Tooling

Plagiarism & AI-Content Detection

Turnitin, Copyleaks, GPTZero — and an honest take on the LLM-detection accuracy debate.

Plagiarism detection has been around for 25 years and is a (rough) solved problem. AI-text detection is not. Multiple peer-reviewed studies (2023-25) have shown all major AI-detection tools produce 5-30% false positives — disproportionately on non-native-English writers. Bake that into your assignment policy. Pair this page with grading & feedback and exam proctoring, but read the honesty note first.

Plagiarism detection (mature, established)

  • Turnitin — paid; the higher-ed monopoly. Site-licence model; integrates with Canvas / Blackboard / Moodle / Brightspace via LTI. Database includes student submissions globally — competitors can't replicate that scale. Includes Turnitin AI Writing Detection (controversial, see below).
  • iThenticate — paid; from Turnitin; focused on academic / journal submissions.
  • Copyleaks — paid + free trial; broader content-check; agency-friendly; multilingual; offers Copyleaks AI Detector.
  • PlagScan / Plagiarism Checker X / Quetext — paid + free tier; smaller-scale.
  • Plagiarism Detector, DupliChecker, Grammarly Premium plagiarism check — consumer-grade.
  • PlagiarismFinder, Compilatio, Urkund (Ouriginal) — EU-popular.

AI-content detection (handle with care)

  • GPTZero — paid + free tier; the brand-name; founded by a Princeton student in 2023. Newer model versions claim better accuracy, especially on long-form. Still: false-positive risk is real.
  • Turnitin AI Writing Detection — bundled in Turnitin; flagged "% likely AI-generated"; Turnitin has acknowledged accuracy limits and recommends it as a signal, not a verdict.
  • Copyleaks AI Detector — bundled with Copyleaks; multilingual; similar caveats.
  • OriginalityAI — paid; SEO / agency market; aggressive marketing; same accuracy caveats apply.
  • Winston AI, Sapling AI Detector, ZeroGPT, Crossplag — alternatives; no single one is reliably better than coin-flip on adversarial / paraphrased text.
  • OpenAI's own detector — was withdrawn by OpenAI in July 2023 due to "low rate of accuracy." This is the most honest data point in the field.

Honest take (read this before you pay)

  • False positives are not rare — published rates of 5-30% across these tools, with bias against ESL students, students with disabilities, and certain writing styles.
  • Adversarial paraphrasing (running output through another LLM, or tools like Undetectable.AI, StealthGPT) defeats detectors easily.
  • Recommended policy — treat AI detection as one signal among many; never auto-fail on it; pair with process artefacts (drafts, version history) and viva / oral defence; teach assignments that are AI-resistant by design.
  • Open letter from EFF, ACLU, and many universities (2023-24) discourages reliance on these tools for academic integrity decisions.

What works better than detection

  • Process portfolios — require Google Docs / OneDrive version history.
  • In-class writing — pen-and-paper or exam-locked browsers.
  • Oral defence — 5-minute viva on the submitted work.
  • Personalised prompts — tied to specific class material the LLM hasn't seen.
  • Annotation / sourcing — require pull-quotes from class readings.
  • Lower the stakes — if AI use isn't catastrophic, the arms race is wasteful.

Self-host / open

Pick this if…

  • Higher-ed default, must integrate with LMS: Turnitin (you probably already have it).
  • Cheaper plagiarism alone, free tier: Quetext or Copyleaks.
  • Code plagiarism in CS class: MOSS or JPlag.
  • AI detection because admin requires it: GPTZero or Turnitin AI — but document the false-positive policy.
  • Better than detection: redesign assignments — viva, version history, in-class writing.

On this page