A 2-minute self-check

Meat proxy, or author?

When you pass along what an AI produced, this plots you on two axes: how well you actually understand it, and whether you'll stand behind it.

Built on Niklas Gruhn's essay "Don't be a meat proxy." His idea, his phrase. Read the original first, that is rather the whole point.

Eight situations

Answer for what you would actually do, not what sounds best. There is no visible "right" answer, each choice moves you on two hidden axes.

Your result

—

Understanding →
Blufferowns it, doesn't get it
Authorgets it, owns it
Pipegets neither
Narratorgets it, won't own it
Accountability →
Understanding
0
Accountability
0

The two ends, in one glance

Same tools, same AI. The gap is what you do before you hit send.

Relaying (Pipe)

  • Passes the output as-is and lets the reader sort it out.
  • Cannot answer a follow-up without going back to the model.
  • Ships the plausible-sounding parts along with the wrong ones.
  • Makes the reviewer do the reading it skipped.

Authoring

  • Read every line and cut what did not hold up.
  • Can defend the claim in their own words, without the model.
  • Tested or checked the parts that mattered.
  • Says what they verified and what they are unsure of.

How this is scored

Transparent on purpose. The essay is Niklas Gruhn's; the two-axis model and the type names below are my elaboration of it, not his. No black box, nothing leaves your browser.

The two axes and four types
  • Understanding measures whether you actually comprehend the AI output, enough to explain and defend it without the model in front of you.
  • Accountability measures whether you put your name on it and take responsibility, versus passing it through or hiding behind "the AI said so."
  • Pipe: low on both. The pure meat proxy, a relay.
  • Narrator: understands it, but won't own it. Forwards it as the model's work so no blame lands on them.
  • Bluffer: owns and ships it, but doesn't really understand it. The dangerous one, vouching blind.
  • Author: understands it and stands behind it. The goal.
Why you can't game it

Each option is a realistic behavior, not a labeled "good" or "bad" answer, and each moves you independently on the two axes. The Author choice almost always costs you time or forces you to admit uncertainty, so the only way to score as an Author is to genuinely do that work. Answer honestly and the result is honest.

The field guide

A short norm you can paste into a code-review doc or a team channel. Use AI all you want. Own what leaves your hands.

Before you send AI output to a human
  • Read it. All of it. If you skimmed, you are relaying, not reviewing.
  • Cut it. Delete the padding and the confident filler. Signal only.
  • Check the claims that matter. Run it, look it up, or say you didn't.
  • Rewrite it in your words. If you can't, you don't understand it yet.
  • Say what you're unsure about. Honest uncertainty beats fake authority.
Applied to code review
  • Do not open a PR you could not explain line by line without the model.
  • If AI wrote it, you still author it. The reviewer is not your debugger.
  • Reviewer's rule: if the author cannot defend a change in their own words, that is the review comment.
The one-line version

Be a filter, not a pipe. The value you add is the judgment between the model and the human.