The first time I sent a document back asking which part of it its author actually stood behind, the answer was an uncomfortable silence. The text was impeccable. Structured, with its executive summary and its three recommendations. Nobody had read the whole thing, including the person signing it.
That behaviour has had a name since August 2026. A meat proxy is a person who forwards text, code or any other AI-generated output without reading it, understanding it or validating it. They act as a relay between the machine and the recipient, and they add nothing along the way.
Developer Niklas Gruhn proposed the term on 3 August 2026, and it spread when Simon Willison documented it that same day. Frank Allenby had used the long form, meat-based llm proxies, in March 2026. The interesting part is not the word but that anyone needed to invent it.
A meat proxy forwards AI output without reading or validating it. The cost does not land where it is generated: it lands on the recipient, who is usually more senior, which is why it appears on no dashboard. BetterUp and the Stanford Social Media Lab measured it across 1,150 US office workers: close to two hours per incident and $186 per employee per month. Controlling it is not about banning AI. It is about deciding who signs.
What a meat proxy actually does
They are the cable. The AI produces, the person forwards without stopping at what they are forwarding, and the work of checking whether the thing is true, whether the data exists and whether the recommendation makes sense in this context moves wholesale to whoever opens the email without having asked for that assignment. The definition in circulation is literal on this point: without reading it, without understanding it, without validating it.
It is worth separating it from two things it gets confused with, because it is not using AI to draft, which is sensible and something anyone with a deadline does now, nor is it delegating a mechanical task that nobody in their right mind would do by hand with a tool sitting right there. A meat proxy appears when a question of judgment arrives, the machine answers fluently, and that answer goes out without anyone having exercised any judgment at all.
The distinction matters because it decides where you intervene, and that is what separates a policy that works from one that only produces paper: if the underlying problem were the tool, it would be fixed with a tool policy, which is exactly what many companies are writing right now. It is not.
Why the company should care, not just the recipient
Because the cost transfers upwards and becomes invisible. The person forwarding looks productive, because they dispatched the matter in two minutes and the system records it as dispatched, while the recipient looks slow for spending the afternoon reconstructing what they were sent, and neither impression matches what happened.
There is a measurement. BetterUp, with the Stanford Social Media Lab, surveyed 1,150 US office workers in September 2025 about what they called workslop, AI-generated content that looks good and has no substance. Forty per cent had received some in the previous month, and each incident took close to two hours to resolve, about $186 per employee per month, more than nine million dollars a year on a ten-thousand-person payroll.
The figure I would look at first is none of those. Fifty-three per cent of recipients reported annoyance and 22% felt offended. The hours come back. The signal that an unread document sends about what you think of other people's time takes far longer to repair, and no budget buys it back.
All of this happens while most organisations still fail to extract value from their AI investment. BCG's study of more than 1,250 companies in 2025 sizes it bluntly: only 5% obtain value at scale and 60% obtain no material value despite substantial investment (BCG, September 2025). That implementation gap is not explained by badly chosen technology alone. Part of it is explained by processes where nobody answers for what the technology produces.
Where it shows up when no code is involved
Almost everywhere, and that is the most common diagnostic error. The phenomenon was named among developers, in pull requests and Slack channels, so it has been read as an engineering problem. Engineering is where it gets detected first, because it has peer review and automated tests, and it is not where it does the most damage.
The serious damage happens where there is no pipeline validating anything:
A report to the executive committee with a conclusion nobody can defend when someone asks where the number comes from. An email to a client promising a deadline that existed in no previous conversation. A functional specification that goes down to development with an invented requirement, and gets built in full before anyone notices the business never asked for it. A reply to a regulator in banking or healthcare carrying an interpretation of the rules that the company has validated with no one.
The four have something in common. They pass without friction the only review that existed in that circuit, which was the sender's own, and they pass it precisely because the sender reviewed nothing, so all four reach their destination carrying the company letterhead and the intact appearance of having been thought through by someone.
Broken code falls over. A plausible report gets approved.
What the UK market says about this
The market is already asking for this. According to Shakers' market analysis, 5,366 of 53,093 active UK tech postings mention code review, against 1,548 that mention Copilot (n=53,093, October 2026). Three and a half mentions of the judgment for every mention of the tool that does the writing.
That ratio is the datum, not the volume. Ten per cent of the corpus names the review of what the machine writes, against under three per cent naming the tool that writes it. When the judgment outruns the tool by that margin, the market is not reacting to a fashion. It is describing its floor.
It is worth saying what that number measures and what it does not. It measures mentions in the wording of the postings, not requirements checked one by one, which are different things however similar they look in a spreadsheet, so the honest verb here is mention and any other would claim more than the data supports.
