Meat Proxy

A meat proxy is the labelled-meme term for the failure mode where a human sits between an LLM and its intended downstream consumer, copying the model's output into Slack/PR comments/code-review threads/feedback boxes without reading, editing, or owning it. The label was introduced by Niklas Gruhn in the essay "Don't be a meat proxy" on Aug 3, 2026 and amplified through X on Aug 4, 2026 — by the time of writing the original post had reached 161.7K views and the term was being used to describe the practice across software engineering, internal-team comms, journalism, and client services. The meme is structurally the inverse of the agentic aspiration: instead of an LLM acting on the user's behalf with proper review, the user acts as the LLM's posting mechanism with no review, the LLM cannot tell what its own output means, and the human cannot either — only the recipient, who now has to interpret a wall of confidently written text from a sender who demonstrably did not write it.

The failure mode has a clean causal chain. It is enabled when an LLM produces plausible-looking output cheaply (low cost per token), incentivised when the human's environment rewards visible turnaround over verified turnaround (a Slack reply now is more legible than a thoughtful one later), and observed at scale precisely in the workplaces where the cost asymmetry is largest — engineering teams with a Claude-Code-adjacent tool, support teams with an AI-draft button, marketing teams shipping a blog post that was generated and lightly edited. The label names what the older productivity literature calls automation bias but adds a cultural layer: the failure is now widely legible as a kind of low-status labour, and the term names both the practice and the participation of the human in the chain, which is why "don't be a meat proxy" reads as advice and not just a description.

Adjacent vocab already in the system that the entry sits near: automation bias (the older literature term — meat proxy is a culture-named subtype), human-in-the-loop (HITL — the framework defending against this failure when it is real and intentional), slop and grindslop (the output-quality side of the same problem — automated output both untested and unread), sycophancy (the model-side bias that makes the failure more common than it would otherwise be), grounding (the architectural counter-measure when the proxy does review at least), unverifiability and output verifier (the stricter versions of the same: every LLM claim mechanically traceable to a primary source). The label is funny because the framing is uncomfortable: the practice is genuinely useful in high-volume low-stakes contexts (a quick acknowledgement reply where the receiver will scan it anyway) and genuinely corrosive in low-volume high-stakes ones (a senior reviewer asked to defend a code-review comment they did not write) — the term forces a context-by-context judgement where one might otherwise ride on inertia.

Tradeoffs in the term itself: the -proxy suffix stretches the conventional reading of "proxy" as a deliberate technical relationship (HTTP proxy, proxy server, proxy voting) into a passive-mistake one, which is part of why the meme lands — the human is implicated as having made the proxy choice. Open questions worth flagging: whether the term will stick in academic HCI / CSCW literature the way automation complacency and AI over-reliance have, or remain a culture-coin; whether the asymmetric naming (no widely-circled term exists for an LLM over-skeptical human) reflects that the practice is the larger failure mode; and whether the next iteration of model UX — explicit provenance markers, citation-by-default, or receipt-style confirmation buttons — will make the relay behaviour less rewarding and the term less needed. The simplest counter to a meat-proxy situation is the simplest test: the next person downstream who asks what the message means should be addressable by name, and the answer should not be "I don't know, the computer wrote it."