I get probably five LinkedIn messages a day that are obviously written by AI.
They’re too long. Too polished in a very specific way. They’re carefully engineered to “start a conversation” I have absolutely no interest in having. Sometimes they’re followed by an email explaining that the LinkedIn message was sent “by mistake,” which is apparently another automated tactic designed to get me to respond.
They’re exhausting.
And the biggest problem isn’t that AI wrote them.
It’s that I’m pretty sure no human really did anything between the AI generating the message and me receiving it.
Someone gave a system a target, let it generate some words, and passed those words along.
They became a meat proxy.
I recently came across the phrase “meat proxy” in a post by software engineer Niklas Gruhn, and it stuck with me. He was describing the increasingly common experience of asking someone a question and getting back a giant AI response essentially verbatim. His point was simple: if all you’re doing is relaying Claude’s output, you aren’t adding much value.^1
A meat proxy is basically a human router:
Someone asks you something.
You ask AI.
AI gives you an answer.
You pass the answer to the other human.
That’s it.
No real interrogation. No judgment. No context added. Maybe no understanding at all.
And I think this points to a much bigger divide that is starting to emerge around AI.
AI can make you substantially smarter, faster, and more capable than you were before.
It can also make you dumber.
The difference is whether you use it to avoid thinking or to think further.
Did you actually read it?
You can usually tell when someone is operating as a meat proxy.
The output is often ridiculously verbose because generating words is free now.
There are five paragraphs where two sentences would have worked.
The answer ignores something that was discussed yesterday.
It confidently says something that is completely wrong.
There’s jargon nobody involved in the conversation would naturally use.
And, yes, there are usually way too many em dashes.
The person sending it may technically be communicating more, but they’re transferring the work of understanding onto the recipient.
Now I have to read your giant message, figure out what you actually need from me, separate useful information from filler, and determine whether any of it is even correct.
That’s not leverage.
It’s cognitive littering.
The simplest test I’ve started thinking about is:
Do you actually understand what you’re sending?
Did you fucking read it?
If I ask why the recommendation says what it says, can you explain it?
If there’s missing context, would you recognize that?
If I tell you part of it is wrong, can you reason through why?
If you don’t understand the output, you aren’t really delegating work to AI.
You’re routing packets.
I give AI an enormous amount of work
None of this means I think we should protect work from AI.
Quite the opposite.
I hand an absurd amount of work to AI.
For the last several months, I’ve been coming back repeatedly to a POC around agent-driven animated characters.
It started as an idea I was curious about and turned into hundreds of hours of painstaking experimentation.
We tried things.
They failed.
We tried them differently.
We tested models.
We explored where automation worked and where it broke.
We figured out what agents could handle, what traditional tooling could handle, and what still genuinely needed a human.
And throughout the process, I kept pushing to make the findings durable.
What did we learn?
What should we document?
What will we forget three months from now?
What would someone else need to understand if they picked this up?
How far can we push this?
What happens if we approach it a completely different way?
Should we rent a GPU and audition a bunch of video models?
Can we generate different narrator voices and blind test them?
What if we change the way we color grade the outputs?
What does the production version eventually need to look like?
Some of those AI sessions have effectively stayed alive for weeks because I keep returning to the same line of thought and pushing it further.
Eventually, I reached a pretty clear conclusion about one part of the workflow.
We still need a human artist.
Not for rigging.
Not for animation.
For the art.
So I found the human.
And before our kickoff, I used AI to help turn everything we had learned into a brief.
That brief contained details that would have been very easy to miss.
For example, if a character is shown in profile, I still need the arm you can’t see. The body parts need to be separated and overlap correctly so the character can eventually be rigged and animated. I need exploded parts as well as an example of the completed character.
Those details matter.
Without them, an artist could deliver something beautiful that is fundamentally wrong for the downstream system.
AI generated the first version of the brief quickly.
And then I spent another 45 minutes going back and forth on it.
Is this clear?
Are we missing any common rigging problems?
Will an artist understand why we’re asking for this?
State that more plainly.
Get rid of the technical jargon.
Make the phases obvious.
Make the deliverables crystal clear so he can scope the work immediately.
Could I have sent the first version?
Sure.
It probably would have been fine.
But after hundreds of hours learning the problem, why wouldn’t I spend another 45 minutes making sure the human entering the process gets the benefit of all that learning?
That isn’t outsourcing thinking.
It’s the payoff from thinking.
AI should let us hand humans better problems
This is one of the things I think we’re missing in the conversation about AI.
The goal isn’t simply to replace human work.
AI can also help us make the human work that remains dramatically better.
The artist in my example shouldn’t have to rediscover everything I already learned.
He shouldn’t spend hours creating a file only for me to realize later that the body parts aren’t structured in a way that makes animation possible.
I already paid that learning cost.
AI helped me package and transfer it.
Now he can spend his time doing the thing I actually need a human to do: making great art.
Good AI use reduces wasted human work.
A meat proxy does the opposite.
The meat proxy saves themselves five minutes and creates twenty minutes of work for everyone downstream.
Prompting isn’t the magic
This is also why I think we overrate prompts.
People talk about prompts like they’re secret incantations.
“Here are my ten ChatGPT prompts that changed everything.”
Maybe.
But unless you’re building reusable skills, agent workflows, or orchestration, I don’t think the individual prompt is usually the interesting part.
Some of my best AI sessions start with what would be an absolutely appalling message to another human.
It’s a brain dump.
Half-formed thoughts. Missing punctuation. Contradictions. Things I’m unsure about. Things I think might matter.
Then AI takes a first pass at understanding me.
And we start working.
That part is the skill.
“I don’t understand enough to decide this. Help me understand.”
“Step back. What other options do we have?”
