This is the reference sheet I use to review my own drafts for machine-written prose. Note that almost nothing described here is wrong in isolation; the tests in §1 are what turn a word into a finding. Moreover, today’s AI tells may not be the same as tomorrow’s. With those caveats, I present to you my “AI Tells Reference Sheet”:
1. Judgment tests
| Test |
Question |
| Density |
How often does it recur? Ignore one “robust” in 2,000 words; flag three in a paragraph. |
| Precision |
Is the word doing specific work, or inflating? |
| Voice |
Would this author plausibly have written the sentence? |
| Cadence |
Is the tell a repeated form (not a repeated word)? Count it across the whole piece; report the pattern once |
| Literalness |
Is a metaphor standing in where a plain phrase was available? Anthropic’s September 2026 prompting guide calls this “mannered prose” (“a dial worth turning” for “a parameter worth varying”) and says Claude produces it unprompted. Most of §2 fails this test |
Severity: High = hard tells (contrast scaffold, “delve,” “impact” as a verb, a confirmed cadence pattern, anything in §2). Medium = defensible words in telltale clusters. Low = borderline, worth a look.
A clean text proves nothing. Only one category (§9) is conclusive on a single hit.
2. Always flag
| Pattern |
Fix |
| “actually,” “genuinely,” “honestly,” “frankly” (intensifiers) |
Cut |
| “carries” (weight, meaning, significance) |
Show what it does |
| “deploy” (outside software/military) |
“use,” “apply” |
| “does real work” |
Say what it accomplishes |
| “earns” (its place, its keep) |
Show the value or cut |
| “impact” as a verb |
“affect,” “change,” or name the effect |
| “is real” (“the tension is real”) |
State the consequence |
| “is worth” (noting, considering) |
Make the point |
| “load-bearing,” “does the heavy lifting” |
Show what it does, or cut |
| “matters” (“context matters”) |
Name the stakes |
| “the move” / “a move” (metaphorical) |
Name the action |
3. Vocabulary
Word tells date. The 2023–24 set is weaker evidence now; the mid-2025-onward set is quieter and current.
Current generation (2025–26)
| Pattern |
Fix |
| “quietly” (building, transforming); “deeply,” “fundamentally” |
Cut; if unannounced, say who missed it |
| “shape” (the future, how we think) |
Name what changed and why |
| “shift” (in thinking, toward) |
“change,” or the specific verb |
| “signal” (send a signal, a strong signal) |
Say what was communicated, to whom |
| “compound” (decisions, benefits) |
“add up,” or the rate |
| “land”/“lands” (the argument landed) |
“works,” “persuades” |
| “pull” (the pull of, feel the pull) |
Name the incentive or person |
| “hold” (space, the line) |
Say what the person does |
| “emphasizing,” “enhance,” “highlighting,” “showcasing” |
“stress,” “improve,” “show” |
| “the work” (do the work, trust the work) |
Name the task |
| “every time” (as emphatic tag) |
State the rule once |
| “you already have” / “you already do” |
Name the practice and how it transfers |
| “built different” / “built to last” |
Name the design choice |
| “significant,” “increasingly,” “consequences,” nominalizations |
Short word; verb form (“grew,” not “experienced growth”) |
Persistent (all eras)
| Cluster |
Examples |
Fix |
| Importance inflation |
crucial, vital, pivotal, essential, key (insight/takeaway) |
Demonstrate or downgrade |
| Hype |
game-changer, transformative, revolutionary, seamless, robust (outside engineering), elevate |
Describe the change |
| Corporate verbs |
leverage, streamline, empower, harness, unlock, foster, ensure (clustered), synergy |
“use,” or the mechanism |
| Quest framing |
journey, embark on, navigate, dive into / deep dive |
“process,” “start,” “handle” |
| Spatial filler |
in the realm/world/landscape of, “the AI landscape,” at its core |
Cut |
| Quantity filler |
myriad, a plethora of, a wide range of, comprehensive, holistic |
“many,” or the number |
| Complexity praise |
intricate, nuanced, multifaceted |
Name the complication |
| Discourse verbs |
underscore, highlight, illuminate, shed light on, explore (as filler) |
“explain,” “show,” or state the thesis |
| Padding |
serves as, stands as, that being said, to put it simply, additionally/furthermore/moreover (stacked) |
“is”; “But”; cut |
| Audience flattery |
“whether you’re a X or a Y” |
Address the reader |
| Legacy language |
watershed moment, indelible mark, lasting legacy, continues to captivate, a testament to |
Concrete events and dates |
Dated (2023–24, still catches pasted or imitated text)
delve, tapestry, intricate, meticulous, boasts, testament, vibrant, “in today’s fast-paced world.”
