Signs of AI writing: the full list we flag

By Sourabh Singh26 September 20268 min read

Every phrase pattern and sentence-level signal in the TextAsMe analyser, published in full, then tested on 200 human and 200 ChatGPT answers to the same questions.

Most lists of AI tells are somebody's hunch. This one is the actual rule set our analyser runs on every draft pasted into TextAsMe, published as-is. When a sentence gets flagged on this site, it is because it matched one of the entries below.

What we found in 400 real answers

We ran the analyser over 200 questions from HC3, a public research dataset (Guo et al., 2023). For each question we took one answer written by a person before ChatGPT existed and one answer ChatGPT gave to the same question. The questions come from Reddit's Explain Like I'm Five, personal finance, computer science and medicine, and every answer is between 120 and 600 words. The table shows the share of answers in which each signal fired at least once.

Signal	ChatGPT answers	Human answers
Any flagged phrase	43.5%	19%
Sentence-length spread below 6	33%	12%
Rule-of-three list	28.5%	10.5%
"a wide range of" and similar	18%	2%
Metronome rhythm (sentence matches the draft's average length)	15%	4%
Stock transition (Moreover, Furthermore, Additionally)	9.5%	4.5%
"It is important to note that"	8.5%	0%
No concrete detail (no number, name or date)	97%	89.5%
No contractions at all	42%	41%

Three things stood out. Rhythm and structure separate the two groups better than any single word does: ChatGPT answers were almost three times as likely to have flat sentence lengths and to fall into lists of three. The strongest single phrase was "it is important to note that", which appeared in 17 of the 200 ChatGPT answers and in none of the human ones. And some signals barely separate at all. Nearly every short answer, human or machine, lacks a number, name or date, and contraction rates were the same in both groups.

LLM vocabulary (20)

Words that are rare in human prose but wildly over-represented in model output.

Pattern	Weight	Why it reads as AI	Write instead
delve into	1.00	"Delve" appears in model output roughly 10x more often than in human writing. It is the single most-cited detector trigger.	dig into
tapestry	1.00	Metaphor almost nobody reaches for unprompted. Near-certain model fingerprint.	mix of
a testament to	0.90	Ornamental praise construction. Very high frequency in generated text.	shows
navigate the complexities	0.95	Stock metaphor pairing. Detectors weight multi-word collocations heavily.	deal with
in the realm of	0.80	Filler prepositional frame that adds no information.	in
underscore	0.60	Formal synonym for "shows" that models strongly prefer.	shows
importance intensifier	0.55	Models reach for maximum-importance adjectives by default. Humans reserve them.	important
complexity adjective	0.70	Signals depth without supplying any. Heavy model marker.	complex
myriad/plethora	0.75	Elevated quantity words that spike detector scores.	many
corporate verb	0.55	Abstract verbs preferred by models over concrete ones.	use
foster	0.55	Abstract verb preferred by models over concrete ones.	build
facilitate	0.55	Abstract verb preferred by models over concrete ones.	help
brochure adjective	0.60	Marketing register bleeding into prose.	solid
pave the way	0.80	Dead metaphor with a fully predictable next token.	lead to
shed light on	0.75	Same problem: the detector predicts "light on" with near-certainty.	explain
at the forefront	0.70	Stock positioning phrase.	leading
ever-evolving	0.80	Time-filler cliché that models use to open paragraphs.	changing
embark on a journey	1.00	Peak generated-essay phrasing.	start
unlock the potential	0.80	Product-launch phrasing in the middle of prose.	get more out of
meticulous	0.55	Over-used praise adverb in model output.	careful

Filler opener (6)

Throat-clearing before the actual point. Models do this to fill the first clause.

Pattern	Weight	Why it reads as AI	Write instead
important to note	1.00	Six words before the sentence starts. Models use this to buy time; humans just say the thing.	Delete it
worth noting	0.95	Same throat-clearing pattern.	Delete it
should be noted	0.95	Passive throat-clearing. Delete it and the sentence improves.	Delete it
in today's world	1.00	The classic generated-essay opening. Detectors flag it on sight.	Delete it
in conclusion	0.85	Announcing the conclusion instead of writing one.	Delete it
one of the most	0.70	Ranking hedge that avoids making a claim.	a key

Stock transition (2)

Moreover / Furthermore / Additionally. Humans rarely stack these; models always do.

Pattern	Weight	Why it reads as AI	Write instead
stock transition	0.85	Formal connectives stacked at sentence starts. Human writers vary or drop them entirely.	Delete it
stock transition	0.60	Same pattern — a connective doing the work a sentence should do.	Delete it

Hedging (3)

Vague intensity with no commitment. Reads as risk-averse, which is how models are tuned.

Pattern	Weight	Why it reads as AI	Write instead
vague intensifier	0.50	Intensity with no measurement behind it.	a lot
vague quantity	0.50	Quantity words that dodge an actual number.	many
hedge phrase	0.60	Models hedge by default because they are tuned to avoid over-claiming.	Delete it

Wordy construction (6)

Three words doing one word's job. Inflates length, lowers perplexity.

Pattern	Weight	Why it reads as AI	Write instead
due to the fact that	0.80	Five words for "because".	because
in order to	0.55	"To" does the whole job.	to
the fact that	0.50	Almost always deletable.	that
a wide range of	0.60	Filler quantity frame.	many
when it comes to	0.65	Four-word topic switch that adds nothing.	for
has the ability to	0.70	"Can" is the whole phrase.	can

Cliché (2)

Phrases so worn that the next word is fully predictable to a detector.

Pattern	Weight	Why it reads as AI	Write instead
game-changer	0.70	Hyperbole that reads as generated marketing copy.	big shift
plays a vital role	0.85	Formulaic importance claim with a predictable tail.	matters for

The 6 sentence-level signals

Phrases are the easy part. Clean all of them out and a draft can still read as generated, because detectors mostly measure the shape of the prose. These are the six measurements we take on every draft, with the thresholds we use.

Signal	What we measure	Where it starts to read as AI
Rhythm	Spread of sentence lengths across the draft (burstiness)	Below 6. Human writing swings between short and long sentences.
Vocabulary	Unique words per 50-word window	Below 70% unique
Voice	Contractions per 100 words	Close to zero in formal writing. In our informal test set it did not separate the groups.
Specificity	Share of words drawn from the 220 most common English words	High share, with no number, name or date in the sentence
Restraint	Flagged phrases from the tables above per 100 words	Every hit counts, and the density adds up across the draft
Openings	Share of sentences that start with a distinct word	Three or more sentences in a row opening the same way

Two structural patterns also get flagged at sentence level: rule-of-three lists, because three-item lists are a model's favourite rhythm, and long clause chains that keep extending a sentence instead of stopping it.

What this list cannot tell you

Whether a person used AI. Plenty of people write "it is important to note" on their own, and detectors misfire most on non-native English writers.
What Turnitin or GPTZero will say. Their models are private; ours measures the same family of signals and says so.
Whether the ideas are yours. Swapping every phrase above changes the text, not the authorship.

For the vocabulary side in more depth, read our post on the words ChatGPT overuses. To see which of these fire on your own draft, paste it into the detector below: every flagged sentence names the rule it matched.

See which of these tells your draft trips, sentence by sentence.

Check a draft free