Meat proxies: what to keep for yourself
You have seen the word by now. On the third of August Niklas Gruhn put a post on his blog titled "Don't be a meat proxy": the colleague who forwards you the model's answer without reading it. His own line is "I don't need a meat proxy in between". Hacker News picked it up, seven hundred comments, and from there it went everywhere. Two weeks, and it was already an insult.
The joke landed. Nobody bothered to say what you are supposed to do instead, and that is the part worth writing down. There is no study behind the term and no statistics: it is three-week-old slang from an irritated engineer. But it landed on a real mechanism.
Why the proxy goes first
Not because the machine is smarter. The arithmetic is simple: if you add nothing between the input and the output, you are not a link, you are latency. Lean manufacturing has a word for that operation, waste, the kind the customer will not pay for because it does not change the product. The person you forward to will open the same model and get the same thing a minute sooner.
Meat proxy describes a position in the value chain. A specialist lands in it, and so does a director.
When the model writes the text, what you add comes down to four things, and the model brings none of them. The first is framing the task. It answers the question you asked rather than the one you needed, and the gap between the two is the work. "Make a deck on the quarter" and "make a deck that has to convince the board not to cut the automation budget, because the effect arrives in year two" are two different assignments, and the second never appears on its own.
The second is context and constraints: everything absent from open sources. How things are done at your company, who reads this, what happened last time, what is off limits. The model fills the gap with the average of the internet, and the average of the internet is exactly what stops being paid for.
The third is acceptance, and it gets skipped most often. Not spelling, but joints: do the numbers agree across sections, is there a contradiction with what you said last month, did the answer quietly drop the inconvenient option.
How far this is from a formality I found out on my own project. Going through an implementation plan with an agent, I ran six rounds of clarifying and checking questions. The first question found seven gaps. The second showed that not a single item in the plan had acceptance criteria: the plan said what to do and never said how we would know it was done. The model was the strongest available, Fable 5 in maximum reasoning mode. Skip those six rounds and the plan travels down the chain unchanged, and I am precisely the proxy this article is about.
And accountability. A sent letter and an assigned task have consequences, and a person answers for them. The model does not accept accountability, and that is how the world is built rather than a technical limitation. Remove these four things and what remains of you is a forward button. That automates completely.
The 70/30 split
Seventy percent of the work goes to the assistant: research, drafts, wording options, proofreading, routine, and arguing against my own position when I ask it to. Thirty stays with me: thinking the task through, setting constraints and context, steering, checking, deciding, and carrying the decision.
I am not counting minutes or lines. I am counting who holds the frame of the task: who sets it, who draws the boundaries, who answers for it.
On a letter that assigns work the line is easiest to see. Form goes to the machine: structure, tone, closing options, a check for ambiguity. What stays with you is what the model does not have. Who this goes to, what that person already knows, how their last attempt at this failed, what counts as done and by when. Simple test: strip out everything you wrote by hand and see whether the assignment is still executable. Usually it is not.
A management decision is harder, because the model happily takes on more than it should. Generating options, finding precedents, mapping risks, playing the opponent, all of that is its work. Yours is the criteria and their weights. Criteria do not delegate. The model will offer its own, they will be reasonable, and they will not be yours. If you cannot say the criterion out loud before you see the options, you are not choosing, you are agreeing.
The second floor of those thirty percent
I used to justify the thirty percent by quality and risk. Somebody has to think it through, set the frame, check and decide, or it comes out fast and wrong.
Now I see a second floor to the rule. Those thirty percent are also the share you keep growing on.
I did not notice it right away. I state a position or a decision and ask for it to be taken apart: where the logic sags, what an opponent would say. I argue, rewrite, take it apart again. Those six rounds on the plan are the same species, and one of them went entirely to my answers: the agent asked, I answered. What is left after each pass is not the text, it is the method. A few months in I caught myself noticing that the structure of an argument now arrives whole instead of being assembled after the fact. I will not claim my thinking changed, I did not measure that. The form in which a thought arrives is what changed.
In the third edition of my book the thirty percent are described as "think it through, set direction and constraints, control, make the final decision". Learning is not in that list, and what I am writing now adds a fifth element to the rule.
Although the mechanism is already in the book, at company scale. It says: "Competencies degrade not because people get dumber, but because they stop training: the machine prepares the decision." I wrote that about organisations that hand decisions to advisers wholesale. Now I see the sentence transfers to a single person, myself included.
The mechanism is not mine. In education, the model where a learner works alongside a master and takes over the method rather than the answers is called cognitive apprenticeship; Allan Collins, John Seely Brown and Susan Newman described it in the late eighties [1]. There, support is withdrawn step by step as the learner takes on more responsibility [2]. Without that phase the learner stays dependent on help, and that phase is the one people forget.
Elizabeth and Robert Bjork arrived at the same place from another direction with desirable difficulties [3]. Conditions that improve performance quickly often fail to make the skill stick; conditions that slow visible progress work for transfer. What makes the work more pleasant today often costs quality a year out.
Both bodies of work are about people learning from people. The transfer to working with a language model is mine, and it is a hypothesis rather than an established fact. I tested it on one person, myself.
