Are We All Meat Proxies Now?

In August the insult was cute. Someone pasted a Claude brick into Slack, a coworker replied "don't be a meat proxy," and the thread moved on.

In October the dictionary card is trending because people are not looking up a joke. They are looking up the job they already have.

We are staring at the wrong villain again.

The villain is not "using AI." We vibe code here. The villain is a workplace that pays for throughput of unread tokens and then acts surprised when the humans in the middle stopped being colleagues. They became cables. The model on one end. A tired recipient on the other. A person in between who never opened the envelope.

This is an opinion piece. The definition is not. We wrote the slang explainer in August. The point of no return essay is the bigger claim: skills you stop practicing erode, and a pause of the frontier does not give them back. This post is the smaller, nastier question that post already answered in one sentence and then kept walking:

Are we all meat proxies now?

Read it and tell me I'm wrong.

Update — October 2, 2026: Tavus published Griffin, a video-to-video Human Interaction Model. After a one-minute call, 26 of 54 participants thought the partner was human — and Griffin-Lite is testers-only because that property is deception. Numbers, VideoFDB scores, and what still runs in production.

TL;DR — the questions this post answers

What a meat proxy actually is

A meat proxy is not a person who uses a model. It is a person who relays the model.

On August 3, 2026, Niklas Gruhn published Don't be a meat proxy. The scene is not science fiction. It is Slack, a pull request, and a WhatsApp argument:

Claude said: [giant response verbatim]

Please don't do this. [...] I can talk to Claude myself. It's going to be faster and I get to control the context. I don't need a meat proxy in between.

He named the cost of the unread dump: AI output is extra work for the recipient — verbose, full of plausible nonsense, jargon-dense. His example sentence from Claude: "NATS control-plane events: stream leader election / R3 quorum re-form during pod churn." He had to look up almost every word. That is not help. That is exporting the reading.

Then the code-review version, which is the part that should scare a staff engineer:

Shipping some code can be done with close to zero effort now: Copy/paste the ticket description into Claude Code. Don't look at the code or read what Claude has written. If there's any feedback from reviewers, copy/paste that into Claude Code as well. [...] That works. But who has done the implementation? The reviewers did, using Claude Code, and you as a meat proxy.

Simon Willison called it "an excellent new term" the same day and quoted the four-step rule. Honesty about coinage: Willison's update notes Frank Allenby used "meat-based LLM proxies" in a similar sense in March 2026. Gruhn is still the essay that named the Slack crime and stuck. We published the etiquette explainer so the slang had a home on explainx.ai. The dictionary card is the lookup surface that is trending now.

The four steps are still the whole product:

1. Read it.
2. Understand it.
3. Validate the part that would hurt if it were wrong.
4. Write it in your own words.

That last step is not style. Gruhn called the rewrite "a decent certificate that you've done the prior steps." Skip it and you are not a colleague with a take. You are TCP.

This is not vibe coding. Vibe coding is how you generate. It is not workslop. Workslop is polished and still useless. It is not cognitive debt. Cognitive debt is what you owe yourself when you cannot reconstruct the code next month. A meat proxy is what you owe other people this afternoon.

If you only remember one row, remember the last one. Generation got cheap. Trust did not.

The question changed

In August you could still treat this as manners. Don't paste unread. Be a grown-up.

By October the dictionary is trending because manners lost to incentives.

We already described the end state in the point of no return piece: a workplace of agents and meat proxies, humans as network cables, tokens as the visible job. Tokenmaxxing made that sentence operational. If the dashboard ranks you by burn, the person who sits with a hard problem for an hour looks idle. The person who pastes looks busy.

The numbers are not vibes. Meta's internal "Claudeonomics" dashboard ranked roughly 85,000 employees by token usage, with one top user reportedly burning around 281 billion tokens in 30 days. Ramp's customer data showed average monthly AI token spend up 13× since January 2025 (our breakdown). Tesla capped spend at $200 per employee per week. Meanwhile METR-class coding studies keep landing nearer 2× than 10×. The leaderboard does not care.

So no: we are not "all" meat proxies in the sense that every human forfeited judgment on the same Tuesday. Yes: we built systems that punish the people who still do the reading. That is a quieter claim and a worse one. Extinction debates still get the cameras. This gets the sprint.

The receipts — this is not just slang

Opinion is allowed. Unsourced opinion is how this topic dies in a quote-tweet. Here is the evidence that unread relay is the cognitive pattern, not only a manners complaint.

