Turnitin Check Glossary: Key AI Detection Terms Explained

Academic integrity language has changed fast. Turnitin itself frames its mission around helping students produce original work while giving educators tools to review submissions against a database of billions of web pages, archived internet content, student papers, and periodicals [1]. That database is the reason a similarity score is almost never zero — common phrases, assignment titles, and reference lists all match something [2].

Meanwhile, AI detection introduced a second, entirely separate vocabulary. Instructors now talk about confidence thresholds, false positives, and word-level accuracy. Students talk about "getting flagged" and "humanizing." The two conversations use overlapping words with different meanings, and that gap is where most confusion — and most panic — lives.

This glossary defines the terms in the order you are likely to encounter them: first similarity and plagiarism vocabulary, then AI detection vocabulary, then the newer terms around evasion and humanization.

Similarity Score (Similarity Index)

The similarity score is the percentage of text in a submission that matches other sources. Turnitin is explicit that it does not check for plagiarism in writing — it checks submissions against its database and highlights matches for a human to review [3]. The score is a tool within a review process, not a verdict.

Turnitin's own guidance breaks the score into colour bands: blue for no matching text, green for one word to 24% matching text, yellow for 25–49%, orange for 50–74%, and red for 75–100% [3]. Note that the colour scheme can vary depending on the integration, which is expected and does not affect the results [3].

The single most important thing to understand: a high similarity score is not automatically misconduct. A paper with extensive direct quotation and a full bibliography can score high and be entirely legitimate. National University's library guide makes this point directly — the similarity index represents all text matches, including references and direct quotes, and the amount of matching text matters less than the exact nature of the matching [2].

Match / Matched Text

A "match" is a specific passage in your submission that Turnitin has identified as similar to a source in its database. Matches are highlighted in the report. The table of contents on the right of the report lets you drill into each match individually [2].

Source Overlap

Source overlap is the broader category: the degree to which your text overlaps with existing sources. It includes legitimate overlap (quotation, common terminology, standard phrasing) and problematic overlap (unattributed copying). The report shows you the overlap; you and your instructor decide which is which.

Repository (Student Paper Database)

Turnitin's database includes a repository of works students have submitted in the past [3]. This is why submitting a paper you wrote for a previous class can generate a match against yourself. It is also why non-repository checking matters: a pre-submission check that does not add your file to that database leaves no trace for a later institutional submission to match against.

Non-Repository Check

A non-repository check processes your document without adding it to Turnitin's student paper database. Turnitin0 operates this way: the file is checked without being added to the student paper database, reports are not shared with third-party databases, and users can delete files from their account. For students running multiple drafts, this is the difference between a private rehearsal and a permanent record.

False Positive (Similarity Context)

A false positive in the similarity context is a match that looks like misconduct but is not — a quoted passage, a standard method description, a reference list. Turnitin's own documentation notes that even quoted text with quotation marks and references will show as a match [3]. This is not a flaw; it is how the tool is designed to work.

AI Detection Score

The AI detection score estimates the proportion of a document written by AI. It is a separate metric from the similarity score and appears in a separate report. Turnitin0 delivers both as downloadable PDFs with every order, so you can compare them side by side.

The 20% Confidence Threshold and the Asterisk

This is the term most students misunderstand. Turnitin shows an asterisk (*%) instead of an exact percentage when AI detection falls below its 20% confidence threshold. In plain terms: if the tool is not at least 20% confident that AI was involved, it declines to give you a number at all.

An asterisk is not a clean bill of health and it is not an accusation. It means the signal was too weak to report. Many instructors read an asterisk as "no AI detected"; a careful instructor reads it as "below threshold."

Confidence Threshold

The confidence threshold is the minimum certainty level a detector requires before reporting a result. Below it, the tool withholds a precise figure. This is a deliberate design choice to reduce false accusations, and it is why two documents with very different amounts of AI involvement can both display the same asterisk.

Word-Level Accuracy

Word-level accuracy measures how often a detector correctly classifies individual words as AI-generated or human-written. It is a stricter metric than document-level accuracy, which only asks whether the whole document was flagged.

Turnitin0's published research gives concrete numbers here. In 180 GPT-5.6-Sol essays totalling 156,955 words, Turnitin achieved 97.88% word-level accuracy, with Physics lowest at 88.81% and Business Administration highest at 99.67%. In 170 Claude Fable-5 essays totalling 131,451 words, accuracy reached 99.01%, with Physics again lowest at 96.52% and Criminal Justice highest at 99.80%. In 180 Gemini 3.5 Flash essays totalling 147,117 words, accuracy was 98.35%, with Information Technology lowest at 94.36% and Business Administration and International Relations highest at 99.82%.

The pattern across all three studies is consistent: detectors are strong on raw AI output, and discipline matters — quantitative and technical subjects produce more variable results than business and humanities writing.

