A draft lands in your review queue from a freelancer, an agency writer or a drafting agent, and it reads clean but hollow. Every paragraph closes on a neat summary, and three sentences in a row run the same length. The word "delve" turns up in the second line. Your readers have learned this pattern as well, and once they spot it they stop trusting the argument underneath. That costs you the authority the piece was supposed to build.
Your team can learn to catch these tells and edit them out before anything ships. Treat each marker as a probabilistic signal. A marker shifts the odds that a model produced a passage, and no single marker, or even a stack of them, proves who wrote it. That limit is also why AI generated text detection tools still misfire on human writers.
The main markers of AI-generated text
AI text tends to give itself away in five layers, and the tells carry more weight when several layers show them at once. An editor should scan for these groups:
- Lexical: Overused style words such as "delve," "underscore" and "showcasing," stock hedges, and transitions like "furthermore" stacked at the start of sentences.
- Syntactic and rhythmic: Sentences that cluster in one length range, predictable clause order, and participial phrases bolted onto the end of a sentence.
- Structural: Lists of three, symmetrical paragraphs, "it's not X, it's Y" contrasts, and conclusions that restate the introduction.
- Tonal: Even, mildly positive, emotionless prose with no visible author behind it.
- Content-level: Generic claims, invented specifics, and citations that point to papers that don't exist or don't say what the text claims.
A single marker is weak evidence, because skilled human writers produce every one of them. A draft that shows tells in four or five layers at once deserves a closer read.
How perplexity and burstiness reveal AI text
Perplexity measures how predictable a passage looks to a language model. Burstiness measures how much that predictability, or the sentence length, swings across the document.
- A model generates text one token at a time. The model assigns a probability to every possible next word and picks from the likely ones. Run a finished passage back through a model and you can score how surprised it is at each word. Low perplexity means the wording is mostly what a model would have chosen itself. People reach for odd words, slang and sudden turns, so human text usually scores higher.
- Burstiness picks up rhythm. Human writers put a four-word sentence next to a forty-word one and let complexity rise and fall. Models hold a steadier pace. A study comparing English news articles with output from six LLMs found human sentence lengths more scattered, while the models bunched into a narrower band.
These two measures shaped early AI generated text detection. GPTZero says it pioneered perplexity and burstiness scoring, then moved to a deep-learning architecture in autumn 2023 and now trains a custom model on human and AI text to look past those two factors. Today's commercial detectors are trained classifiers, but what they learn still traces back to the gap between predictable machine word choice and messier human choice, so a plain, careful human writer can look "predictable" and get flagged.
How to spot and edit out AI writing tells step by step
Scan a draft in layers, starting with individual words and finishing with facts, and flag every tell before you rewrite anything. If you fix as you go, you tend to swap one stock phrase for another.
Step 1: Flag overused words and stock phrases
Search the draft for the vocabulary models overuse, then replace each hit with the word you'd say out loud. The evidence here is the strongest of any marker. An analysis of PubMed abstracts from 2010 to 2024 found "delves" appearing 28 times as often in 2024 as pre-ChatGPT trends predicted, with "underscores" and "showcasing" also far above trend. In a separate comparison, GPT-4o models used "tapestry" and "intricate" far more often than humans.
Group the hits by what they do in the sentence, because each group needs a different fix:
| Function | Words and phrases to flag | Plain replacement |
|---|---|---|
| Inflated vocabulary | "delve," "tapestry," "intricate," "underscore," "showcase," "pivotal," "realm" | look at, mix, detailed, show, key, field |
| Hedges and throat-clearing | "it is important to note," "it is worth noting that," "plays a significant role in" | Delete the opener and state the fact. Name the actual role. |
| Stock transitions | "furthermore," "moreover," "additionally," "notably" | "also," "and," "but," or no transition at all |
Treat the hedges as editor-flagged habits. Research on hedging rates is thin, so cut them because they delay the point. Studies of LLM vocabulary do find heavier use of transitions like "additionally" and "furthermore," and you'll see them most at the start of a paragraph.
Step 2: Check sentence length and rhythm
Read the draft aloud. Wherever three sentences in a row land at the same length and shape, rewrite one of them. Your ear catches what counting misses, especially repeated subject-verb openers and sentences that all end on a trailing "-ing" clause.
