5 AI Writing Patterns Detectors Flag (And How to Fix Them)
The five statistical patterns AI detectors actually score: uniform sentence length, AI vocabulary, transitions, em dashes, parallel structure. Fix workflow inside.
AI detectors do not look for plagiarism. They do not have a database of ChatGPT outputs to match against. They are statistical classifiers, and they read your prose for the specific fingerprints that language models leave behind by default. Five of those patterns do most of the work in flagging text as AI. Knowing what they are means you can fix them by hand in fifteen minutes — or hand them to a structural rewriter that targets each one explicitly.
Why these five patterns and not others
We pulled the technical writeups from each major detector — GPTZero on perplexity and burstiness, Turnitin's model documentation, the Wikipedia Signs of AI Writing project that the platform's volunteer editors maintain — and looked for the patterns each detector explicitly mentions or implicitly weights through training data. Five patterns appear in every list. Other patterns matter, but these five carry most of the signal weight in our 50-essay corpus testing.
Each one is a side-effect of how language models work. ChatGPT and Claude pick the highest-probability next token at every step, which by definition produces text that is statistically average — and "statistically average" looks like uniform sentence lengths, predictable vocabulary, formal transitions, balanced clauses. These are not bugs in the LLM; they are the model behaving exactly as designed. Detectors catch them because the math is unmistakable.
Pattern 1 — Uniform sentence length (the biggest signal)
Take any paragraph that ChatGPT-4 produces and count the words per sentence. The default behavior of the model produces sentences that hover between 18 and 25 words each. A whole paragraph at that length reads smooth in your head — nothing wrong with it on the surface — but the variance is too low. Detectors measure this as the coefficient of variation (CV%) of sentence lengths. AI text usually sits below 30%. Real human writing runs 50–90%, depending on the writer.
GPTZero formalizes this as "burstiness." Turnitin builds it into the per-token probability scoring indirectly — uniform-length sentences produce uniform per-sentence perplexity, which raises the document-level AI score. Sapling and Originality.ai both weight burstiness heavily. So the sentence-length pattern is the single most reliable signal across detectors.
How to fix it
Find any three consecutive sentences in your draft that are roughly the same length. Cut one in half — split it at the comma or "and" and turn it into two sentences. Glue another two together with a comma or a dash. Drop a four-word fragment somewhere in the paragraph. The math changes immediately. CV% jumps from 25% to 60% with two edits, and every detector reads it as more human.
The economic impact of climate change is substantial and far-reaching. Rising sea levels threaten coastal infrastructure across the globe. Agricultural yields decline as temperatures continue their upward trajectory. Insurance industries face mounting claims from extreme weather events.
Climate change is expensive. Coastal cities are flooding, crops are failing in places that used to grow them reliably, and the insurance industry is buckling under the claims. Big numbers. The IPCC report from last year put the annual cost in the trillions, and that estimate keeps revising upward.
Pattern 2 — AI vocabulary clusters
Language models have favorite words. Not random favorites — the words appear at frequencies 50–100x what you'd find in normal student or journalist writing. GPTZero's published top-10 list includes delve, crucial, tapestry, foster, leverage, realm, navigate, nuance, elevate, and landscape. Independent analysis adds furthermore, moreover, endeavor, comprehensive, vital, and pivotal. The phrase "crucial role" alone appears 182x more often in AI text than human text, per academic linguistic analysis published in 2024.
Three or four of these words in the same paragraph is enough to push a sentence into the AI bucket. The detector is not running a string match — it is reading word-context combinations through its trained weights, which learned that "delve into the realm" almost never appears in pre-2023 human writing and almost always appears in post-2023 LLM output.
How to fix it
Keyword search through your draft for the cluster: delve, crucial, tapestry, realm, foster, leverage, navigate, nuance, elevate, landscape, furthermore, moreover, endeavor, comprehensive, vital, pivotal. For each match, swap to a plainer word that means the same thing. Delve into becomes look at. Crucial becomes important or just gets cut. Tapestry becomes set or mix. Realm becomes area. The fix is mechanical and takes about three minutes per thousand words.
Pattern 3 — Transition word density
Real student writing connects ideas implicitly or just starts the next sentence. ChatGPT-style writing leans on logical connectors — furthermore, moreover, additionally, however, in conclusion, it is important to note, consequently. Each one signals a structured argument. LLMs produce these at rates above 5 per 100 words; natural human writing usually sits below 2 per 100. The density itself is a signal.
This pattern is the easiest to fix and the most-overlooked. Most students focus on vocabulary while leaving the transitions intact, which means the score drops by a few percent rather than below the threshold. Cutting transition words alone — without changing anything else — usually moves a 60% AI score down to 35%.
How to fix it
Search and delete. Find every furthermore, moreover, additionally, however, therefore, consequently, and in conclusion. Delete them. The sentence usually still works without the connector. If it doesn't, replace with a simpler alternative — also, but, so. Or merge the sentence with the previous one using a comma. Five minutes of search-and-replace shifts the score significantly.
