Understanding AI Detection: How Turnitin Works

How Turnitin Detects AI Writing

Turnitin's AI detection works by analyzing two key statistical properties of text: perplexity and burstiness. Understanding these concepts helps you understand why AI-generated text gets flagged.

Perplexity: How Predictable Is the Text?

Perplexity measures how predictable a piece of text is. AI models are trained to produce the most probable next word, which makes their output highly predictable (low perplexity). Human writing is more varied and surprising (higher perplexity).

• AI text: Low perplexity — words follow predictable patterns
• Human text: Higher perplexity — more varied word choices and structures

Low perplexity (AI-like):  "The results of the study indicate that the
data suggest a significant relationship between the two variables."

Higher perplexity (human-like):  "We ran the numbers twice. The data
told a slightly different story the second time."

Burstiness: How Varied Is the Sentence Structure?

Burstiness measures the variation in sentence length and structure. Humans naturally write with a mix of short and long sentences. AI tends to produce sentences of similar length (low burstiness).

• AI text: Low burstiness — sentences tend to be similar in length
• Human text: High burstiness — natural variation from very short to very long

What Triggers Detection?

1. Uniform sentence length — Sentences all around 15–25 words
2. Overly formal tone — AI tends to be consistently formal throughout
3. Predictable transitions — "Furthermore," "In addition," "Moreover" used excessively
4. Perfect grammar throughout — No typos, no natural variations
5. Balanced paragraph structure — Each paragraph roughly the same length

What Turnitin DOESN'T Flag

• Heavily edited AI text with significant human intervention
• Text with natural imperfections (typos, informal language, varied structure)
• Translation from other languages
• Technical terminology that happens to be repetitive

AI Detection vs Plagiarism: Two Different Offences

Students often blur these. They are separate scores, separate rules, and separate consequences:

The Detection Score

Turnitin gives a percentage score (0–100%) indicating how much of the text it believes was AI-generated. Most institutions set thresholds at:

• Below 20%: Generally acceptable
• 20–50%: May trigger review
• Above 50%: Likely flagged for investigation

Worked Example: Why One Passage Flags and the Other Doesn't

Here are two paragraphs on the same idea. Only one trips the detector. Spot the difference:

PARAGRAPH A (flagged):
"Furthermore, the adoption of mobile payment systems has significantly
transformed financial transactions in developing economies. Moreover,
these systems provide enhanced accessibility for unbanked populations.
In addition, they reduce the cost of financial services. Consequently,
regulatory frameworks must evolve to accommodate this growth."

PARAGRAPH B (not flagged):
"We watched it happen on the street. A trader in Kumasi sells her
maize with a Momo code on a tin — no bank account, no card, no queue.
Mobile money didn't just change payments here; it made banking
invisible. And that, honestly, is the harder thing for regulators:
the old rules simply don't see the transaction happening."

• A = low perplexity (every word is the "most likely" next word) + low burstiness (uniform ~18-word sentences) + stock transitions
• B = high burstiness (a 4-word sentence beside a 30-word one) + high perplexity (surprising word choices) + a concrete, specific detail

The Three Limits of AI Detection

Every detector has hard limits. Knowing them keeps you realistic — about both evasion and false alarms:

• Probability, not truth: the score is the model's confidence, not a fact about who wrote it. It cannot "prove" anything by itself
• Length matters: below roughly 300 words, statistical signals are too weak for reliable scoring — short answers are genuinely hard to judge
• Newer models erode the signal: each generation of AI writes with higher perplexity, so detector thresholds are a moving target for everyone

How KNUST Departments Use the Score

Understanding the enforcement side helps you calibrate risk:

• Most departments treat the AI score as a screening signal, not a standalone verdict
• A flagged score typically leads to a viva-style interview about your work — knowing the material is the real protection
• Turnitin access is via the library/IT systems; students can't usually run their own institutional checks, so use the free alternatives in the Premium tier for pre-checks
• Past cases show appeals succeed when students produce draft history and research notes

Quick Reference Card

One glance at this card answers the three questions that come up most:

Reading the Turnitin Report Correctly

Most students misread the report. The key facts to understand:

• The AI score is per sentence — Turnitin highlights individual sentences it thinks are machine-written, not just a whole-document number
• A low overall % can still contain heavily flagged sections — supervisors look at those highlights, not the headline number
• The report shows two separate scores: Similarity (plagiarism) and AI. They are unrelated — you can have 2% similarity and 95% AI

How to read a report:
1. Open the AI writing report icon on the right of the document
2. Look at highlighted sentences, not the overall % first
3. Cross-check which detector type was used (AI writing vs AI paraphrasing)
4. Note the sentence count flagged vs total sentences

False Positives: Why Real Work Gets Flagged

• Academic template language — "This study aims to investigate..." repeated across sections mirrors AI patterns
• Non-native English writers — formulaic phrasing used to be safe in exams is statistically machine-like
• Lists and structured writing — bullet-heavy, evenly-spaced text scores high
• Field-specific repetition — lab reports and methods sections are inherently repetitive
• Short submissions — the detector is less reliable under ~300 words

Turnitin vs Other Detectors

Detection Models in 2026

• Turnitin's detector now covers GPT-4, Claude, Gemini, Llama and paraphrasing tools
• Paraphrase (spinner) detection is separate from AI-writing detection — reworded text still gets flagged if the rewriter keeps sentence rhythm
• Detectors can be defeated by mixing: AI for structure, human for sentences — which is exactly how responsible assisted writing should work anyway