False Positive (AI Context)

A false positive in AI detection is human-written text incorrectly flagged as AI-generated. This is the single biggest fear in the field, and the evidence is genuinely reassuring. In 504 human-written graduate-level PLOS essays totalling 135,712 words, Turnitin achieved 100.0% word-level accuracy with no false positives across 18 majors. In 340 human-written ESL undergraduate essays totalling 263,329 words, Turnitin again achieved 100.0% word-level accuracy across all domains, majors, and word-count buckets.

That said, false positives are a recognised risk across the industry. SubmitSense, an independent checking service, acknowledges that AI detectors can produce false positives or false negatives, that human writing can be flagged, and that AI detection should not be the sole basis for any decision and requires human judgment. That is an honest position, and it is worth remembering when you read any detector's marketing.

False Negative (AI Context)

A false negative is AI-generated text that a detector fails to flag. This is the failure mode that matters for institutions rather than students. The same SubmitSense disclosure covers it: detectors can miss AI content as well as over-flag human content.

AI-Polished Writing

AI-polished writing is human-written text that has been run through an AI tool for grammar, clarity, or style improvements. It sits in the awkward middle ground between fully human and fully generated, and it is where detectors struggle most.

Turnitin0's research on this is the most striking finding in the set. In 500 AI-polished graduate essays totalling 132,275 words, Turnitin achieved only 47.54% word-level accuracy — barely better than a coin flip — and some majors scored 0%. If you use AI to "just clean up" your prose, you are operating in the least predictable zone of the entire detection landscape.

Perplexity and Burstiness

Perplexity

Perplexity measures how surprised a language model is by the next word in a sequence. Human writing tends to have higher perplexity — we choose unexpected words, make idiosyncratic transitions, and occasionally write something a model would never predict. AI writing tends to have lower perplexity because models gravitate toward the statistically likely next token.

Low perplexity is not proof of AI authorship. Technical writing, legal prose, and heavily templated academic formats all produce low perplexity naturally. This is one reason Physics consistently scores lowest in Turnitin0's accuracy studies: physics writing is formulaic by necessity.

Burstiness

Burstiness measures variation in sentence length and structure. Human writers are bursty — a long, clause-heavy sentence followed by a short one. AI output tends to be uniform, with sentence lengths clustering around a comfortable average.

Burstiness is the more intuitive of the two concepts. If you read your draft aloud and every sentence takes roughly the same breath to deliver, your burstiness is low.

Perplexity and Burstiness Together

Detectors typically combine both signals with many others. Neither is decisive alone, which is why no one can give you a reliable formula for "how much perplexity do I need." The honest answer is that these are contributing features, not thresholds.

Humanization and Evasion Terms

AI Humanizer

An AI humanizer rewrites AI-generated text to read more naturally and to reduce detector signals. Turnitin0's humanizer preserves meaning, citations, headings, and.docx formatting — a meaningful detail, because a humanizer that mangles your citation style creates more work than it saves.

Evasion Rate

Evasion rate is the percentage of text that successfully avoids being flagged. Turnitin0's research on humanization reports that across 174 humanized essays totalling 204,736 words, the service achieved 76.44% word-level evasion against the Turnitin AI detector, with Education at 100% and English lowest at 55.41%.

Read that range carefully. A 76.44% average with a 55.41% floor in one subject means humanization is a probability improvement, not a guarantee. Any service claiming otherwise is overpromising.

Word-Level Evasion

Word-level evasion applies the evasion concept at the individual word rather than the document level. It is a more granular and more honest measure than "did the whole document pass," because partial detection is the realistic outcome in most cases.

How to Use These Terms in Practice

Vocabulary is only useful if it changes what you do. Here is the practical sequence.

Before you submit anything, run a non-repository check. Turnitin0 accepts.docx,.pdf, or.txt files, English documents only, with a word count greater than 300 and less than 30,000, and a file size under 20 MB. You get two downloadable PDFs — the AI detection report and the similarity/plagiarism report — and the turnaround is under 15 minutes in 98% of cases, with most orders finishing in 5–15 minutes and a guaranteed 30-minute delivery during rare queue spikes.

When you read the reports, apply the vocabulary. A yellow similarity band (25–49%) with matches concentrated in your reference list is a non-issue [3]. A 40% AI score concentrated in your literature review is a rewrite target. An asterisk means below the 20% confidence threshold — better than a number, but not a guarantee.

When you revise, distinguish between the two problems. Similarity problems are solved by paraphrasing and citation. AI detection problems are solved by adding genuine human variance: your own analysis, your own examples, your own sentence rhythm. If you need help with the second, the AI humanizer is the targeted tool.

