AI Checker Glossary: Perplexity, Burstiness and Detection Terms Explained by Turnitin0

Academic integrity conversations have shifted from "did you copy this?" to "did a machine write this?" That shift created a vocabulary problem. Instructors, students, and integrity officers use the same words to mean slightly different things, and vendor marketing makes it worse by inventing terms that sound technical but describe nothing measurable.

A shared glossary fixes three practical problems:

• Misread reports. A student sees a highlighted sentence and assumes the whole essay is condemned. Understanding sentence-level versus document-level scoring prevents that panic.
• Bad arguments. "The detector is wrong" is not a defense. "The detector's confidence fell below its reporting threshold, which is why the report shows *% instead of a number" is a defense that references the actual mechanism.
• Smarter revision. You cannot lower perplexity or raise burstiness if you do not know what those measurements describe.

Turnitin0's own research program exists because these terms need evidence behind them, not just definitions. Six published reports test how the detector behaves on real corpora, and the numbers are specific enough to be useful.

The Core Vocabulary at a Glance

These two terms come from language modeling, not from academic integrity. Detectors borrowed them because they separate human and machine writing reasonably well.

Perplexity

Perplexity measures how surprised a language model is by each word in a sequence. If a model assigns high probability to every word choice, the text has low perplexity. If the word choices are unpredictable, perplexity rises.

Human writing sits in the middle. We use common words most of the time, then reach for something unexpected: an odd verb, a specific noun, a clause that does not follow the expected pattern. Language models do the opposite. They select high-probability tokens, which produces smooth, low-perplexity prose.

This is why AI text often reads as competent but flat. Nothing is wrong, and nothing is surprising.

Low perplexity is not proof of AI authorship. Technical writing, legal boilerplate, and heavily edited academic prose can all score low because their vocabulary is deliberately constrained.

Burstiness

Burstiness measures variation. In practice, it looks at how much sentence length, clause structure, and rhythm change from one sentence to the next.

Human writers are bursty. A long, subordinate-clause-heavy sentence is followed by a short one. A paragraph opens with a claim, then wanders, then snaps back. Machine output tends toward uniformity: sentences cluster around a similar length, paragraphs follow a similar shape, and transitions repeat.

High burstiness is a human signal. Low burstiness is a machine signal. Neither is decisive on its own, which is why detectors combine several measurements before producing a score.

Tokens and Word-Level Scoring

A token is the unit a model reads. It is usually a word or part of a word. Detectors do not score documents as single objects; they score tokens and words, then aggregate upward.

This distinction matters because Turnitin0's benchmarks report word-level accuracy and word-level evasion rates. For example, in a benchmark of 180 essays generated by GPT-5.6-Sol across 30 majors, Turnitin reached 97.88% word-level accuracy, correctly flagging 153,620 of 156,955 words as AI-generated. In a separate benchmark of 170 essays produced by Claude Fable-5, accuracy reached 99.01%, with 130,151 of 131,451 words correctly identified.

Those percentages describe words, not verdicts. A document can contain flagged words and unflagged words at the same time.

Sentence-Level and Document-Level Scores

Detectors typically produce three layers of output:

1. Token or word level — the finest grain, used to highlight specific passages.
2. Sentence level — grouped highlights that show which sentences carry the strongest signal.
3. Document level — a single aggregate figure, often expressed as a percentage of the text flagged.

Reports that show only a document-level number hide the useful information. Sentence-level highlighting is what tells you where to revise.

Definitions of perplexity and burstiness explain the inputs. These terms explain the decision.

AI Detector vs. AI Checker vs. AI Scanner

These three words are used interchangeably in marketing, and the difference is mostly branding. Functionally, an AI detector classifies text as machine-generated or human-written. An AI checker does the same thing, sometimes with additional reporting. An AI scanner usually implies batch or file-based processing.

What matters is not the label but the output format. A tool that returns a bare percentage is less useful than one that returns a highlighted report you can act on.

Confidence Threshold

A confidence threshold is the score below which a detector declines to report a precise result. Turnitin applies a 20% threshold: when AI detection falls below that level, the report displays *% instead of an exact percentage.

This is a deliberate design choice, and it is the single most misread element in AI detection reports. A *% does not mean "0% AI." It means the signal was too weak to report a specific figure. Students who understand this avoid two mistakes: assuming a clean result, and assuming a hidden accusation.

False Positives and False Negatives

A false positive occurs when human-written text is flagged as AI-generated. This is the error type that damages students, because it shifts the burden of proof onto the writer.

A false negative occurs when AI-generated text passes as human. This is the error type that damages institutions.

Detector quality is usually described by how it performs on both. Turnitin0's benchmark on human-written graduate-level essays from the PLOS corpus is instructive: across 504 essays totaling 135,712 words, spanning 18 majors and four domains, Turnitin achieved 100.0% word-level accuracy, correctly classifying every word as human-written. That is a direct measurement of false-positive resistance on that corpus.

Word-Level Accuracy and Evasion Rate

Word-level accuracy is the share of individual words classified correctly. Evasion rate is the inverse framing used to evaluate humanizers: the share of AI-generated words that escape detection.

Turnitin0's humanization benchmark tested 174 humanized essays totaling 204,736 words and recorded an overall word-level evasion rate of 76.44% against the Turnitin AI detector. Results varied sharply by discipline: Education essays reached 100%, while English essays were lowest at 55.41%.