Three conditions, or it does not work
"Do it for me" and "do it with me" end differently. In the first you get a text; in the second you pay attention and take the method. The same hour with an assistant ends in opposite places, and no usage report tells them apart.
The shape of the request is the second condition. "Improve my text" trains the model to please you: it will happily polish weak logic on request. What trains you is "break my argument". A leading question starts the nodding, and you get an echo instead of a review.
The third condition is the "now do it yourself" phase. If support is never withdrawn, partnership turns into dependency: constant help replaces the skill instead of building it. None of the three switches on by itself; each takes a decision from you.
The same thing at company scale, only more visible
In the book I go through a case where staff mastered a corporate chatbot and stopped thinking. A guest asks, the employee asks the bot, the employee relays the answer. Service did not drop: replies got faster and better written than before. A year later it turned out there was nobody to handle a non-standard question. The people on the floor had lost the habit, and the bot does not understand questions like that.
From the outside it looks like a successful rollout. Speed rises, complaints fall, the effect shows up in the report. The failure arrives from the direction nobody measured.
The error is the same one as in accepting an agent's work. With no criteria, "done" in the report means one thing: it finished typing. A report on the work gets accepted in place of the result, and at company scale that runs for years.
So when you decide what to hand the assistant inside a process, ask about speed and then immediately ask who will be able to check this in a year. No answer means what got automated is your expertise rather than an operation.
The price I am not hiding
Partnership is not faster than delegation. It costs more.
Training and saving draw on the same pocket, your attention. I have written before that AI moves attention off tasks where a mistake is cheap and onto tasks where it is expensive. Here is what that frame was missing: freed attention does not redistribute itself. Part of it has to come back into your own thirty percent.
At industry scale I have described this trade already: if AI takes the junior engineer's work, where does the senior capable of verifying it come from ten years later. "We risk winning three years of productivity and losing a generation of expertise." For an individual manager the trade is the same, only faster and quieter: a year of productivity in exchange for your own expertise.
Percentages are not surrendered by decision. Nobody says to themselves, I will give away my development. They go in pieces, and "no time to dig in today, I will take it as is" is the right call in every single instance and a losing one over a year. The feedback is delayed: the thirty percent you gave away do not spoil the result at all for a while. The signal that you stopped growing arrives not tomorrow but the day something has to be done without the assistant.
How to check yourself
The feeling of "I am engaging with it" cannot be trusted. From the inside, understanding and the illusion of understanding feel identical; both produce clarity.
So the check has to be external. Once a quarter take a task of the kind you normally do with the assistant and do it without. The question is not whether it came out well. The question is whether the structure arrived on its own.
Between quarters, four errors do the work; I have written them up separately. Taking agreement for verification. Asking a leading question and getting confirmation of what you fed in. Asking with no criterion. Failing to notice there is nothing to answer with. They share a root: a question with no ground under it. The form is there, the footing is not.
Two cases are visible immediately. The model answers "yes, all good" to any material you give it, which is a misconfiguration rather than praise. Or you spot an error and ask the same instrument that produced it to confirm. Both look like verification and neither is.
The third one bothers me most. Every time I catch the model it is not because the question was good but because I know the subject better than it does. Remove that layer and the question is left without an arbiter: the answer will look like verification, and there will be nobody to verify the verification.
Both extremes are expensive
It is easy to read this as "use AI less", and that reading is wrong. Whoever avoids AI is losing speed and quality today. Nobody is taking the tool away; everyone will have it, the machine will absorb everything average, and the only thing worth paying for will be what you add on top.
So there are two extremes and both are expensive. Staying away from the tool costs you speed now. Handing it everything costs you yourself in a year. The first leaves you a subject. The second makes you an interface, and interfaces get optimised.
What I am not claiming
That AI changes thinking. I did not measure thinking; I observed the form in which a thought arrives.
That partnership is more effective than delegation. It costs more attention, and that is not a figure of speech.
That this holds for everyone. One person, one role, one set of tasks, no measurement. This is a look at a mechanism, not research.
Sources
Origin of the term: Niklas Gruhn. Don't be a meat proxy. Blog post at gruhn.me, 3 August 2026.
Quotes on degrading competencies and on a generation of expertise: "Artificial Intelligence. Freefall", third edition. The six-round case, the four errors and the tests for imitation of verification are covered in "Questioning AI: form and environment".
1. Collins A., Brown J. S., Newman S. E. Cognitive apprenticeship: Teaching the crafts of reading, writing, and mathematics // Knowing, Learning, and Instruction / ed. L. B. Resnick. Hillsdale: Erlbaum, 1989. P. 453–494.
2. Collins A., Brown J. S., Holum A. Cognitive apprenticeship: Making thinking visible // American Educator. 1991. Winter. The term scaffolding was introduced earlier: Wood D., Bruner J. S., Ross G. The role of tutoring in problem solving // Journal of Child Psychology and Psychiatry. 1976. Vol. 17. P. 89–100.
3. Bjork E. L., Bjork R. A. Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning // Psychology and the Real World. New York: Worth Publishers, 2011. Ch. 5. P. 56–64.