What Real Users Report

The vocabulary above is abstract until you see how it plays out. Turnitin0's user feedback is consistent on the operational side. Raini Dipré (CA) described the process as easy, fast, and efficient, with the report arriving much faster than expected. May Zin (SG) received a complete report in about 20 minutes and could download the AI and similarity score reports at the same time. Daniela Pellegrini (GB) has used Turnitin0 several times, reports quick delivery, and found the humanizer helpful when revising. Shubham Pachauri (IN) found it quick and helpful for checking and improving academic writing, and specifically liked the humanizer for making text sound more natural.

Others describe straightforward utility: B C (US) used the site for assignments, plagiarism checking, and awareness of AI; Shawn Thakur (AU) found it easy to use and on time; Encrypted (GB) called it the best site for Turnitin scans, authentic and simple to use; Taksh Patel (AU) called it a great service, 100% legit and working.

Across the platform, Turnitin0 reports 100,000+ AI and similarity reports delivered to 20,000+ students worldwide, primarily in the United States, United Kingdom, Canada, Australia, New Zealand, and Ireland, with a 4.9/5.0 satisfaction rating. Of humanizer orders, 98.2% are re-checked with Turnitin.

Honest Limitations

A glossary that only lists strengths is marketing, not reference material. Here is what Turnitin0 does not do.

Turnitin0 is an independent service and is not affiliated with Turnitin, LLC. It handles English documents only, for both checking and humanization. The checking service requires a word count greater than 300 and less than 30,000, with a file size under 20 MB; the humanizer accepts files under 90 MB. There is no free word quota or free trial for the humanizer. On Trustpilot, the profile carries a 4.3/5 TrustScore with an "Excellent" label but only 9 reviews, all within the last 12 months, and the company has not recently invited customers to review — so those reviews may not be representative.

None of these limitations change the core recommendation, but you should know them before you rely on the service.

How Turnitin0 Compares on the Terms That Matter

The competitor landscape is worth understanding in glossary terms, because different services define the same words differently.

T-checker (turnitinaichecker.ai) offers both similarity and AI reports by default, supports.pdf and.docx, and covers English, Spanish, Japanese, and Arabic for AI detection. It states documents are never stored or added to any repository, with files automatically cleared within 24 hours. Its word range is 320–29,999 and processing takes about 5–20 minutes. The trade-off: no independent user feedback is available to verify the vendor's claims, and credit rules can deduct a credit even when only one check type is run.

SubmitSense (aiturnitinchecker.com) supports DOCX, PDF, TXT, and RTF, accepts 400–30,000 words at up to 40 MB, claims most reports in 5–15 minutes and 99% within 30 minutes, and deletes files within 24 hours with reports accessible for 30 days. It is notably candid that AI detection can produce false positives and false negatives and requires human judgment. Again, no independent user feedback was available.

TurnitChecker (turnitchecker.ai) offers private non-repository processing, downloadable PDFs, and an AI refinement service for flagged English passages taking 10–30 minutes. Its limits are tighter — 400–28,000 words, 10 MB maximum, essay/thesis format only — and it notes that new standard purchases are temporarily limited to single checks while report delivery is monitored.

ZeroGPT is a free detector with a broad tool suite including a humanizer, plagiarism checker, paraphraser, and summarizer, with a 15,000-character input limit and batch file upload. It is a useful first-pass tool, but it is not a Turnitin report and does not replicate what your instructor sees.

MyDetector.ai offers free detection with sentence-level highlights, file uploads in TXT, DOCX, PDF, and PPT, and a 200,000-character limit, though it enforces an unspecified daily usage quota.

humanizeai.pro is a free humanizer supporting.txt,.docx,.pdf, and.md with custom style and tone settings, but it requires a Captcha and may need multiple iterations.

EssayDone positions itself as a broader AI writing assistant claiming a 10x speed boost, automatic citations, and bypassing 12+ detectors including Turnitin. All of these are vendor homepage claims with no independent verification provided.

The pattern is clear. Free detectors are useful for a first signal. Multi-language services add coverage. But if you need the actual instructor-facing report — the one with the same layout, the same similarity bands, and the same AI detection output your professor will open — a dedicated non-repository checking service is the only category that delivers it.

Quick Reference Table

Conclusion

The vocabulary in this glossary exists to help you make better decisions, not to intimidate you. Similarity scores measure overlap, not guilt [3]. AI detection scores measure statistical signals, not authorship. Asterisks mean below-threshold confidence. Perplexity and burstiness explain why human writing looks different from generated writing. Evasion rates describe probabilities, not guarantees.

The practical conclusion is the same one this article opened with: run your own draft through the Turnitin check service at Turnitin0 before you submit, read both PDF reports with the terms above in hand, and revise with the AI humanizer if the AI detection report shows a problem. The reports mirror what your professor sees, delivery is under 15 minutes in 98% of cases, and your file is not added to Turnitin's student paper database. Understanding the terminology is the first step; seeing it applied to your own writing is the one that actually protects your work.