The edits are mechanical. Split a long sentence, merge two short ones, cut one entirely, or let a fragment stand. Vary where paragraphs end so they don't all close on the same beat.
Polish is the weakest tell in this layer. A draft with no typos and textbook grammar might come from a model, or from a writer who ran Grammarly and had a good editor. People misread this cue badly. In a PNAS study of AI-written self-presentations, participants rated grammatically flawed texts as more likely to be AI, when those texts were in fact less likely to be AI. So don't add errors to make copy look human. Fix the rhythm and leave the grammar clean.
Step 3: Break formulaic structure
Structural patterns survive word-level edits, so they need their own pass. A study of 225 argumentative texts found that LLM output used the tricolon, a list of three, far more often than human experts did, and it was the strongest differentiating marker the authors measured.
Each pattern has one fix:
- Rule-of-three lists: Use as many items as the material has. Two is fine. Four is fine.
- Symmetrical paragraphs and bullets: Vary paragraph length, and turn a bullet set back into prose when the items depend on each other.
- "It's not X, it's Y" framing: Delete the denial and state Y.
- Summary conclusions: End on the last new fact or the next action. A Scientific Reports comparison found every ChatGPT essay in its corpus opened its conclusion with "In conclusion," while human essays varied.
Step 4: Restore tone and personal voice
Flat, neutral prose usually means the author is missing. The models have no stake in the argument, so nothing in the text pushes back or takes a side. The news-text study linked above also found human writers expressed stronger negative emotions, such as fear and disgust, and less joy than the models did.
Put the author back with three kinds of material:
- Anecdote: A specific moment from your team's own work, with the tool, the date or the number attached.
- Opinion: A claim you'd defend in a meeting, stated without three hedges around it.
- Specific detail: A name, figure or example a competitor couldn't paste into their own post.
Here's a quick test. If you can swap a competitor's name into a paragraph and it still reads true, that paragraph needs a fact only you have.
Step 5: Verify facts and citations
Content-level errors are the one tell you can check outright, and they do the most damage once published. A Scientific Reports study of short literature reviews written by ChatGPT-3.5 and ChatGPT-4 documented fabricated and erroneous citations across 636 references on 42 topics.
Run this check on every draft:
- Open every link and confirm the page exists.
- Find the claim on that page, with the same figure, population and period.
- Confirm DOIs resolve to the paper named, not an unrelated article.
- Search quoted phrases verbatim to confirm someone said them.
- Cut generic claims that no source supports, or anchor them to one that does.
Step 6: Compare a human and an AI passage side by side
Seeing the markers in context trains the eye faster than any list. The pair below is illustrative, written for this guide from one prompt: "Write a short paragraph explaining why content briefs fail."
The AI-style version:
In today's fast-paced content world, briefs play a pivotal role in shaping successful outcomes. However, it is important to note that many briefs fall short. Furthermore, they often lack clarity, context, and direction. It's not about writing longer briefs, it's about writing smarter ones. In conclusion, a well-crafted brief is the foundation of compelling content.
Each marker in that passage maps to a step above:
- "In today's fast-paced content world" is a scene-setting opener with no information.
- "Pivotal" and "well-crafted" are inflated vocabulary.
- "It is important to note" and "Furthermore" are a stock hedge and a stock transition.
- "Clarity, context, and direction" is a reflexive list of three.
- "It's not about X, it's about Y" is the contrast frame.
- "In conclusion" opens a summary that adds nothing.
- The whole passage could describe any company's briefs.
The human version:
Most briefs fail before the writer opens them. The brief says "mid-market finance teams" and nothing about which report they're stuck on at month-end, so the draft comes back addressed to everyone. A longer template won't fix that. The brief needs one sentence on the reader's actual problem and one on what they should do after reading.
The edited AI version:
Briefs usually fail on context. They name an audience like "marketing leaders" and stop there, so the writer guesses at the problem and the draft drifts into generalities. Adding sections to the template won't help. Add one line on what the reader is stuck on and one on what they should do next, and the draft has something to aim at.