Pattern 4 — Em dashes everywhere
This one became famous in 2024 when journalists noticed they could spot ChatGPT in published articles by counting em dashes. Rolling Stone wrote it up and the pattern has not gone away — Claude and GPT-4 both deploy em dashes in places where most human writers would use commas, parentheses, or just a period. The signal is one em dash per paragraph or higher, especially in places where a comma or colon would fit naturally.
Detectors do not specifically search for em dashes (it would be too easy to game), but the punctuation contributes to the per-token probability scoring. An em dash in a low-probability position raises the perplexity slightly; an em dash in the predictable AI position lowers it. Across a paragraph with three or four em dashes in AI-typical positions, the cumulative effect is real.
How to fix it
Count your em dashes. If there are more than two per paragraph, replace at least half with commas, colons, or parentheses. The reading rhythm doesn't change much — em dashes and commas both signal a brief pause — but the statistical fingerprint does. This fix takes thirty seconds and contributes about 5–10% to the score reduction in our testing.
Pattern 5 — Symmetrical parallel structure
"We must consider the social, the economic, and the cultural dimensions of this issue." That sentence is a Claude or GPT-4 specialty. Three balanced parallel elements, all the same shape, the same length. Real human writing breaks parallelism — one of those clauses gets a different shape, or runs longer than the others, or trails off. The model defaults to perfect symmetry because perfect symmetry is the highest-probability completion of "we must consider the X, the Y."
This pattern appears most strongly in argumentative or analytical writing, where students are using ChatGPT to organize an argument. The triplets and quartets of parallel constructions stack up. Detectors do not flag parallelism directly, but the combination of parallel structure with even sentence length and AI vocabulary lights up the classifier.
How to fix it
Find any three-or-more parallel construction (three nouns in a row, three "we must" phrases, three balanced clauses). Break the symmetry. Make one element longer than the others, or restate it as a question, or just delete one. The sentence is shorter and reads more like real prose. Detector signal drops accordingly.
How detectors actually weight these patterns
No detector publishes the exact weights, but our reverse-engineering work across 50 humanized essays plus the technical writeups from each platform produces a rough ranking. Burstiness (Pattern 1) carries the most weight on every detector — fix uniform sentence length and the score drops more than from any other single fix. AI vocabulary (Pattern 2) is second, and it compounds with the burstiness signal — paragraphs with both uniform length and AI vocabulary score higher than the sum of the parts. Transitions, em dashes, and parallel structure each contribute 5–15% individually, but they add up.
This is why fixing all five patterns gets your score under 5%, while fixing only one or two leaves it at 30–50%. The detector is reading a combined signal. Each pattern feeds the classifier evidence; missing one pattern leaves enough evidence to flag.
Refrazr fixes all five — automatically
The Refrazr engine runs your text through pattern analysis on all eight scoring dimensions (the five above plus passive voice, short sentence presence, and contraction frequency), tells the LLM exactly which patterns are present, and rewrites at the structural level. Free 500 words/day, no signup.
Try Refrazr free → Read the technical methodologyPatterns detectors don't (yet) catch
Three patterns are AI-typical but currently fall below the detection threshold. Hedging language is one — phrases like it can be argued that, it is worth considering, some scholars suggest. LLMs use these constantly to avoid taking positions; humans use them less. Detectors seem to weight hedging weakly because formal academic writing also hedges. So this pattern is real but not currently a strong signal.
The second is meta-commentary — sentences like this essay will explore or in this paper, we examine. Pure meta-commentary is an LLM tic. Detectors miss it because journal-style academic writing uses similar structures. The third is the rhetorical-question opener: What is the role of X in modern society? at the start of a paragraph. Common in ChatGPT, less common in real student writing, but detectors do not specifically score for it.
These three patterns are not safe to use freely — they signal AI to a careful reader even when the detector misses them. But the math of evading detection currently does not require fixing them. Focus on the five primary patterns and the score moves where you need it.
Manual fix vs structural humanizer — which to use
By hand, fixing all five patterns on a 1,000-word essay takes about 45 minutes if you are careful. The fix is mechanical and reversible — you can do it before submitting and have version-history evidence of your edit pass, which protects against false-positive disputes. The downside is the manual rewrite usually misses one or two patterns, which keeps the score around 10–20% rather than under 5%.
A structural humanizer like Refrazr runs all five fixes (plus three more) automatically, in the right combination for your specific text. It analyzes your input first — measures CV%, counts AI vocabulary density, scores transition density, em dash frequency, parallel structure ratio — then sends the LLM specific instructions to break the patterns it found. Post-processing handles residual signals. Across 50 humanized essays in March 2026, structurally rewritten text scored below 5% in 47 of 50 cases. The manual fix, on the same essays, scored under 20% in 40 of 50.
The honest tradeoff: manual fix preserves your authorial voice exactly, takes time, leaves a few patterns unfixed. Structural humanizer rewrites the prose entirely, takes fifteen seconds, lands lower scores. For high-stakes submissions, run both — manual fix first to preserve voice, then a humanizer pass to clean up residual signals. Fifteen minutes total, scores reliably under 5%.
Run both passes — manual edit + Refrazr
Free 500 words/day. Paste your draft after your hand pass, click humanize, copy the result. The combined approach is the most reliable workflow we know of, and it preserves your meaning while clearing the structural patterns detectors flag.
Try Refrazr free → Test before and after