AI detection and personal statements
Short, formal, heavily revised, and written to a template. Every one of those pushes a score up, and a personal statement has all four.
An application essay is close to the worst-case document for a detector, and almost nobody checks one before sending it. Worth knowing why the profile is so exposed.
Four properties, all working against you
It is short. A few hundred words. Every detector is substantially less reliable under that length, because sentence-length variation needs a sequence of sentences to exist at all. See why detectors are unreliable on short text.
It is formal. You are writing to impress an institution, so the register is careful and the constructions are conventional. Conventional is predictable, and predictable is the signal.
It has been revised more than anything else you have written. Ten drafts, feedback from a teacher, a parent, a friend, an advisor. Every pass smooths the prose, and smoothing removes the variation a classifier reads as human.
Our corpus puts human academic writing at a within-document sentence-length standard deviation of 8.7 words and generated text at 6.4. A tenth draft sits closer to the lower number than a first draft does. See why good writing gets flagged more.
It follows a template. Opening hook, formative experience, what you learned, why this course, closing tie-back. Everybody is taught the same shape, so everybody produces the same shape, and a repeated structure reads as regular.
The connective problem is worse here
Statements are full of signposting because they are taught as structured arguments. "Furthermore", "Moreover", "Additionally", "Ultimately", "Overall". These open 2.4% of human sentences in our corpus and 5.8% of generated ones.
In a 600-word statement, four of them is a high density in a very small sample.
What moves it, and improves the statement
The changes here are the same ones a good advisor would ask for anyway.
The specific detail. "I developed an interest in materials science" is the generalised version. The version with the actual failed experiment, the actual date, the actual thing that broke, is both more human to a classifier and better to read. Specificity is the strongest human signal there is and it is the first thing lost in redrafting.
Sentence length variation. Put a four-word sentence next to a long one. Statements written to a uniform rhythm read as flat to a reader as well.
Cut the signposting. In 600 words the structure is visible without it.
Break the template somewhere. Start in the middle of the story rather than announcing it.
Full list in what makes text read as machine written.
Check it before it goes
The document is short, you only send it once, and you will not see any score anybody else generates. That combination makes a pre-send check unusually cheap relative to what it is worth.
Ours is free and unlimited with no account, takes .pdf and .docx, scores every paragraph separately so you can see which part of the statement reads flattest, and never stores your text. Wrong-flag rate on genuine human writing 0.4%, published, against Turnitin's 4%.