find-ai-tells — spot the telltale signs of AI-written prose

Scan prose for the patterns that make text read as machine-generated, report each one with its location and a plain-English revision, and (on request) apply the fixes. The same catalog doubles as a self-check on my own writing.

A “tell” is a heuristic, never proof. Any single one is innocent — delve is a real word, an em-dash is correct punctuation, three items is sometimes just three items. The signal is clustering and mechanical repetition: the same rhetorical move every paragraph, every bullet built **Term:** gloss, a triad in every sentence. The goal is to de-slop — cut the filler and the reflexes — not to ban words or flatten a real human voice.

When this fires

  • “find AI tells”, “find-ai-tells”, “ai-tells”, “de-slop this”, “remove the AI tells”, “make this not sound AI-generated”
  • “does this sound like AI / ChatGPT?”, “was this written by AI?”, “check this draft for tells”
  • As part of any PR/MR review where the diff touches narrative prose — docs, READMEs, lecture notes, textbook chapters, commit/PR descriptions making a substantive claim. Run alongside (not instead of) accuracy review (fact-check-prose) and style review (use-preferred-style).
  • Standing self-check (no invocation needed): before I present any non-trivial prose I authored — a PR/issue description, a commit body, README or doc text, a vignette paragraph, a long chat answer meant as deliverable prose — I first run my draft against the catalog below and cut what I find. Code, terse status lines, and short conversational replies are exempt.

Procedure

  1. Identify the target. A path, a PR/MR (gh pr diff N / glab mr diff N), pasted text, or — for the self-check — my own draft before I send it.
  2. First pass, grep the cheap lexical/typographic tells (catalog below has a ready-to-run command). This finds the mechanical ones fast and cheaply.
  3. Second pass, read for the rhetorical and structural tells that grep can’t see — copula avoidance, participle commentary, antithesis, triads, both-sidesing, hollow conclusions, uniform rhythm.
  4. Apply the deletion test — remove the rhetorical framing and check if any concrete fact remains (a specific name, date, measurement, trade-off, or actionable decision). If nothing remains, the passage has zero information content.
  5. Report (external targets only — skip for the self-check). A table — tell · location (file:line or quoted snippet) · why it reads as AI · suggested revision — followed by a one-line density verdict: is this an isolated word or a pervasive pattern? Don’t cry wolf on a single innocent em-dash.
  6. Offer to apply (external targets) / just fix it silently (self-check). For an external target, on request rewrite in place with Edit, preserving the author’s meaning and voice. When scanning my own draft, skip the report (step 5) and simply cut the tells before presenting.

The catalog

A. Lexical tells (overused words & phrases)

Reflex vocabulary that LLMs reach for far more than human writers:

  • Verbs: delve, leverage, harness, utilize, foster, facilitate, navigate, underscore, embark, unlock, elevate, spotlight, “shed light on”, “dive into”.
  • Nouns/imagery: tapestry, testament, realm, landscape, beacon, journey, treasure trove, plethora, myriad, game-changer, deep dive, “own-goal”, “the whole of it”.
  • Adjectives: robust, seamless, holistic, nuanced, multifaceted, intricate, comprehensive, pivotal, crucial, vital, paramount, vibrant, bustling, cutting-edge, state-of-the-art, ever-evolving, rich (as in “rich history”).
  • Stock frames: “in today’s fast-paced world”, “in the realm of”, “at the heart of”, “when it comes to”, “more than just”, “stands as a testament to”, “plays a crucial/pivotal role”, “about what”.

Quick first-pass grep (case-insensitive, prints file:line). Define the pattern once, then run it through whichever tool is on hand:

tells='delve|leverage|utilize|seamless(ly)?|robust|holistic|nuanced|multifaceted|intricate|tapestry|testament|realm|landscape|beacon|plethora|myriad|pivotal|crucial|paramount|underscore|foster|harness|embark|unlock|elevate|game-?changer|cutting-edge|state-of-the-art|ever-evolving|treasure trove|fast-paced|in the realm of|at the heart of|more than just|shed light|dive in(to)?|deep dive|serves as|stands as|the whole of it|own-goal|about what'
rg -niE "\b($tells)\b" <target>          # ripgrep
grep -rniE "\b($tells)\b" <target>       # no ripgrep — same pattern, via grep