4. Constructions
| Pattern |
Example |
Fix |
| Contrast scaffold |
“It’s not X. It’s Y.” / “not just X, but Y” |
Assert Y; at most one per piece |
| Negative parallelism |
“No fluff. No filler. Just results.” |
One plain sentence |
| Antithesis overload |
“not only/but also,” “on the one hand/on the other,” more than twice |
Vary sentence logic |
| Neat juxtaposition |
Mirror-image clauses (“catches what the tool got wrong, not what it never saw”) |
Keep one; make the second clause add information |
| Rule-of-three padding |
“clear, concise, and compelling” |
Cut to distinct items |
| Compulsive counting |
“three reasons why,” “the four things that matter” |
Drop the count |
| Trailing participial tack-on |
“…, highlighting the importance of X” / “…, reflecting a broader trend” |
Delete, or promote to a supported sentence |
| Comma-clipped tail |
“It’s a good rule, in theory.” / “They shipped it, eventually.” |
Fold in or commit |
| Question pivot |
“The result? …” / “The catch? …” |
State the result |
| Suspense pivot |
“Here’s the thing.” / “And here’s where it gets interesting.” |
Cut |
| Engagement bait |
“Let that sink in.” / “Read that again.” |
Cut |
| Abstraction as equation |
“The tension is the point.” / “That’s the real question.” |
Ask the question or point at the thing |
| Invented concept label |
“the supervision paradox,” “workload creep” |
Replace with the claim; support it |
| Unnamed-truth claim |
“the part nobody acknowledges,” “to name it plainly” |
State the point |
| Unearned universal |
“nobody talks about this,” “everyone gets this wrong” |
Specific, defensible claim |
| False range |
“from innovation to cultural transformation” |
List the items |
| Despite-challenges formula |
“Despite criticism, X remains widely respected.” |
Specifics, or pick one half |
| Both-sidesing |
“While X has advantages, it also has drawbacks” |
Take the position the evidence supports |
| Compliment sandwich |
“Great point! That said…” |
Say what’s wrong |
| Vague authority |
“Experts argue,” “Observers have noted” |
Name the source or drop it |
| Unquoted sources |
No named person quoted anywhere in a piece that gestures at expertise |
Quote someone, or claim it in your own voice |
| Second-person validation |
“You’re not just X, you’re Y.” / “Are you ready to go deeper?” |
Cut, or praise something specific |
| Profundity closer |
“It’s less a map than a compass.” / chiastic paradox as kicker |
End on the finding |
| Hedging stack |
“It’s important to note that, in some cases, X may potentially…” |
One hedge max. Style only; never evidence of authorship (corpus evidence is contested) |
5. Cadence and structure
Invisible sentence by sentence. Read structural positions across the whole piece and count.
| Pattern |
What to count |
Fix |
| Short emphatic topic sentences |
Paragraph openers like “Speed wins.” / “It worked.”; one-sentence dramatic paragraphs. Two or more = High |
Keep the strongest; vary the rest (fold, demote, let substance lead) |
| Uniform paragraph rhythm |
Every paragraph 3–4 mid-length sentences |
Vary deliberately |
| Fluff sandwich |
Generality / content / restated generality |
Keep the middle sentence |
| Paragraph-final restatement |
“Ultimately, …” / “In short, …” |
Cut |
| Anaphora abuse |
“It’s about X. It’s about Y. It’s about Z.” |
Keep one; recast the rest |
| Synonym cycling |
tool → platform → solution → offering |
Reuse the right term |
| Dead metaphor |
One metaphor revisited in every section |
Keep the first use; go literal after |
| Announce-then-answer |
“This piece will examine…” / “Having established X, …” |
Start with the first claim |
| Signposting |
“more on that below,” “as we’ll see” |
Delete |
| Tie-back |
“as mentioned above,” callback to the intro |
Restate the substance in a clause or delete |
| Stakeless competence |
No names, numbers, dates, tools, prices; flawless grammar in casual register |
Author supplies detail |
6. Punctuation
| Signal |
Detail |
Fix |
| Em-dash overuse |
Readers assume AI on sight (the “ChatGPT hyphen”), so it costs credibility regardless of author. As authorship evidence it has collapsed: in the Economist’s 2026 corpus only Claude out-dashed humans; ChatGPT used fewer; GPT-5.1 suppresses them by instruction |