There is a smaller and more precise signal. 262 postings explicitly mention AI-generated code, 0.49% of the corpus. That is small, and it is also the exact niche of this problem: one London posting for an AI engineer role, advertised at up to £300k, states it better than any internal policy I have written when it opens with this isn't a 'vibe coding' role and, in the same breath, demands shipping high-quality code at speed with the latest AI tools and models. It is one case, not a trend. But it is the market drafting the antithesis of the meat proxy in the body of a job posting.
How to control it without banning AI
Banning it does not work. It pushes usage into the shadows and leaves you without traceability exactly when you need it most, with the same volume of unreviewed output and no way to know where it came from. The control is a process control.
The first costs nothing and is the sharpest discriminator: answer in your own words. This is not a matter of style. A summary written by the sender, in their words and not the model's, is the only cheap proof that they have read, understood and validated what they are sending, and it has the property that there is no way to fake it without ending up doing it. Whoever cannot summarise what they are sending has not read it.
The second is to ask about the decision, not the tool. It does not matter whether they used a model; what matters is whether that person can defend why this approach and not another, whether they know the limits of what they are proposing, and whether they considered the obvious alternative, and all of that is checked with one concrete question before approving anything.
The third is to stop treating a green pipeline as the last word: automated tests validate syntax and behaviour, which is a lot and is why we have them, but this problem is one of judgment, and no test measures judgment however well written it is.
And the fourth is to spell out where it matters. Without a matrix like this one, each person improvises their own threshold and the organisation has none:
| Surface | Unreviewed output | Minimum control to demand |
|---|---|---|
| Internal draft, notes, exploration | Acceptable | None |
| Internal communication between teams | Discouraged | First-person summary |
| Client, partner, supplier | No | Identifiable named reviewer |
| Specification going into construction | No | Validation of each requirement's origin |
| Production, contract, regulator | Banned | Signature with name and accountability |
Missing is the part almost nobody instruments: measure at the receiver. The useful metric is not how much a team produces but how many hours your senior profiles spend reconstructing what reaches them. Until that is measured, the cost stays invisible by design, which is exactly how it has worked since it existed.
Who answers when a machine wrote the text
The same person who sent it. That is the entire policy, and it fits in one sentence.
The organisation that does not write it down ends up with a chain of accountability broken at one very specific point: the moment someone presses send without having exercised judgment, and the recipient assumes there was judgment because there is a name in the sender field. Nobody is lying, and yet the system stops having the checkpoint everyone took for granted.
This gets worse with agents. BCG's own study estimates they already represent 17% of the value generated by AI in 2025, on the way to 29% by 2028. The more output the machine produces, the more the chain depends on someone, at some point, answering for it. Work moves from roles to skills and agents, and in that movement the competence that goes up in price is not writing faster. It is deciding what goes out under your name.
That is the link with certified AI-skills talent, which is where you find out whether someone exercises that judgment, instead of discovering it when the report is already in front of the committee. We treat this same displacement, from the code side, in our piece on code review as the bottleneck. And when what gets forwarded unchecked is a software dependency, the problem stops being one of quality and becomes one of security, as happens with slopsquatting.
The question that orders all of this is not whether your team uses AI. It is whether, when something goes out under your company's name, there is someone able to defend it.
Frequently asked questions about the meat proxy
- What is a meat proxy?
A person who forwards text, code or other AI-generated output without reading it, understanding it or validating it, acting only as a relay between the machine and the recipient. Developer Niklas Gruhn proposed the term on 3 August 2026, and it spread that same day when Simon Willison documented it on his blog.
- Does using AI to draft make me a meat proxy?
No. The difference is whether you exercise judgment before sending. Drafting with AI, reading the result, correcting it and being able to defend it is normal work. Forwarding without reading, when what was asked for was a decision, transfers to the recipient the burden of checking whether the thing is true.
- What does it cost a company?
BetterUp and the Stanford Social Media Lab measured it across 1,150 US office workers in September 2025: 40% received content of this type within a month, with close to two hours per incident and about $186 per employee per month. On a ten-thousand-person payroll that exceeds nine million dollars a year.
- Is it only a problem for technical teams?
No, and that is where the usual diagnosis fails. It was named among developers because engineering has peer review and detects it earlier. The greater damage happens where there is no pipeline: reports to leadership, client emails, functional specifications and replies to a regulator.
- Is it controlled by banning the use of AI?
Banning pushes usage into the shadows and removes traceability without reducing unreviewed output. It works better to require a first-person summary of what is being sent, to ask about the decision rather than the tool, and to define in writing on which surfaces unreviewed output is acceptable and on which it is banned.
- Do UK companies ask for the ability to review what AI generates?
The wording of the market does. According to Shakers' market analysis, 5,366 of 53,093 active UK tech postings mention code review against 1,548 that mention Copilot (n=53,093, October 2026). When the judgment outruns the tool by three and a half to one, that is not a fashion. It is the floor of the market.