“This sounds technical. Explain it plainly.”
“You’re assuming something I didn’t say.”
“What would make this fail?”
“Try this a different way.”
“What am I missing?”
“No. That still isn’t clear.”
I’ve written before about stubbornness being one of the most important traits when working with AI.
You keep pushing.
The first answer isn’t sacred.
Neither is the fifth.
You’re building context together.
I don’t have all the answers when I begin, either. I’m developing my own understanding as we work.
The collaboration makes both sides of the context richer.
I just have a slightly higher context window.
AI has made me think more, not less
The most interesting thing about AI for me personally is that I’m doing more thinking than I did before.
Not less.
AI makes exploration cheap.
Before, you might have enough time to investigate two approaches.
Now you can investigate ten.
You can pressure test them.
Prototype several.
Ask another model for a different perspective.
Build a blind evaluation to compare the outputs.
Take the winner and push it another three levels.
Then ask what the final version of the system might look like so you don’t accidentally paint yourself into a corner with the first version.
The economics of curiosity changed.
That’s a huge deal.
A lot of work used to stop at:
“We got it working.”
Now you can ask:
“How good can we actually make this?”
Correctly used, AI gives you the ability to take almost everything further.
That creates more opportunities to think, not fewer.
Cognitive atrophy vs. cognitive compounding
This is where I think the real danger sits.
Modern life removed a lot of physical labor.
That’s good.
I’m glad I don’t have to carry rocks around all day to survive.
But removing physical labor had a side effect: if we don’t deliberately exercise our bodies, they deteriorate.
So we invented gyms.
I wonder whether AI is about to create the same problem for our brains.
We are removing intellectual friction at an incredible pace.
You don’t have to struggle through the first draft.
You don’t have to research every source yourself.
You don’t have to remember every command.
You don’t have to manually structure the presentation.
You don’t have to stare at the blank page.
Again, I think this is overwhelmingly good.
But friction was also exercise.
There’s already some evidence that this isn’t just a metaphor. A 2025 study of 319 knowledge workers found that greater confidence in generative AI was associated with less self-reported critical thinking, while AI use shifted cognitive effort away from producing the work and toward verifying, integrating, and overseeing it. The authors specifically raised the risk that routine reliance on AI could reduce opportunities to practice independent problem-solving.^2
And if you consistently use AI to remove the need to understand, question, edit, decide, or reason, I think those muscles will atrophy.
The alternative is cognitive compounding.
And the research isn’t uniformly pessimistic. In one 12-week study of 240 first-year university students, researchers deliberately used AI to offload lower-order writing tasks while requiring students to critique, revise, reflect, and evaluate. Those students showed larger gains in critical-thinking measures than the comparison group. The important variable wasn’t simply whether they used AI. It was how they used it.^3
You use AI to remove the mechanics so you can spend more time understanding the problem.
You use it to challenge your assumptions.
You explore more possibilities.
You build things you previously weren’t capable of building.
You enter new domains and develop real working knowledge of them.
You move faster, so you get more repetitions.
And each round leaves you more capable than the one before it.
Same technology.
Completely different trajectory.
I worry about the 22-year-olds
This matters for everyone, but I’m especially worried about people entering the workforce right now.
Entry-level work is already being squeezed by AI.
A lot of the tasks we historically gave junior employees are exactly the tasks machines can now do extremely well.
That means young people have to become more valuable faster.
And becoming a meat proxy is probably the worst possible response.
If your contribution is taking something your boss said, putting it into Claude, and returning Claude’s response, there isn’t much of a moat there.
The API is cheaper than you are.
The opportunity is to become the person who can take a problem and go dramatically further with it.
Understand it.
Challenge it.
Prototype it.
Ask questions nobody asked yet.
Find the edge cases.
Learn the domain.
Use AI to acquire capabilities you didn’t have six months ago.
That is marketable.
And it compounds.
I’m a dramatically different person than I was at the beginning of this year.
I still think of myself as a product person, but “product person” no longer captures what I can do.
I build.
I prototype.
I work deeply with technical systems.
I create content.
I can move into domains where I previously would have needed someone else just to get started.
Apparently I can even spend an unreasonable amount of time using AI to think through my wardrobe.
None of that happened because I discovered the right prompt.
It happened because I kept using AI to go further.
Use your brain
I don’t think the answer is to draw some hard line around which decisions humans must make forever.
I’m increasingly comfortable letting AI make a lot of decisions.
That line will keep moving.
But I do think you need to understand the thing you’re responsible for.
You need enough context to recognize when something is wrong.
You need enough judgment to know when something is confusing.
And when your work reaches another human, you should be making their life easier, not handing them a pile of machine-generated words and expecting them to figure it out.
So use the AI.
Use the hell out of it.
Give it work.
Let it automate things.
Let it write first drafts.
Let agents operate independently where they’re good enough to do it.
But don’t use it as an excuse to stop thinking.
Ask another question.
Push one level deeper.
Make it explain the thing you don’t understand.
Challenge the answer.
Cut the five paragraphs down to two sentences.
Actually read what you’re about to send.
AI gives us an opportunity to become much more capable humans.
It also gives us an opportunity to become meat proxies.
Your habits decide which one happens.
Footnotes
1. Niklas Gruhn, “Don’t be a meat proxy,” August 3, 2026. Read the original post
2. Hao-Ping Lee et al., “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers,” CHI Conference on Human Factors in Computing Systems, 2025. Read the paper
3. Hui Hong, Poonsri Vate-U-Lan, and Chantana Viriyavejakul, “Cognitive Offload Instruction with Generative AI: A Quasi-Experimental Study on Critical Thinking Gains in English Writing,” Forum for Linguistic Studies, 2025. Read the study
Leave a Reply