Don't be a meat proxy

Original Article ↗ Hacker News Discussion ↗

Generative AI is increasingly being used as a crutch in workplaces, with people pasting raw LLM output into chats, docs, and code reviews instead of thinking, validating, and communicating in their own words. Commenters describe this as turning humans into “meat proxies” that merely relay AI responses, offloading the hard work of verification and understanding onto colleagues and degrading code quality, documentation, and trust. Many argue organizations need explicit norms or policies—such as “if you ask for human attention, show human effort”—to ensure AI remains an assistive tool rather than a way to shirk responsibility.

Core complaint: “Meat proxies” and slop grenades

  • Many describe coworkers pasting raw LLM output (often prefixed with “Claude said…”) into chats, PRs, docs, or tickets.
  • This offloads the hard part—understanding, verification, and summarization—onto others.
  • People liken it to throwing “slop grenades,” “BI slop,” or being a “very expensive keyboard.”
  • Brandolini’s law is cited: generating bullshit is cheap; refuting it is expensive. LLMs radically amplify this asymmetry.

Impact on work quality and roles

  • Senior engineers report being reduced to AI babysitters or rubber‑stamping unread AI PRs.
  • Some see “agentic engineering” loops (error → paste to AI → retry) as shallow, with little real understanding.
  • Others say LLMs let good engineers focus on design and glue work, and can lift weak ones to a minimum bar.
  • There’s concern about “vibe coders” and entire orgs losing foundational skills, creating future maintenance crises.

Etiquette and policy proposals

  • Suggested norms:
    • “If you are asking for human attention, demonstrate human effort.”
    • Never send raw LLM output; provide a human-written TL;DR and your own judgement.
    • Make “meat proxy” behavior socially costly (ignoring such messages, asking “Did you read this?”).
  • Some orgs add handbook rules, require “attention requests” with questions about docs, or propose AI Codes of Conduct.
  • Debate over consequences: some advocate making repeat offenses fireable; others say culture and management incentives are the real problem.

When LLM relaying is acceptable

  • A minority argue forwarding AI answers can be fine when:
    • The question is trivial and the asker clearly didn’t try; akin to “let me Google that for you.”
    • An expert prompts once, carefully vets a large doc, then shares it instead of everyone querying separately.
  • Even then, many insist the sender must have read and understood the output.

LLM style, jargon, and cognitive load

  • Complaints that newer models are more terse, tense, and jargon‑dense, making text harder to parse.
  • Some mitigate via custom instructions (ELI5, Simplified Technical English, “sound kinda dumb but be correct”).
  • Others note hallucinations are less frequent but still real; verification and domain knowledge remain essential.

Broader concerns: laziness, intelligence, and culture

  • Long subthreads debate whether technology (LLMs, internet, short‑form content) is driving declining attention, IQ, and “reverse Flynn effect.”
  • Some foresee an “intelligence caste”: those who offload thinking to AI vs those who keep exercising their brains.
  • Others argue human+machine capability is what matters; historical tech (writing, calculators, compilers) also shifted which skills are practiced.