Which Sentences Get Flagged as AI? Sentence-Level AI Detection

AI detectors flag a sentence when it looks more predictable than people usually write, and the usual suspects are generic openers, signposts and summary lines. Sentence-level AI detection scores each sentence (or a short window of sentences) separately, then combines those scores into a document result. Document-level detection scores the whole text at once. Because most real drafts are mixed, the way a tool rolls up a few flagged sentences can decide the verdict on the whole.

If you want to know which sentences are AI-flagged in your own draft, you need a sentence-level view. This post explains how the two approaches work, why a handful of sentences can tip a document, why rewriting only those sentences is safer than rewriting everything, and what changes when you combine several detectors.

Two ways to score a text

A document-level detector reads the full passage and returns one number, often phrased as a probability that the text is AI-generated. It has a lot of evidence to work with, so the estimate is relatively stable. The drawback is that it cannot tell you where the signal came from.

A sentence-level detector breaks the text into units and scores each one. Many tools then highlight the flagged units and report the share of text they cover. This is closer to how writing actually happens: people draft some paragraphs themselves, ask a model for help with others, and edit across both. Research has followed. SeqXGPT, presented at EMNLP 2023, built a dataset of documents mixing human and AI-modified sentences and reported that earlier methods struggled at the sentence level, while its approach worked at both sentence and document level (Wang et al., EMNLP 2023).

We expected attendance to fall after the fee change, and in two of the three districts it did. These findings highlight the important role that economic factors play in shaping educational outcomes. The third district is the puzzle: attendance rose by four points, and the head teachers we interviewed credited a bus route that opened the same term. This suggests that transport access may moderate the effect of fees. Overall, the results underscore the need for a comprehensive approach to policy design.

Two of five sentences flagged, one borderline. A document score would blur this into a single number.

Notice which sentences get flagged in the example. They are the generic summary lines, the ones that could close almost any paper. The sentences with specific places, numbers and a named cause read as human. That pattern is common, and it is useful: it tells you exactly where revision effort belongs.

The noise problem with single sentences

Sentence scores are informative but individually uncertain. One sentence has few tokens, so any predictability estimate rests on thin evidence. Many tools smooth this by scoring overlapping windows, so a sentence's label depends partly on its neighbors.

Turnitin has been open about this effect. In a post on its false positive rates, it described a sentence-level false positive rate of around 4%, and reported that 54% of false positive sentences sit directly next to actual AI writing (Turnitin). In other words, human sentences near flagged ones are more likely to be flagged too.

The practical reading: treat a single highlighted sentence as a prompt to look, not as a finding. Patterns across a paragraph are more meaningful than any one line.

Why a few sentences can tip the document

Every tool must turn sentence scores into a document verdict, and the choice of rule matters more than most users realize. Common options include the average sentence score, the share of sentences above a threshold, or a separate threshold on that share. Under a share-based rule, a 20-sentence abstract with four flagged sentences sits at 20%, which may or may not cross a reporting line. Fix two of them and the share halves.

Clusters matter too. Because neighboring sentences influence each other in windowed scoring, three generic sentences in a row can pull a fourth, human sentence over the line. Break up the cluster and the neighbor often clears on its own. This is why a small number of edits in the right places can change a document result more than a large number of edits spread evenly.

For academic writers, the typical culprits are predictable: opening context sentences, signposting ("This section discusses"), and closing summaries. We list fixes for those in how to revise AI-sounding sentences. Reference lists and quoted passages are a separate case; see why Turnitin flags references and citations as AI.

Why rewrite only the flagged sentences

If detection works sentence by sentence, revision should too. Rewriting a whole document to fix four sentences has three costs.

• Meaning risk scales with coverage. Every rewritten sentence can lose a hedge, a number or a citation link. Rewriting 4 sentences gives four chances for drift; rewriting 40 gives forty. See preserving meaning and citations.
• Your voice disappears. The sentences that read as human are usually the ones with your specific observations. A global rewrite flattens them.
• New flags can appear. A rewriter applied everywhere can make previously clean sentences look uniform, raising scores elsewhere.

1. Score each sentenceFind which units read as machine-like and which already read as human.
2. Leave human sentences aloneReturn them exactly as written, word for word.
3. Rewrite the flagged onesChange structure and rhythm while keeping claims, numbers, citations and terms.
4. Re-score in contextCheck the new sentences inside the full paragraph, since neighbors affect scores.
5. Repeat where neededGive a second pass to any sentence that is still flagged.

Aggregating several detectors

Different detectors use different training data, thresholds and units of analysis, so they often disagree on the same sentence (we cover the reasons in why AI detectors disagree). Combining them raises a design question that mirrors the sentence-to-document one: does a sentence count as flagged if any detector flags it, or only if most do?

The Liang et al. study shows how much that choice matters. Running seven detectors on human-written TOEFL essays, the authors found that 97.8% were flagged by at least one detector, while 19.8% were flagged by all seven (Liang et al., Patterns 2023).

• Flagged by at least one detector97.8%
• Average across detectors61.3%
• Flagged by all seven19.8%

Source: Liang et al., "GPT detectors are biased against non-native English writers", Patterns 2023

Same essays, same detectors, three very different numbers. A strict "any detector" rule is cautious about missing AI text but flags almost every human essay here. A "unanimous" rule is kinder to human writers but misses more. Neither is universally right; what matters is knowing which rule a report uses and reading it with that in mind. For writers, checking a draft against several detectors gives a more honest picture than trusting any single score, because it shows where tools agree and where a flag is likely noise.

How Academic Humanizer uses sentence-level scoring

The Humanizer in Academic Humanizer is built on this sentence-level logic. It scans your draft for sentences that AI detectors flag, rewrites only those, and returns sentences already read as human exactly as you wrote them. If nothing is flagged, the output equals the input and no credits are charged. The AI Detection Report pairs each flagged sentence with its rewrite and compares the draft and the output on the same measure, the share of sentences read as human, with results checked against Turnitin, GPTZero, ZeroGPT, Copyleaks, Sapling and Originality.ai. If a few sentences are still flagged, the report says so and you can run those passages again.