How the Detector Scores Sentence by Sentence

Under the hood, Turnitin runs the model over the document and produces a per-sentence likelihood. Understanding the mechanics explains a lot of "weird" results:

• Each sentence gets its own probability; the overall % is the fraction of sentences above threshold, not an average
• Short sentences (under ~15 words) are evaluated with lower confidence — that's why a 5-word sentence rarely gets highlighted
• Quoted material and common-knowledge sentences are scored but marked separately
• Heavily edited AI text often scores per-sentence mixed: half the paragraph human, half machine

Interpreting a 40% score on a 10-sentence paragraph:
= 4 sentences above the machine-likelihood threshold
Find the 4 highlights → those are what a supervisor will challenge

What Supervisors Actually Do With the Report

Knowing the human behind the tool removes most of the fear:

• Most lecturers open the AI report only when the similarity score is also high, or when the writing style is conspicuously uniform
• The highlight map, not the headline %, drives decisions — a 15% overall with one fully-flagged paragraph gets more attention than a flat 40%
• Supervisors routinely ask students to explain flagged sections in person before any misconduct process starts
• In practice, a clear draft history + a calm conversation resolves most first-time flags

The Paraphrasing Detector, Explained

Many students are surprised to learn Turnitin checks two things at once:

• AI writing indicator — is the text generated directly by a model?
• AI paraphrasing indicator — was existing text rewritten by a spinner (QuillBot, Wordtune, etc.)?

Paraphrasing detection catches text that was originally human but machine-reworded, and text reworded by AI from a cited source. Rewriting the words does not reset the signal — the sentence rhythm usually survives the rewrite.

Common Mistakes When Interpreting Scores

• Ignoring the highlighted sentences — the number is not the whole story
• Assuming 0% means safe — a 0% score only means no single sentence crossed the threshold
• Panicking at 40% — if the flagged sentences are templates and methods boilerplate, a supervisor can override it
• Checking only once — Turnitin results can change slightly if you re-submit after editing; check the final version only

Pre-Submission Checklist

• Draft written yourself — AI used only for outline, structure, or grammar checking
• Sentence lengths varied — mix of very short and longer sentences every paragraph
• Transition words rotated — not "Furthermore" three times per page
• Personal evidence added — your data, your experiments, your observations
• Real citations included — every claim maps to a real source you read
• First-person voice present — where your discipline allows it
• Draft history saved — version timestamps as evidence if challenged
• Report checked sentence-by-sentence — not just the overall %
• Flagged boilerplate explained — if templates are flagged, be ready to justify them
• Plagiarism score checked separately — similarity and AI scores are independent

Troubleshooting FAQ

• "My 100% human essay was flagged" — Pull your draft history with timestamps and raise it with your lecturer. False positives are documented and supervisors can override them.
• "I used AI for grammar only and got flagged" — Grammar-checking a paragraph you wrote rarely flags it; heavy restructuring by AI does. Keep edits word-level, not sentence-level.
• "Which score matters — similarity or AI?" — Both. Check the AI score for machine-written text and the similarity score for copied text. Different offences, different consequences.
• "Can the detector tell which AI wrote it?" — No. It reports probability, not a model fingerprint.
• "Is my AI-paraphrased work safe?" — Not automatically. Turnitin has a separate paraphrasing detector and reworded text often keeps AI rhythm.
• "Why did my AI score drop after I edited?" — Because the detector re-scores sentence by sentence on the submitted version. Editing rhythm is exactly what lowers the score — that's the mechanism working as designed.
• "Can my supervisor see my AI score from last year?" — Reports are tied to submissions. A new, clean submission starts fresh; old reports matter only if reopened during an active review.
• "What should I do if flagged unfairly?" — Stay calm, gather evidence (drafts, notes, references), and request a review. Most institutions have an appeal process for AI flags.
• "Do I need to declare that I used AI?" — Check your department's policy. Many courses now ask for an AI-use declaration — being transparent is almost always safer than being caught.
• "What's the worst that can happen?" — Repeated, deliberate submission of AI text as your own work can lead to academic misconduct proceedings, including suspension. One flagged essay with evidence behind it usually ends at a conversation.
• "Does Turnitin detect my English-improved text?" — If you wrote the ideas and only used AI for grammar, the rhythm is yours and detection rarely triggers. The flag appears when AI rewrites whole sentences and keeps its own rhythm.

Appeal Letter Template

If you genuinely wrote the work and were flagged, use this structure when requesting a review:

Subject: Request for review of AI-detection flag — [COURSE] — [YOUR NAME]

Dear [LECTURER / DEPARTMENT HEAD],

My submission "[ASSIGNMENT]" received an AI-detection score of [X]%.
I am writing to request a manual review because I authored this work.

Evidence I can provide:
1. Google Docs / Word version history showing the document built up
   over [DATES] — timestamps predate any AI session
2. My hand-written research notes and drafts ([describe])
3. Specific data/observations in the essay that only I could know
   (e.g. [ONE EXAMPLE])

I understand the score is a statistical tool and welcome the
opportunity to discuss my work in person.

Regards,
[YOUR NAME] | [STUDENT ID]

Extra Resources

Free Pre-Check Tools to Try

• GPTZero (gptzero.me) — free tier with per-paragraph breakdown
• Writer.com AI detector — free, quick sentence-level check
• CopyLeaks — limited free credits, high accuracy on Claude/GPT-4

Pre-check your own draft before submission, then apply the writing-style fixes from our related lesson to any flagged passages.