Meat proxy: the copy-paste that turns your colleagues into a serving hatch

Painted illustration in blue, green and pink tones: in a restaurant kitchen full of plants, a waiter carries off a dish under a cloche while a small robot cook bustles at the stove.
AuthorSébastien Klifa
Date13 November 2026
Reading time8 minutes

Tuesday, 9:32 am, a Slack notification. A colleague has just shared a summary of an article in a team conversation. Three pages, bold headings, bullet lists, emojis, emojis, emojis. With a little note: “I made you a quick summary.” Impressive. Especially since the article in question was published… four minutes ago.

You start reading. The sentences are correct, it’s fairly well structured, but you struggle to follow the thread and even more to see what you’re supposed to take away from it. The pile-up of “it’s not X, it’s Y” and of sentences with three adjectives confirms the diagnosis. At the bottom, there’s even a line left over that must have been pasted by mistake: “Would you like me to prepare a shorter version?” It’s another f***ing Claude summary. And when you ask your colleague what he really thinks of it, he can’t answer. He doesn’t know what to take away from it either. He asked Mr. Claude or Mr. GPT to summarize the article and shared the result as is.

This way of acting as a serving hatch between an AI and other people has a name: the “meat proxy”, a relay made of flesh and blood. Niklas Gruhn describes this behaviour in his post “Don’t be a meat proxy” (opens in new window), published in August 2026. Among other things, he tells how he asks a question on Slack and gets back “Claude said:”, followed by a very long answer copied over word for word.

Using an AI isn’t the issue. The problem starts when you pass on what it produced without having looked at what it says. Sometimes you haven’t even read it. Sometimes you skimmed it and it looked fairly serious. Either way, you leave it to someone else to work out whether the answer is worth anything.

You’ve probably already…

Seen a big block of AI-generated text or code land in a team discussion, without a word of explanation (developers call this a “slop grenade”). Asked someone to explain the work they’ve just finished and realized they can’t. Found out that a quote or a report went to the client as is, with approximations a proofread could have caught. And as for LinkedIn, don’t get us started. Between the lecturers, the prompt sellers, the design revolutionaries who no longer need software and the generated images filled with text that isn’t really text, some people would do well to read what they post.

It happens outside work too. You ask a friend for advice, and he answers with a wall of text copy-pasted from a chatbot. You wanted his own opinion, because he knows you and your situation. If he used an AI to help him think, great. But you’d like to know what he thinks instead of having to guess it in the middle of all those paragraphs.

Painted illustration: at a café table surrounded by plants, a man hands his friend a thick stack of printed pages.

And yet the temptation is understandable. You’re short on time, the answer comes fast and it’s nicely presented. You may also feel less comfortable writing than the AI. Except that you’re sending this answer to someone who is counting on you to have done a minimum of sorting.

A bit hurtful, isn’t it?

You asked a person a question and you get the answer of a machine that this person hasn’t even read. It can give the impression that your question wasn’t worth two minutes of their time, when you were counting on them and on what they know about the subject.

The philosopher Kenneth Boyd examines this question in an article for the Prindle Institute (opens in new window). For him, handing the question off to a chatbot without giving it that minimum of attention means not taking seriously the person who asked it. And it also says something about what you think of your own expertise. You’re the one who was asked the question. If the person wanted an AI’s opinion, they probably have a subscription too. Some people add “ChatGPT says so”. Then you know where the answer comes from, but still not whether it’s right.

The nonsense filter

Boyd also points out that when you pass on information, you take responsibility for it. He gives the example of a search on a forum. You find three answers to your problem: one off topic, one that solves it and one that advises you to smash your keyboard with a hammer. You’ll sort through them, or at least we hope so for your keyboard’s sake. You didn’t invent the solution, but you did at least check whether it fit.

Painted illustration: in close-up, the hand of a man in a suit brings a hammer down on his computer keyboard.

That’s your nonsense filter. The meat proxy doesn’t even do that: he pastes and he sends. Yet an AI can state false information with confidence, or a recommendation that doesn’t fit the situation. If nobody takes a moment over it before sending, it’s up to the recipient to notice.

And fine words can impress us. We hesitate to challenge an answer that looks so well built, especially on a subject we don’t know well. Yet sometimes we know of a constraint the AI didn’t take into account, a decision already made, a detail of the trade. What we know about the situation, the precise context, matters, and not just a little, even if we put it less elegantly.

This blind trust can go a long way. Everyone has heard of an AI recommending a restaurant that never existed. Or worse, hikers getting lost in the heights above Cauterets (opens in new window) because they had asked ChatGPT for their route. If you don’t see the problem when you use it for yourself, you’re even less likely to see it before passing it on to others.

Time saved by one, work left to the other

The sender may have saved time. The recipient now has to read all of it, look for what answers their question, check whatever seems doubtful and sometimes ask for clarification. The text is ready, but it still has to be understood, and it’s the recipient who takes that on.

Researchers at BetterUp Labs and the Stanford Social Media Lab (opens in new window) use the term “workslop” for this AI-generated content that looks polished but lacks the substance needed to move the work forward. In particular, they describe the extra effort that falls on colleagues who have to rework it.

Gruhn recounts a case from software development. A developer hands his ticket to Claude Code, sends the code without reading it, then passes the reviewers’ comments back to the AI. The code works (sometimes). It’s the reviewers who have to understand what was produced, spot the problems and ask for fixes. The biological intermediary just handles the back-and-forth between the two.

Not to mention receiving three pages for a question that two sentences would have answered.

Painted illustration: an office worker has fallen asleep with his head on his desk, under a long roll of paper coming out of a printer.

And what about us?

We use AI a lot too. It helps us clean up texts, smooth them out, go from our notes and talks to a written text. But before sharing them, we have to be able to explain what we claim in them. A sentence that sounds good may have changed the meaning of what we wanted to say, and it’s up to us to catch that.

After several months of failures and situations like these, we ended up writing it down in black and white in our internal rules on AI: an AI-written summary shared within the team is no longer than half a page. What has to stay in it is the goal and the result, what works and what doesn’t. AI may keep helping us summarize documents. But if we have something interesting to tell the team, half a page is enough to get the substance across. For the rest, we can talk about it directly, between real people. It has drastically cut down the practice in-house (the meat proxy, mind you, not talking to each other as real people).

Also, our AI use statement details what we use it for, from the site’s code to translations and some of the blog’s illustrations. Disclosing this help also lets readers know how the content was prepared. The rest, proofreading and validation, is done by us before publishing.

What if the 9:32 colleague was you?

Next Tuesday, when your summary is ready to go out at 9:32, it can easily wait until 9:40. No one will hold it against you for taking a few minutes to understand what you’re about to share, check what deserves checking and form your own opinion. Even if it’s tempting to just delete the last line.

Sources

Want to discuss your project?

Everyone talks about user experience, service design, ergonomics… It’s not very clear, but you’d like to explore all that or get better at it.

Contact us