ai-tells

AI tells

A tell is an involuntary giveaway. Not a rule violation, a habit that leaks the writer's nature. You are hunting tells, not scoring "AI-ness": detectors are unreliable, humans score near chance on this task, and no single tell is proof. Density is the signal. Three tells in a paragraph is a rewrite. One em dash is nothing.

Distilled from Wikipedia:Signs of AI writing (read 2026-07-24), filtered to what applies to first-person prose. Its wikitext, DOI/ISBN, category, template, and edit-summary sections are Wikipedia-specific and deliberately dropped.

Run order

1. Grep pass — mechanical, catches the lexical tells.
2. Read pass — the structural tells no grep sees (§ Read pass).
3. Rewrite, then re-grep. Fixing one tell routinely plants another: killing repetition invites synonym-churn, killing "serves as" invites "functions as".

Done when: grep is clean or every surviving hit is a named deliberate keep, every Read-pass item has been walked one at a time, and the prose still sounds like its author rather than like nobody.

Grep pass

Two passes, because hard-wrapped prose splits multi-word phrases across newlines — in a file wrapped at ~70 characters, rather than is invisible to a line-based grep about half the time. Run pass B whenever the target is wrapped; skip it only when the file is one paragraph per line.

F=path/to/draft.md   # set to the file under audit

# A1. Model junk markers. Any hit is unconditional deletion.
grep -noE 'contentReference|oai_?citation|oaicite|turn[0-9]+(search|view|image)|attributableIndex|\[cite: |grok_(card|render)|ppl-ai-file-upload|attached_file|:::writing' "$F"

# A2. Lexical tells, as a frequency table. Read the counts, not the list.
grep -nioE '\b(serves? as|stands? as|functions? as|operates? as|boasts?|underscor(es?|ing)|showcas(es?|ing)|emphasiz(es?|ing)|highlight(s|ing)|foster(s|ing)|enhanc(es?|ing)|leverag(es?|ing)|pivotal|crucial|vital role|testament|tapestry|delve|intricate|meticulous|vibrant|robust|seamless(ly)?|enduring|realm of|landscape|navigat(e|ing|ion))\b' "$F" \
  | cut -d: -f2- | tr 'A-Z' 'a-z' | sort | uniq -c | sort -rn

# A3. Vague attribution. Collides with fact-grounding; treat every hit as a sourcing bug.
grep -niE '\b(experts? (say|argue|agree|note)|industry reports?|observers have|some critics|studies show|research shows|widely (held|regarded|considered))\b' "$F"

# A4. Typography. Count em dashes against the deliberate keeps below.
printf 'em dashes: %s\n' "$(grep -o '—' "$F" | wc -l | tr -d ' ')"
grep -nE $'[‘’“”]' "$F"   # curly quotes/apostrophes: match the file's own convention

# B. De-wrapped phrase pass. No line numbers, so it answers "is it in here",
#    then you locate it by searching the phrase.
tr '\n' ' ' < "$F" | grep -oiE \
  "not (only|just|merely|simply)[^.]{0,60}\bbut\b|\b(it|this|that)(.s not|( i)?sn.t) [^.]{0,40}[.,] +it.s|\bnot a [^.]{0,40}, not a\b|\brather than\b|despite (its|the|this)[^.]{0,80}(challenge|obstacle|hurdle)|faces (several|numerous|various)|\bnavigate the landscape\b|in today.s [^.]{0,20}world|it.s worth noting|in conclusion" \
  | sort | uniq -c | sort -rn