Comma, semicolon, colon, or new sentence. Never space an em dash |
| Punctuation scarcity |
Fewer commas and semicolons than human prose, almost no parentheses; long compounds joined by “and.” Economist: “and” is the single most overused word across all four models; thin punctuation beats any word as an indicator |
Break compounds; subordinate the subordinate clause |
| Contrast starvation |
Opposing clauses joined by “;”, “and,” or “while,” almost never a coordinating “but.” Zero or one “but” in a long piece is the signal |
Write “, but” where the logic turns |
7. Formatting
| Pattern |
Fix |
| Bullet bloat (prose chopped into fragments with bolded lead-ins) |
Recombine into prose; bullets only for true lists |
| Bold-itis (mid-sentence emphasis bolding) |
Unbold; rewrite for emphasis |
| Colon-subtitle headers (“Precision: Why Word Choice Matters”) |
Plain descriptive headings |
| Title Case headings; Wh-word headings (“What This Means,” “Why It Matters”) |
Sentence case; name the content |
| Emoji decoration (🚀 bullets, ✅/❌ lists, emoji-capped headers) |
Delete |
| Unicode decoration (→, box-drawing, decorative bullets) |
Words for relationships; ordinary list markup |
These are era tells. By Anthropic’s own account, Fable 5.1 uses bold less and reaches for headers and lists less often than earlier Claude models. Bullet bloat and bold-itis still catch 2024 and 2025 output; as evidence against current Claude they are weak, as the em dash became.
8. Openers and closers
| Pattern |
Fix |
| “In today’s fast-paced world / digital age / ever-evolving landscape” |
Start with the subject |
| “Let’s dive in” / “Let’s unpack this” / “Let’s explore” |
Begin |
| “It’s important to note that” / “It’s worth noting that” |
State the point |
| “Keep in mind that” / “Remember that” |
Cut |
| “At the end of the day” |
Cut or state the conclusion |
| “In conclusion,” / “In summary,” (short pieces) |
End on the last point |
| “By following these steps, you can…” |
Cut |
| “…and that’s exactly the point” / “…and that’s okay” |
Cut |
| “I’m excited / honored / thrilled to announce” |
State the news |
| “I hope this email finds you well” |
Open with the reason |
| “I hope this helps” / “Great question” / “Would you like me to…?” (in a deliverable) |
Cut; assistant-voice leakage |
9. Paste-through artifacts (conclusive)
Exact strings no human types. One hit settles authorship. Current as of August 2026; a miss tells you nothing.
| Artifact |
Source |
contentReference, oaicite, oai_citation, turn0search0, attributableIndex |
ChatGPT |
[cite: 1], [span_1](start_span) |
Gemini |
grok_card, grok_render_citation_card_json |
Grok |
| Stray lenticular brackets, dagger symbols |
DeepSeek |
attached_file, ppl-ai-file-upload |
Perplexity |
:::writing |
Unclassified |
utm_source= on pasted links |
Any model with web search |
10. Model tendencies (2026)
Economist corpus, August 2026, except where noted.
| Model |
Tendency |
| Gemini, Claude |
Polysyllables, rare words, nominalizations |
| ChatGPT, Claude |
Most “not X but Y” scaffolds and rule-of-three triplets per 1,000 sentences |
| Claude |
Only model to out-dash human writers |
| ChatGPT |
Markedly fewer em dashes than humans |
| Grok |
“causal,” “empirical,” “correlate”; inverts negative parallelism into “X rather than Y” |
| Claude Fable 5.1 (Anthropic prompting guide, September 2026) |
Fewer stock phrases than earlier Claude; denser prose, longer sentences, fewer paragraph breaks; mannered metaphor (“earns its keep”); less bold, fewer headers and lists; reproduces source passages without quotation marks when summarizing |
11. Review procedure
- Grep §9 strings first.
- Grep §2, §3, §8 as candidates; judge each on §1 tests.
- Close read for §4 constructions, paragraph by paragraph.
- Second read for §5–6: paragraph openers as a set, “but” count, “and”-chained compounds.
- Sort findings High / Medium / Low; report cadence patterns once, with instances listed.
- Run your own replacement text against the same tables.
- If a model drafts the replacement, put Anthropic’s mannered-prose paragraph in the prompt. The short form, “Please remove all mannered prose,” also works.