B. Rhetorical tells (sentence-level reflexes)

  • Copula avoidance (dodging “is” and “are”) — refusing plain statements of fact in favor of dressed-up constructions (“serves as a”, “stands as a”, “acts as a”, “marks a”). Replace with “is” or “are”.
  • Trailing “-ing” participle commentary — tacking a participial phrase onto the end of a factual clause to simulate analysis (“…thereby highlighting the importance of X”, “…ensuring seamless alignment”, “…paving the way for”). Cut the participle clause or split into a concrete second sentence.
  • The negation-reversal antithesis — the single biggest tell: “It’s not just X, it’s Y” / “It isn’t about X; it’s about Y” / “This isn’t merely X — it’s Z”. Grep: rg -niE "(it'?s|this is) not (just|only|merely|about)".
  • Formulaic sentence openers — a declarative fronted by an underspecified reference (bare demonstrative “This is”, “That is”, “These are”, “Those are”, or “The one that”), a fronted wh-clause (“Who”, “What”, “Where”, “When”, “Why”, “How”, “Which”, “Whose”), a partitive quantifier (“Some of the”, “Many of the”, “All of the”, “None of the”), or a fronted subordinate clause/preposition (“While”, “Although”, “Despite”, “Because”, “Since”). Split on sentence-final punctuation before matching; a line-start grep sees only the sentences that happen to begin a line: tr '.!?' '\n' < <target> | grep -icE "^ *(this|that|these|those|the one that|who|what|where|when|why|how|which|whose|while|although|despite) ". Fix: name the noun the demonstrative stands for, front the subject and drop the copula, cut “of the”, or lead with the main clause and qualify after.
  • Contrastive closes, copula clefts, and stock metaphors — “rather than” and “not the same as” ending a claim by default, unnecessary copula clefts (“is what”, “is where”, “is when”, “is how”), “carries”/“carry” and “tells”/“doesn’t tell”/“does not tell” standing in for a plain verb, “load-bearing” standing in for “essential”, “the whole of it”, “own-goal”, mid-phrase conjunctives (“, however,”), and indirect “about what”. Each is ordinary English at one or two hits, so count density instead of banning them: grep -ioE "rather than|not the same as|load-bearing|carr(y|ies)|\btells\b|does(n't| not) tell|the whole of it|own-goal|is (what|where|when|how)|\babout what\b|, however," <target> | wc -l. Fix: state the claim positively, use direct verbs (“the lease stops collisions”, not “the lease is what stops collisions”), drop copula clefts, and replace metaphor cliches with plain wording.
  • Convoluted sentences — clause nesting the reader must hold open across two or more embedded clauses. Cue: the ~25-word bar use-preferred-style step 2 sets, or three or more commas plus a subordinator (“which”, “whose”, “because”, “while”, “so that”). Fix: split at the outermost clause boundary. Automated checking: python3 scripts/check-sentence-complexity.py <file> computes sentence word count, subordinator nesting, readability formulas (ARI, Coleman-Liau), and sentence burstiness deterministically.
  • Thin punctuation — long sentences glued with “and”, while commas, semicolons, colons, and parentheses stay rarer than a human writer’s. Count them per 1,000 words and count “and” against them. (The Economist, “How to spot AI writing”, August 2026: across 55,940 sentences, sparse punctuation beat the em-dash as a marker, and only Claude used em-dashes more often than the human writers did — so judge the em-dash tell below per model, and re-check it as models change.)
  • Rule of three, mechanically — triadic lists everywhere (“fast, reliable, and scalable”; “discover, explore, and master”). One triad is rhetoric; a triad in every sentence is a tell.
  • Sweeping range frames — “from X to Y”, “whether you’re X or Y”.
  • Signposting filler — “It’s worth noting that”, “It’s important to note”, “Importantly,”, “Notably,”, “That said,”, “At the end of the day,”, “It’s essential to understand that”.
  • Hedging stacks — piled modals: “may potentially”, “can sometimes help to”.
  • Conversational scaffolding (chat-style) — “Certainly!”, “Absolutely!”, “Great question!”, “Let’s dive in”, “Let’s explore”, “I hope this helps!”, “Feel free to…”.
  • Transition-adverb pileup — “Moreover,”, “Furthermore,”, “Additionally,”, “Consequently,” opening consecutive sentences.
  • Hollow conclusion — an “In conclusion,” / “In summary,” paragraph that only restates, adding nothing.