Reporting the range rather than a single headline number is the honest approach, and it is worth remembering when any vendor claims universal bypass.

AI-Polished Text

AI polishing means a human wrote the draft and a model rewrote it for fluency. This is a distinct category from full generation, and detectors handle it differently.

Turnitin0's benchmark on 500 AI-polished graduate essays totaling 132,275 words found Turnitin's word-level accuracy at 47.54%, correctly flagging 62,879 words. In other words, polishing is substantially harder to detect than generation. That finding is relevant to anyone trying to understand why a lightly edited draft produces an ambiguous report.

Report and Platform Terms: What You See in the Interface

Once a detector has made its decision, the platform has to present it. These are the terms that appear in that presentation.

Similarity Score vs. AI Score

A similarity score measures textual overlap with existing sources. It is a plagiarism metric, not an AI metric. An AI score estimates the likelihood of machine generation. They are produced by different systems and answer different questions.

A high similarity score can occur in entirely human writing that quotes heavily. A low similarity score says nothing about AI involvement. Confusing the two is the most common error in reading an integrity report.

Repository vs. Non-Repository Checks

A repository check adds the submitted document to a database that future submissions are compared against. A non-repository check does not.

This distinction has real consequences. If you submit a draft to a repository-based system before your final submission, your own draft can later match your own final paper. Turnitin0 runs non-repository checks: your file is checked without being added to Turnitin's student paper database, and reports are not shared with third-party databases. You can also delete files from your account.

Turnitin Report Types

Turnitin0's checking service returns two downloadable PDFs in a single order: a Turnitin AI detection report and a similarity/plagiarism report. The reports are formatted identically to what professors see in their learning management system, which is the point. You are not looking at a simplified consumer summary; you are looking at the same artifact.

The service accepts.docx,.pdf, or.txt files, English only, with a word count greater than 300 and less than 30,000, and a file size under 20 MB. Turnaround is under 15 minutes in 98% of cases, with most orders completing in 5 to 15 minutes and a guaranteed 30-minute ceiling during rare queue spikes.

AI Humanizer

An AI humanizer rewrites text to reduce machine-detectable patterns, typically by increasing burstiness and varying vocabulary. Turnitin0's humanizer is a separate product from the checking service, and 98.2% of humanizer orders are re-checked with Turnitin.

Two honest caveats belong here. There is no free word quota or free trial for the humanizer. And humanization is not a guarantee: the 76.44% overall evasion rate in Turnitin0's benchmark means a meaningful share of AI words still register, with wide variation by subject.

How to Use This Vocabulary in Practice

Definitions are only useful if they change what you do. Here is the practical sequence.

Step 1: Check before you submit. Run your draft through a Turnitin check service that returns instructor-identical reports. Turnitin0 has delivered over 100,000 AI and similarity reports to more than 20,000 students across the United States, United Kingdom, Canada, Australia, New Zealand, and Ireland, with a 4.9/5.0 satisfaction rating.

Step 2: Read the right layer. Look at sentence-level highlights first. A document-level number tells you whether to worry; highlights tell you what to fix.

Step 3: Interpret *% correctly. If your report shows *%, the signal fell below the 20% confidence threshold. That is a weak-signal result, not a clean bill of health and not an accusation.

Step 4: Revise for burstiness, not for synonyms. Swapping words does not change perplexity much. Varying sentence length, breaking parallel structures, and adding specific detail does.

Step 5: Re-check. Revisions can move a score in either direction. Verify rather than assume.

What Real Users Report

Turnitin0's user feedback is consistent on the operational details. Raini Dipré (CA) described the process as "easy, fast, and efficient," adding that the report came back faster than expected. May zin (SG) noted the report was complete in about 20 minutes and highlighted the ability to download AI and similarity reports at the same time. daniela pellegrini (GB) has used the service several times and found the humanizer helpful during revision. Shubham Pachauri (IN) valued the humanizer for sounding more natural while preserving original meaning. Shawn Thakur (AU) and Encrypted (GB) both emphasized ease of use and on-time delivery, and Taksh Patel (AU) called the service legitimate and functional.

These are operational endorsements, not claims about detection outcomes. That distinction is worth keeping.

Honest Limitations

Turnitin0 is an independent service and is not affiliated with Turnitin, LLC. The checking service handles English documents only, with the word count and file size limits noted above. The humanizer has no free tier. On Trustpilot, the company has not recently invited customers, so the nine reviews on its profile, all from the last 12 months, may not be representative. Anyone evaluating the service should weigh those facts alongside the benchmarks.

Conclusion

The vocabulary of AI detection is not decorative. Perplexity and burstiness describe the signals. Confidence thresholds, false positives, and word-level accuracy describe how those signals become a verdict. Similarity scores and repository status describe how the verdict is presented and stored. Learn the terms, and a report stops being a black box.

The practical recommendation stands: use the Turnitin AI detector from Turnitin0 to see instructor-identical AI and similarity reports before you submit, and use the AI humanizer when revision is the right response. Turnitin0 is transparent about its limits, publishes its benchmark methodology, and gives you two reports in one order without adding your work to a repository or requiring a subscription. Start with the definitions, then verify them against your own draft at Turnitin plagiarism checker.