Avoid AI Tells

Be specific. Regression to the mean is the root tell. LLM text smooths concrete facts into generic statements that could apply to almost any topic on the same subject. Almost every tell below is a symptom of that. Before worrying about any individual phrase, ask: would this sentence still be true and meaningful if I swapped the subject for a different one? If yes, it's too generic — replace it with a specific fact.

These are descriptive signs, not banned words. LLMs learned them from humans, so humans use them too. The goal is NOT to mechanically scrub every flagged word — that produces stilted text and its own kind of tell (over-correction). The goal is to write naturally and specifically, then check the draft against these patterns and fix the ones that are doing real damage. Context decides. An em dash is not a crime; forty em dashes in one document is.

Self-check loop after drafting: (1) cut puffery and filler, (2) replace vague claims with specific ones, (3) vary structure, (4) strip any conversational/chatbot residue, (5) verify every citation is real.

Puffery — undue emphasis on importance/symbolism

The single most common tell. AI inflates the subject's significance by tying it to broader themes.

• Flagged phrasings: stands as a testament to, plays a vital / crucial / pivotal / significant role, underscores its importance, marking a pivotal moment, a rich tapestry of, leaves a lasting legacy, is more than just, represents a broader shift toward, cementing its place, serves as a reminder.
• Fix: state what the thing is or did, with a fact. Not "the festival stands as a testament to the region's rich cultural heritage" → "the festival has been held annually since 1922 and draws about 40,000 attendees."

Promotional / travel-brochure language

• Flagged: breathtaking, stunning natural beauty, must-see, nestled in, captivates visitors, rich cultural heritage, vibrant, enduring, hidden gem, boasts.
• Fix: drop the adjectives of admiration. Neutral, verifiable description only.

Editorializing and disclaimers

AI tacks on commentary and "cover-your-ass" caveats, especially on medical/legal/safety/sensitive topics.

• Flagged: it is important to note that, it is worth noting, it should be noted, no discussion would be complete without, it's essential to understand, one must consider.
• Fix: if the point matters, just state it. If it doesn't, cut it.

Vague attributions (false consensus)

• Flagged: some critics argue, many experts believe, observers have noted, industry reports suggest, it is widely regarded as, studies show (with no study). This is also where source-to-text integrity breaks — e.g. presenting one reviewer's opinion as a consensus.
• Fix: name the source, or cut the claim. One named critic beats "some critics."

Superficial "-ing" tail clauses

Sentences that end with a tacked-on present-participle clause editorializing about significance.

• Flagged: ..., highlighting the need for..., ..., reflecting a broader trend..., ..., emphasizing its role in..., ..., showcasing..., ..., underscoring..., ..., paving the way for....
• Fix: end the sentence at the fact. Delete the trailing interpretation.

Negative parallelisms

• Flagged: not only X but also Y, It's not just X, it's Y, It's not X; it's Y. AI uses this to manufacture contrast, often "correcting" a misconception nobody held.
• Fix: use sparingly if at all. Make the positive claim directly.

Rule of three

Reflexive three-item lists and tricolons (fast, reliable, and affordable). Humans use it; AI overuses it to the point of rhythm.

• Fix: vary list length. Use two, or four, or recast as a sentence when three isn't real.

Conjunction / transition overload

• Flagged: opening sentences repeatedly with Moreover, Furthermore, Additionally, In addition, However, On the other hand.
• Fix: most can be deleted outright. Let sentences connect by content, not by signposting.

Section summaries / conclusions

• Flagged: In summary, In conclusion, Overall, a closing paragraph that restates everything. Common in essays/textbooks, but a strong tell in reference prose and anywhere a summary wasn't asked for.
• Fix: don't append a conclusion unless the format genuinely calls for one. End on substance.

Overused vocabulary (shifts by model era)

Treat as a watch-list, not a ban-list — flag clusters, not single uses.

• GPT-4 era: delve, tapestry, testament, intricate / intricacies, interplay, pivotal, meticulous, vibrant, landscape, underscore, bolster, garner, boasts, crucial, enduring, foster, realm, navigate (figurative).
• GPT-4o era: align with, enhance, fostering, highlighting, showcasing, underscore, vibrant.
• GPT-5 and later: emphasizing, enhance, highlighting, showcasing — plus a newer cluster of undue emphasis on notability and media coverage: has been profiled in, featured in national [media], standalone "Media coverage" / "Online presence" sections, cited as an example of. delve is effectively dead; underscore is fading.
• Fix: prefer the plain word. delve into → examine; leverage → use; navigate the complexities of → handle.

The "Challenges" / "Future Prospects" formula

A rigid end-of-document section that opens "Despite its [praise], [subject] faces challenges..." and closes with a vaguely upbeat outlook or a call to action. The tell is the formula, not the mention of challenges.

• Fix: don't bolt on speculative "future outlook" or "challenges" sections unless asked. Cover difficulties where they're relevant, with specifics.

Rigid, formulaic outlines

Identical section shapes, every paragraph opening the same way, uniform paragraph length. Often a symptom of "generate an outline, then fill each heading."

• Fix: vary paragraph length and openings. Let structure follow the material.

• Excessive boldface — bolding product names, key terms, and phrases mechanically throughout. Fix: bold rarely, for genuine emphasis only.
• Title Case In Headings where sentence case is the house style. Fix: match the surrounding convention; default to sentence case.
• Over-listing — bulleted/numbered lists where prose is expected. Fix: use lists only when the content is genuinely enumerable; otherwise write paragraphs.
• Leaked Markdown — literal **bold**, ## headings, or - bullets pasted into a context that doesn't render Markdown. Fix: match the target format exactly.
• Em dash overuse — not wrong in itself, but a dense scatter of — is a recognized signature. Fix: keep them purposeful and sparse; commas, colons, or periods often work.
• Curly quotes / apostrophes (" " ' ') and stray emojis in headers or list items, when the surrounding text uses straight quotes and no emoji. Fix: match the document's existing punctuation style.

Conversational artifacts that should never survive into a finished document:

• As an AI language model..., I cannot... / refusal text, As of my last knowledge update / knowledge-cutoff disclaimers.
• I hope this helps, Certainly!, Sure!, Let me know if you'd like..., Here is the... framing, Would you like me to....
• Unfilled placeholders: [insert X here], [Your Name], [Company], [date], fill-in-the-blank prompt instructions left in the text.
• Fix: the deliverable addresses its reader, not the person who prompted the model. Strip every trace of the chat frame and fill or remove every placeholder.

The real danger isn't style — it's fabricated facts in confident prose.

• Hallucinated references that look plausible but don't exist; broken links; fake DOIs; ISBNs with invalid checksums; a cluster of 404s not findable in the Internet Archive.
• Citations that exist but don't actually support the claim attached to them.
• Fix: never invent a citation. Only cite sources you can verify exist and that genuinely support the claim. If you can't verify, say so or omit the claim. Flag any statistic/quote/date you're not certain of rather than presenting it as fact.

• Don't robotically purge every em dash, every "however," every list — natural writing uses all of these.
• Don't mangle grammar or add typos to "seem human."
• Don't strip specificity to dodge a flagged adjective; specificity is the cure, not the disease.
• Remember the false-positive reality: skilled human writers, and especially newer/non-native editors, legitimately use many of these patterns. These checks improve writing on their own merits — clearer, more specific, less padded — which is the point regardless of detection.