The MIT number is the private version of Gruhn's Slack crime. You cannot quote the essay. You also cannot explain the paste. Teachers called it soulless. Coworkers call it "Claude said." Same gap. Different room.

None of those papers use the words "meat proxy." They did not have to. They measured output without ownership. Gruhn named the social move that turns that measurement into a coworker problem.

Code review is where the joke stops being a joke

The PR is the cleanest lab.

You open a 400-line agent diff. You cannot explain why this helper exists. Review comments bounce straight back into the model. The reviewer does the implementation using your agent, with you as the proxy between their notes and the weights.

We already asked whether you should review every line of AI-generated code. Risk-based review is a real answer. "Never read it" is not. It is how a team discovers that review was the only place design still happened — the same de-skilling pattern we keep finding when juniors ship fast and cannot debug.

If the outage comes at 2 a.m. and the only person who understands the service is a vendor model with a rate limit, you do not have an engineering team. You have a subscription with salaries attached. The meat proxy is how that company talks to itself during daylight.

"Are we all" is the wrong grammar — "is the job" is the right one

Individuals still choose. I use AI every day. Quitting is not the point. Which thinking you refuse to hand over is the point. Our how to use AI guide already made "never forward what you haven't read" rule 7. This post exists because rule 7 is now the one workplaces try to write out of the job description.

Ask the question at three scales:

Parents already know the child version. Bastani's PNAS field experiment is the clean split: unrestricted GPT-4 during practice, then the exam without it, and scores fall. A tutor that only offered teacher-designed hints did not do that damage. The adult version is a standup that treats "I asked the agent" as a status. Same mechanism. Older humans. Higher payroll.

Ethan Mollick's split still holds: deskilling yourself on annoying tasks is fine; deskilling a whole team is not. Meat-proxy culture erases the split by making the annoying task reading.

The career-ladder joke was not a joke

In September the slang grew a satirical org chart: LLM at the top, then meat proxy, wrangler, principal, director, VP of proxy infrastructure, CTO. The reply that landed was "please put human at top." Uncomfortable because it was a one-notch exaggeration of a real reporting line.

If your company already measures tokens, you do not need the meme. You are living the chart. The model sits above the humans because the metric sits above the humans. Change the metric or admit the job.

What to do before the paste becomes the culture

I am not telling you to type everything by hand. I am telling you the default settings of the modern stack are built for volume, and volume without a certificate is a relay.

1. Kill the token leaderboard. Measure shipped work, defects, cycle time. Never tokens. We have watched this disease long enough.
2. Own the diff. You do not have to retype every line. You do have to answer why this approach, what breaks if this constant changes, which test fails if the model invented a helper that already exists.
3. Rewrite before you send. Slack, email, docs, strategy. If it would embarrass you to read aloud, it is not ready.
4. Label unverified transcripts. That is HAPI — openly the interface — not a meat proxy pretending the paste is a human answer.
5. Run AI-off drills. Once a month, ship something small with no assistant. If you cannot, the usage pinch is not a pricing problem. It is a skill problem with a pricing problem.
6. Keep a fallback skill and a fallback vendor. Limits move when a lab is ahead. Dependence is how they get priced.

If you want structured practice on the "use the tools and still understand the line" side, that is how we run live workshops.

So — are we all meat proxies now?

Not all. Enough that the dictionary is trending for a reason.

If you still read the output, you are not one. If your team still rewards the person who caught the bad number, you are not a proxy farm. If a child in your house still has to get stuck and get unstuck without an answer machine, they are not doomed to the paste.

The point of no return is when those sentences stop being choices and start being nostalgia. A meat proxy is how you get there one unread forward at a time.

The slang was a warning. Treat it like one.

Related reading

• Meat proxy: don't forward AI you haven't read
• AI is taking us to a point of no return
• How to use AI: the fundamentals
• What is tokenmaxxing?
• Cognitive debt: retype LLM code?
• Should you review every line of AI-generated code?
• When answers get cheap, trust is the job
• Does AI make you dumb?
• 2x, not 10x: coding with LLMs
• Sources: Gruhn, Don't be a meat proxy · Willison, Aug 3 · MIT Media Lab, Your Brain on ChatGPT · Anthropic, AI assistance and coding skills · Microsoft Research / Carnegie Mellon, 319 knowledge workers · Bastani et al., PNAS

Opinion piece, October 1, 2026. Definition and quotes follow Gruhn (August 3, 2026). Willison's note on Allenby's March 2026 phrasing is included so coinage is not over-claimed. Study figures match the sources cited here and in the linked explainx.ai research posts. Follow @explainx_ai.