C. Structural & typographic tells

  • Forced tables and forced structure — tabulating content that is not multidimensional data, or bulleting narrative paragraphs to mimic analytical rigor.
  • Lead construction (title as an object) — opening a document by defining its own title as a noun in the world (“X is a comprehensive guide that outlines…”).
  • Em-dash overuse — multiple — per paragraph used as a default connector. (Em-dashes are correct; the tell is frequency, not presence.) Grep: rg -n '—' <target> and judge density.
  • Bold-leading bullets — every list item shaped **Term:** explanation, applied mechanically rather than where emphasis helps.
  • Emoji section headers / bullets — ✅ 🚀 🎯 🔑 decorating headings or list items in otherwise plain prose.
  • Uniform rhythm — every section the same length, every paragraph 3–4 sentences; conspicuously even, sanded-down structure.
  • Over-signposting — “Let’s take a closer look”, “Now, let’s…”, a heading before every two sentences.

D. Tonal / content tells

  • Promotional register — marketing gloss on neutral material (“a powerful, intuitive solution that empowers you to…”).
  • Reflexive both-sidesing — “on one hand… on the other hand…” balance where the writer should just take a position.
  • Vague universals with no specifics — claims with no names, numbers, or concrete detail; the absence of specifics is itself a tell.
  • Restating the prompt before answering; over-explaining the obvious.
  • False precision — confident, invented-looking statistics.

Reporting format

| Tell | Location | Why it reads as AI | Suggested revision |
|------|----------|--------------------|--------------------|
| "it's not just a library, it's a platform" | README.md:12 | negation-reversal antithesis | "It's a platform, not just a library." (or cut) |
| em-dash ×5 in one paragraph | intro.md:4–9 | em-dash used as default connector | split into sentences; keep one |

Then a density verdict, e.g.: “Pervasive — antithesis + triads in most paragraphs; reads strongly AI. Recommend a rewrite pass,” vs. “One stray ‘delve’; otherwise clean.”

Relationship to other skills

  • memories/ai-writing-patterns.md — the underlying research catalog (20 core empirical patterns synthesized from community research, statistical detection metrics including burstiness and TTR, and multi-tier anti-slop architectures).
  • shared/writing/ai-tells.md — the condensed always-loaded statement of this catalog. Keep the two in step: a tell added there belongs here too, with the grep that finds it.
  • simplify / tidy — the code-side analogues (cut dead code / cruft); this skill is the prose-side analogue (cut filler / reflexes).
  • grade-work — when reviewing a deliverable, fold an AI-tells pass into the quality check.
  • memorize / ums — if a new recurring tell surfaces, add it here and note it; keep the catalog living.
  • find-overlap — the sibling read-only scanner: this skill scans prose for AI tells; find-overlap scans a corpus for duplicated/redundant content. Same detect-and-report posture, different signal.
  • check-info-quality / ciq — another sibling scanner: this skill checks how text is written; ciq checks what it claims (stale, irrelevant, or misleading/out-of-context content). Run both over the same target.
  • fact-check-prose — the accuracy-side sibling: this skill flags prose that reads as machine-generated; fact-check-prose checks whether prose claims and reasoning are actually true. Run both on a review pass.

Anti-patterns

  • Don’t treat any single tell as proof of AI authorship — it’s a heuristic; weigh clustering, not isolated hits.

  • Don’t robotically purge correct words (delve, robust) or correct em-dashes when they’re well-used in context.

  • Don’t rewrite until the text is flat and voiceless — de-slopping removes filler, it doesn’t sand off a real human style.

  • Don’t flag code, identifiers, or quoted source material as prose tells.

  • Don’t skip the self-check on my own draft, then ship prose full of the very tells this skill exists to catch.

  • Don’t replace a figure of speech without first asking what it was hedging. plain-prose already says never to trade an honest hedge for false confidence, and the metaphor case evades that rule because no hedge word is deleted: a simile or an as-though construction hedges through its form, so the “precise equivalent” that replaces it states as mechanism what the original only likened. The tell is a replacement that names a process the source, or the surrounding argument, treats as contested. Keep the comparison, or restate the hedge explicitly (“read as though X”, “on one account X”) — do not let concision decide a live question. (ucdavis/lbt#7, 2026-09-15: a course page’s “read like Hebrew wearing Greek clothes” was replaced with “put Greek words into Hebrew sentence patterns”, asserting the Semitic-interference mechanism that the next class in the same course exists to contest. Caught in adversarial review, and the shipped wording restores the hedge as “read as though Greek words had been fitted to Hebrew sentence patterns”.)

Back to top