What makes text read as machine written
The specific, checkable properties that make prose look generated: sentence rhythm, connective density, hedging, and vocabulary that models over-select.
Text does not read as machine written because of any single word. It reads that way because of a set of statistical habits, and most of them are measurable in your own draft without any tool at all.
Uniform sentence length
This is the strongest single signal, and the easiest to check.
Language models produce sentences of strikingly consistent length. Human writing does not. A person writes a long, qualified, subordinate-clause-heavy sentence, then a short one. Then another long one. The variance is high and it is largely unconscious.
Count the words in each sentence of a paragraph and take the spread. Measured across our own corpus, the standard deviation of sentence length within a document has a median of 8.7 words for pre-2020 arXiv writing and 6.4 for generated text. Generated prose clusters; human prose scatters.
Our own detector uses this directly: the standard deviation of sentence length is one of its sixty stylometric features, and in practice it is among the most durable signals in the whole system, because it survives paraphrasing that changes every word.
Connective density
Generated prose signposts far more than people do. The usual set:
Moreover, Furthermore, Additionally, In addition, It is important to note that,
It is worth noting that, In conclusion, Overall, Ultimately
Measured on the same corpus, 2.4% of human sentences open with one of these against 5.8% of generated sentences, more than double the rate.
A person writing a paragraph uses one or two of these at most, and often none, because the logical relation is usually clear from the content. Generated text opens sentence after sentence with them.
Cutting them is the single highest-value edit available, and it usually improves the prose regardless of who wrote it. Removing a leading "Moreover," rarely costs any meaning.
Hedging without commitment
Models hedge, because hedging is safe. The result is stacked qualifiers that assert almost nothing:
This may potentially suggest that there could be a possible relationship
Each qualifier is individually reasonable. Together they leave a sentence that cannot be wrong because it does not claim anything. Human technical writing hedges too, but usually once and specifically: "suggests a relationship, though the sample is small."
Vocabulary that gets over-selected
Certain words appear far more often in generated text than in comparable human writing. In academic and technical registers the recurring set includes:
delve, leverage, robust, seamless, comprehensive, intricate, pivotal, realm,
underscore, testament, landscape, tapestry, navigate, foster, harness, unlock
None of these is wrong. Any of them can appear in perfectly ordinary human writing. It is the density that signals, not the presence, and this is what our lexical model measures: the distribution of word and short-phrase choices across a document, rather than any word on a list.
Structural tells
The three-part list. Generated prose reaches for triads constantly, often where two items or four would be more natural.
The symmetrical paragraph. Topic sentence, three supporting sentences of similar length, closing sentence that restates the opening. Repeated for every paragraph.
The restating conclusion. A final paragraph that summarises what was just said without adding anything.
Even paragraph length. As with sentences: human sections vary, generated sections do not.
Checking your own draft
Four passes, no tools required:
- Count words per sentence for one paragraph. Is the range narrow?
- Count sentences opening with a connective. More than one or two per paragraph is high.
- Look for stacked hedges. Reduce each to one qualifier or commit to the claim.
- Read the last paragraph. Does it add anything, or restate?
If you want the same judgement made statistically, our detector scores each passage separately rather than only the whole document, so it shows you which paragraphs carry the signal. It is free and unlimited.
What this does not tell you
These are properties of text, not evidence about a person. Human writing in a formal register, writing by non-native English speakers, and heavily edited technical prose all carry several of these markers naturally. That is precisely why detectors produce false positives, and why a score is a probability rather than a finding. See why AI detectors flag human writing.