How AI Detectors Treat Quotations, Block Text, and Boilerplate Differently

You quote a paragraph from a Supreme Court opinion, paste in your APA reference list, and add the standard confidentiality footer your professor requires. Then you run the whole thing through an AI detector and the score climbs. Did the detector catch a robot? No. It caught the parts of your document that look like every other document, and that’s a very different thing.

Key takeaways

• Most AI detectors score text statistically and don’t “know” a sentence is a quotation unless the tool was built to strip cited material first.
• Quoted passages from books, papers, or famous speeches often read as low-perplexity (predictable) text, so they can raise an AI-likeness score even though a human typed them.
• Block quotes, reference lists, headers, and legal boilerplate are repetitive and templated, which is exactly the pattern many detectors associate with machine writing.
• A detector that blends your boilerplate into one overall percentage can give a misleading number; sentence-level or per-section views are far more useful.
• If your score jumps in a section that’s mostly quotes or stock language, that’s a cue to read the heatmap, not a verdict on your authorship.

What a detector actually “sees”

It helps to remember what these tools do under the hood. An AI detector doesn’t read for meaning. It measures how predictable your word choices are and how much your sentence rhythm varies. Two stats do most of the work: perplexity (how surprising the next word is) and burstiness (how much sentence length and structure jump around). Human writing tends to be a little unpredictable and uneven. A lot of machine-generated text comes out smooth and even. If you want the full mechanics, we wrote a whole piece on perplexity and burstiness.

Here’s the catch. The detector applies those measurements to whatever text you give it. It usually has no idea that one paragraph is a direct quote from Toni Morrison and the next is your own analysis. To the scoring model, it’s all just tokens. Whether quoted material gets special treatment depends entirely on whether someone built that handling into the specific tool, and most don’t advertise it clearly.

That’s the whole problem in one sentence: detectors score the surface of the text, and quotations, block text, and boilerplate have surfaces that don’t behave like normal prose.

Why quotations confuse detectors

A direct quotation is, almost by definition, polished. A published sentence has been edited, often heavily. Famous lines have been repeated so many times that a language model can practically predict them word for word. That low surprise is exactly the signal that pushes a perplexity-based detector toward “AI-likely.”

Think about it from the model’s side. Ask it to guess the next word in “Four score and seven years ago, our fathers brought forth…” and it nails it. Low perplexity. The detector reads that low perplexity as machine-like, even though Lincoln wrote it and a student merely typed it into an essay. The detector isn’t wrong about the statistics. It’s just answering a different question than the one your teacher cares about.

This is why quote-heavy writing, literature essays, legal memos, theology papers, anything that leans on canonical text, tends to score higher than you’d expect. It isn’t a flaw in your authorship. It’s a mismatch between what the tool measures and what borrowed language looks like.

The two-document trick (a quick diagnostic)

Say you turn in a 1,200-word literary analysis and a detector returns 40 percent AI-likely. Before you panic, split the document. Run your original prose in one check and the block quotes in another.

Often you’ll find your own writing comes back low and the quoted passages carry most of the score. That tells you the number is being inflated by borrowed text, not by your sentences. Just don’t strip quotes out of the actual assignment to game the result; that’s how accidental plagiarism happens, and it misrepresents your sources. The split is a diagnostic, not an editing instruction.

Block text behaves like its own little document

Block quotes get even weirder. When you indent a long passage, you’ve created a chunk of uninterrupted, professionally edited text with no contractions, no asides, no mid-thought corrections. That’s smooth, even, low-burstiness writing. A detector looking at that block in isolation sees the calmest, most uniform paragraph in your whole paper.

Pulled quotes, epigraphs, and excerpted dialogue do the same thing. They interrupt your natural rhythm with a stretch of someone else’s careful prose. If a detector scores section by section, those blocks can light up on their own. If it scores the whole document at once, they drag the average in a direction that has nothing to do with how you wrote.

Boilerplate is the quiet score-raiser

Boilerplate is the stuff nobody really “writes” so much as reuses. Reference lists. Citation formatting. Methods sections. Legal disclaimers. The “About the author” blurb. Cookie notices. The standardized header on a lab report.

All of it shares one trait: it’s templated. Everyone produces it roughly the same way because it’s supposed to be uniform. That uniformity reads as low variety and high predictability, which is the same fingerprint detectors associate with AI output. So a perfectly human, hand-typed bibliography can score as suspiciously machine-like, purely because bibliographies are formulaic by design.

A small scenario. A freelancer sends a client a proposal. The body is clearly their own voice, casual and specific. But the document also has a boilerplate scope-of-work section and a standard payment-terms paragraph copied from the company template. The client runs it through a detector and gets a middling overall score. The original pitch was fine; the recycled legal language is what muddied the number. Without a per-section view, the freelancer has no way to see that.

Why the overall percentage can mislead you

Here’s where all three problems compound. Many detectors hand you a single percentage for the whole document. That number is an average, and averages hide structure.

Imagine a 2,000-word report: 1,500 words of original analysis that scores low, plus 500 words of quotes and boilerplate that score high. The blended figure might land somewhere in the middle, high enough to worry you, low enough to be ambiguous, and not actually descriptive of either part. You end up reacting to a statistic that doesn’t map to any real section of your writing.

This is the strongest argument for using a detector that shows a heatmap at the sentence or token level rather than one flat verdict. When you can see which spans are flagged, the pattern usually becomes obvious: the hot zones are the quotes and the stock language, and your own prose is calmer. The heatmap turns a vague accusation into a readable map. You can dig into more on how detection works if you want to understand what those highlights mean.

It’s worth remembering detectors output probabilities, not proof. OpenAI even retired its own AI text classifier in July 2023 because its accuracy was too low to be reliable. And a 2023 study by Liang and colleagues at Stanford found detectors disproportionately flagged writing by non-native English speakers. These tools are useful signals, not judges, and that’s doubly true on documents full of borrowed and templated text.

What to do with this knowledge

If you write or edit: check your draft, then check it again with the quotes and boilerplate set aside, so you understand where the score comes from. If you’re a teacher or editor: read the highlighted spans before you read the percentage, and expect quote-heavy or template-heavy documents to score higher for boring, structural reasons.

The goal isn’t to dodge detection. It’s to read the result honestly. A raised score on a literature review, a legal brief, or a heavily sourced article is often just the math reacting to predictable, borrowed text. Knowing that keeps you from misreading a number, whether you’re defending your own work or evaluating someone else’s.

Bottom line

Quotations, block text, and boilerplate aren’t AI, but they share the statistical fingerprint that AI detectors hunt for: predictable, polished, repetitive language. A single overall percentage smears all of that together and can leave you staring at a number that describes your sources more than your writing. Read the per-sentence view, separate the borrowed parts from your own, and the picture usually clears up fast.

Want to see exactly which lines are pulling your score, quotes, boilerplate, or your actual prose? Run a check and look at the heatmap, not just the headline number.