A viral blog post names the next AI workplace dysfunction: the 'meat proxy'

The "meat proxy" is the AI workplace dysfunction that turns reviewers into unpaid implementers

The "meat proxy" names a workplace pattern that engineers say is already degrading their code reviews: a colleague pastes raw AI output into a Slack thread, pull request, or decision document without reading it, and the cost of comprehension falls on everyone downstream. The term originated in a blog post published August 3, 2026 at gruhn.me that reached the front page of Hacker News, where a commenter described a colleague asking: "I asked Claude what happened, and it spit out this 300 line response. Can you read it for me and see if it's right?" 1 2.

When someone relays unmodified AI output, they have not added value. They have inserted themselves as a pass-through between the model and the recipient, who must now parse verbose, jargon-heavy text they could have generated themselves with better context control 1. The sharpest example is code review. A developer can paste a ticket description into an AI coding tool, ignore the output, forward reviewer feedback back into the tool, and iterate until the PR passes 1. The author shipped nothing. The reviewers, working through the AI, did the implementation.

Why the meat proxy is a measurement failure, not a discipline problem

The meat proxy is the logical endpoint of a workplace that measures output volume. If a developer's performance is judged by tickets closed and PRs merged, pasting a ticket into an AI tool and forwarding the result is a locally rational strategy. The developer maximizes throughput. The cost is borne by the reviewer.

The arithmetic exposes why this is structural, not behavioral. SmartBear's analysis of a Cisco Systems team established that defect detection in code review degrades sharply past 400 lines of diff 3. Bryan Finster, who writes about lean continuous delivery practices, notes that this ceiling does not rise no matter how many reviewers a team adds 4. A single AI-assisted feature can produce a 600-line diff in one prompt 4. That puts reviewers 50% past the threshold where their ability to catch bugs has effectively collapsed.

The trend is accelerating, not stabilizing. Greptile, which builds an AI code review tool, reports that median pull request size grew 79% in a single year across the codebases it analyzes 5. The 400-line ceiling has not moved. Each percentage point of growth widens the gap between what teams produce and what a human reviewer can meaningfully evaluate.

Greptile's own review data makes the consequences concrete. Across 2.2 million pull requests, 69.5% contained at least one flagged issue. Of those flagged issues, 48% were logic errors, not syntax or style problems 6. Logic errors are the failure mode that reads as syntactically clean. They are the bugs a reviewer operating in a degraded detection zone will rubber-stamp. Finster cites Dave Farley's application of the Nyquist-Shannon sampling theorem to this problem: if defect detection does not exceed the production rate, defects are missed systematically, not occasionally 4. AI has raised the production rate. The human detection rate is capped at 400 lines per session and has not moved.

Here is the cost structure the meat proxy creates:

• The author invests near-zero comprehension. They pasted a prompt, skipped the output, and forwarded it.
• The reviewer must achieve full comprehension of code they did not write, did not prompt for, and have no iteration context for, inside a detection zone that empirical research shows is already degraded.
• The information asymmetry runs the wrong direction. The author held the conversation with the AI and chose not to engage. The reviewer receives the result cold, without the back-and-forth that would explain why the code looks the way it does.

The reviewer is not reviewing. They are implementing. They are doing the comprehension work the author skipped, under conditions worse than the author would have faced.

The fix is not an AI policy memo

Most teams respond to the meat proxy with a written rule: read AI-generated code before submitting it for review. This addresses the symptom, not the metric. If the performance system rewards PRs merged and tickets closed, the incentive to paste and forward has not changed.

The structural fix requires redefining what counts as shipped. The gruhn.me post frames the personal standard: writing a response in your own words is, as the author puts it, a "decent certificate" that you have read, understood, and validated the output before passing it along 1. The team-level version is harder. It means accepting that AI can generate code faster than humans can review it, and that the bottleneck has moved from production to comprehension.

Finster, quoting a LinkedIn post, frames the structural shift bluntly: "AI did not eliminate the coding bottleneck. It just moved it to a different part of the process." His prescription is automated checks that can match the production rate, with human judgment reserved for context no tool can replicate 4.

The practical test is already available to any team. Before a PR enters review, the author should be able to answer specific questions about their own diff. If they cannot, the work is not done. It is a prompt wearing the costume of a finished feature.