Step 7: Weigh markers together before judging
Weigh every marker against how often skilled or non-native human writers produce it. Most tells are common in human prose, so their reliability per document is low:
| Marker | Also common in skilled or non-native human writing? | Reliability as a signal |
|---|---|---|
| Overused style words | Yes, in academic and polished prose | Medium, stronger across a corpus than in one document |
| Uniform sentence length | Yes, in technical and non-native writing | Low to medium |
| Rule-of-three lists | Yes, a long-standing rhetorical device | Low |
| Em dashes | Yes, a staple of edited prose | Low |
| "It's not X, it's Y" contrast | Yes, in editorial and marketing copy | Low to medium |
| Polished grammar, no typos | Yes, in any edited copy | Very low |
| Flat, neutral tone | Yes, in corporate and technical writing | Low |
| Generic, interchangeable claims | Yes, in weak human writing too | Medium |
| Fabricated or misattributed citations | Rare in careful human work | High, and checkable |
My rule of thumb is to treat one or two markers as noise. When tells cluster across three or more layers and a content-level problem turns up, raise it with the writer as an editing conversation, never as an accusation. Trained readers judge the same way. In one ACL 2025 study, annotators who frequently use ChatGPT for writing were accurate detectors, and they relied on AI vocabulary combined with formality, originality and clarity. Annotators who rarely used LLMs were poor detectors and overconfident.
Why AI detection tools get it wrong
Detectors produce a probability estimate. They can't establish who wrote a passage.
Jisc's guidance for further education colleges says AI detection systems cannot conclusively prove that text was written by AI, generate false positives and are easy to defeat. The same page warns that detection can unfairly discriminate and compound existing bias.
- OpenAI's own experience bears this out. Its classifier correctly identified 26% of AI-written text on a challenge set of English texts and wrongly labeled human writing as AI 9% of the time. OpenAI withdrew the tool on July 20, 2023, citing its low rate of accuracy.
- Bias shows up where perplexity does its work. A Stanford study ran seven widely used GPT detectors on 91 TOEFL essays and 88 US eighth-grade essays. The detectors were near-perfect on the native writers but flagged non-native essays at an average false positive rate of 61.22%. The essays all seven detectors flagged had significantly lower perplexity. Non-native writers tend to use a narrower vocabulary, and narrower vocabulary reads as predictable.
The same logic applies to style habits. Writers used the em dash and the tricolon long before chatbots, and flagging them penalizes exactly the people who learned rhetoric well. For a content team, the practical rule is to never reject a freelancer's draft on a detector score alone. Read it against the layered check above.
Which markers survive paraphrasing and humanizer tools
Lexical and rhythmic tells are the easiest to scrub. Content-level tells tend to survive, and humanizers often create new ones.
Prompting helps less than people expect. A study of models imitating everyday writers from samples of their work found they could approximate user styles in structured formats like news and email, but struggled with nuanced, informal writing in blogs and forums. Paraphrasing goes further against detectors, since Krishna et al. showed that their DIPPER paraphraser evaded watermarking, GPTZero, DetectGPT and OpenAI's classifier, and Sadasivan et al. found that paraphrasing a text repeatedly degraded detection again.
Humanizer tools chase the detector score and damage facts along the way. An audit of 11 commercial humanizers found that every tool changed or dropped checkable details such as numbers, names and quoted spans in a large share of documents. Most of the changes hit quoted spans. Research on two-dimensional detection reaches a related finding: humanizers mostly alter surface expression, while content features stay detectable.
So after a humanizer pass, expect fewer "delves" and more uneven sentences, alongside the same generic claims, the same three-part structure and possibly a misquoted source. Manual editing still matters because it fixes the reasons a reader distrusts the text, and it catches the factual damage a humanizer introduces.
How to keep AI tells out of future drafts
Prevent tells upstream by writing them into the rules every writer and drafting agent works from. Editing the same phrase out of every draft wastes your review time. A rule written once stops it at the source.
Build these into your style guide and briefs:
- Banned phrase list: The words from Step 1, each paired with its plain replacement.
- Rhythm rule: Vary sentence length, and never let three sentences in a row share a shape.
- Structure rules: Count list items and end on the last new fact or the next action.
- Voice references: Two or three of your best published pieces, plus notes on what makes them sound like you.
- Verification rule: Every statistic carries its source, population and period, and someone opens every link before publishing.
When you fix the same tell twice in one month, turn it into a rule. In the GrowthX platform, review runs on notes left on highlighted passages, and a note scoped to "Every article" becomes a rule the writer keeps for later drafts, alongside the Writing Profile every drafting agent reads before it starts. Start by listing the five edits your reviewers make most often this week, and if you want to see those edits carried forward automatically, book a demo.
