent, confident, wrong output is the hardest kind to catch, and both posts are describing the same person from two directions — someone who can't tell the difference between AI sounding right and AI being right, and who therefore just forwards it.

For a manager, this reframes the real adoption question. It's not "does my team use AI" — nearly everyone does by now. It's "does my team's output improve between what the AI produced and what actually goes out the door," or is the AI's first draft also the final one, with a human name attached to it.

What to do differently Monday morning

Three questions work as a quick self-check before anything AI-drafted goes to someone else:

1. Did you change more than a typo? If the sentences you sent are the sentences the model wrote, you haven't reviewed it — you've forwarded it.
2. Could you explain why it's right, not just that it sounds right? If your only basis for confidence is the model's tone, that's not verification.
3. Would a colleague be able to tell your version from a raw AI draft? If there's nothing in it that came from what you specifically know, there's no reason for your name to be on it instead of the model's.

A comms lead at a 90-person healthcare startup

Priya runs internal and external comms for a healthcare startup and uses AI to draft weekly updates for the leadership team — pulling from product, sales, and support notes into one summary. For months her process was: paste in the notes, get a draft, skim it, send it. It read fine. Nobody complained.

What changed her habit wasn't a mistake — it was noticing that she couldn't remember what the last three updates actually said, because she hadn't really written any of them. Now she does one thing before sending: she adds a single line the model couldn't have known, usually a read on how a specific stakeholder will react to one item in the update. It takes two minutes. It's also the only part of the update that's actually hers, which is the part her CEO has started asking her about directly.

A senior financial analyst at a 15-person industrial distributor

Marcus has built forecasts for the same distributor for six years and knows which customers pay early, which vendors slip their delivery windows every March, and which line items in the P&L are noise. When he started using AI to draft the narrative sections of quarterly forecasts, the first drafts read well — clean structure, plausible numbers, confident language.

The expertise showed up in what he caught, not what he wrote. One draft attributed a revenue dip to "seasonal softness" — a phrase that sounds like analysis but explained nothing. He knew the actual cause was a single customer's late payment, a one-off, not a trend. A less experienced colleague reading the same draft would have had no reason to question it; it sounded like every other quarter's language. The fix wasn't a better prompt. It was that six years of knowing this specific business let him catch a confident sentence that happened to be wrong.

The takeaway

Prompting skill gets you a draft faster. Reading it like you'll have to defend it — and knowing enough about the subject to catch the sentence that sounds right but isn't — is still the part that has to be yours.