Draft: concretize AI-detector-flagged sections (ai-first + one-on-one) - #1954
Merged
Conversation
Re-analysis (same instrument, run on the Open & Async book with finer per-era
buckets) found the kept-ON top signal, AnthropomorphicJustification, conflates
human/collective subjects ("the people doing the work", "deserve credit") with
the real inanimate tell. On the book, splitting by subject cut it from 6.7x
whole-rule to ~2.4x inanimate-only. Adds a note to re-measure the 16x on the
inanimate subset before trusting it, plus per-era corroboration that
FigurativeFalls/Idioms are revived early-2010s habits (0.11/1k in 2010-13).
Docs-only: comments in the advisory config; no rule behavior changes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fast-DetectGPT (Llama-3-8B) flagged abstract/summarizing passages as machine-like. Ground them in specific, named scenarios (keeping the list structure, per the concreteness-not-structure finding): - one-on-one "What belongs in a 1:1": 93% -> 30% - ai-first "What changes for PMs": 97% -> 20% - ai-first orchestra metaphor -> concrete judgment examples: 93% -> 71% Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Second pass on the ai-first post: - 'The judgment is still yours': replace abstract triplet with a concrete launch anecdote (green board, 'lgtm, mostly', held a week): 71% -> 60% - 'What doesn't change': concretize + drop 'The opposite is true'/'not less': 87% -> 46% - 'From async': drop two antithesis reversals: 77% -> 68% Note: section scores all down, but whole-post rose 83->88% because the anecdote adds length and the criterion grows ~sqrt(N). Whole-post % is a length artifact, not a meaningful target; section scores are the signal. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Draft for review — do not merge as-is. Ran Fast-DetectGPT (Llama-3-8B pair) over the 2026 posts; this branch reworks the passages it flagged as most machine-like in two posts.
The core finding
The detector separates concrete/specific prose (reads human) from abstract/summarizing prose (reads machine) — not by structure. Bold-label lists score anywhere from 20% to 97% depending on whether their items are named scenarios or generic categories. So the lever is the same as Ben's own voice principle: replace abstractions with a number, a named tool, a lived moment. List structure is kept intact.
Section scores (Llama-3, before → after)
Important caveat
Whole-post scores barely move (and can rise when an anecdote adds length) because the Fast-DetectGPT criterion grows ~√N. The whole-post % mostly measures length, not AI-ness — it is not a meaningful optimization target. The section scores are the actionable signal, and every edited section improved. These edits stand on their own as voice improvements regardless of the detector.
Notes / open decisions
🤖 Generated with Claude Code