AI content detector — a signal, not a verdict
AI detection is coming soon — it's on our roadmap, not live yet. When it ships, every AI detector — including ours — will be wrong sometimes. Noplag will return a calibrated likelihood score, not a yes/no verdict, and publish its false-positive rate on non-native English writers. We're building coverage for ChatGPT, Claude, Gemini, GPT-5, and Llama. Read what the score will actually mean before you act on it — especially if you're making a decision about a student or a job applicant. Detection is one signal among many. It should never be the only one.
Detection signatures for every major LLM.
Each model leaves a different statistical signature. We track them per-release and publish per-model calibration so you know what the score reflects for the LLM that wrote the text you're checking.
ChatGPT
gpt-4o, o1, o3 families. We track output drift across releases; per-model calibration retrained within 30 days of each release. Report cites the model family the signal best matches.
Claude
Sonnet, Opus, Haiku tiers across 3.x and 4.x. Detection signal varies by mode (thinking vs. fast); calibration noted per mode in every report.
Gemini
1.5 Pro, 2.0 Flash, 2.5 Pro. Multilingual signal is weaker for non-English text; report surfaces that as 'low-resource confidence' rather than hiding it.
GPT-5
OpenAI's August 2025 release. Behavioural signature differs from gpt-4o; we retrain against the model card's reference outputs within 30 days of release.
Llama
3, 3.1, 3.3 family. Self-hosted variants are harder — finetuning shifts the signature. Report says 'likely LLM-generated, family Llama' rather than naming a specific version.
Every AI detector over-flags non-native English writers. Here's what we do about it.
Liang et al. (Stanford, 2023) measured a 61% false-positive rate on TOEFL essays — six times the rate for native writers. Every detector. Including ours, before we corrected for it.
The Stanford paper
Liang et al. (2023) tested every commercial AI detector against TOEFL essays by non-native English speakers. The benchmark detector flagged 61% as AI-generated — six times the rate for native writers. Same texts, same task.
What we recalibrated
We retrained the classifier with ESL-style features stripped — uniform sentence length, simpler vocabulary, fewer idioms aren't weighted as AI signals. Style-blind features only. We predict from structure, not from polish.
Residual error rate
After the fix: a documented, per-language confidence interval rather than a single headline number. Per-language tables publish when the feature clears evaluation — exactly what an institutional audit would ask for.
When to override
If you're evaluating an ESL writer and the score lands in the 20–80% band, treat it as advisory only. We surface 'low confidence — interpret with caution' explicitly in the report so the user can see when not to act.
What we still miss
Hybrid writing — LLM-drafted, native-editor-polished — is harder than either pure case. The report flags 'mixed signal, likely hybrid' rather than guess; some of the residual error lives here.
Where we measure
Public TOEFL essay sets, Pan-Hispano academic submissions, and CEFR-graded multilingual writing. Same data Liang et al. used, plus 8 newer corpora collected 2024-2026. References publish with the methodology when the feature ships.
A probability, not a verdict.
The score is a likelihood between 0% and 100%. Above 80%: high confidence AI-generated. Below 20%: high confidence human. The 20–80% band is where you treat it as one signal among many, not as proof. Use detection to start a conversation, never to end one.
High confidence the text was authored by a human. The false-positive rate for this band publishes per language when the feature clears evaluation.
The model isn't confident. Treat the score as one input among many — recent writing samples, prior drafts, in-class output. Do not act on this band alone.
High confidence the text was produced by a large language model. Still not a verdict. Combine with context (prompts, browsing history, oral defense) before any decision.
- 5LLM families targeted at launch
- 61%ESL false-positive rate, Stanford '23
- 30dplanned retrain cadence after each model release
- v1.2release target for AI detection
- 2024detection R&D begins
We're on the integrity side. We don't ship a tool to defeat detection.
A humanizer is a tool to defeat AI detection — it rewrites LLM output to lower the detection score. We could build one. We don't. The whole point of this product is to surface AI use, not help hide it. If detection is going to be useful, the people running detectors and the people building humanizers can't be the same vendor. We've picked our side. When we mention humanizers in this product, it's to tell you they exist and may be in use upstream of the text you're checking — not to sell you one.
Read the full positionHow we compare to GPTZero, Originality.ai, and Winston.
Education-focused, no plagiarism layer, opaque calibration. AI detection only; if you also need plagiarism in one report, that's two tools instead of one.
Detection + plagiarism + open-source engine. Per-model calibration data published. Same report, same vendor, auditable end-to-end.
'97% accuracy' marketing. Published false-positive rate undisclosed. English-only calibration. Closed source — you can't audit the score behind the score.
Honest about ESL bias (we cite Liang 2023 by name). Multi-language calibration. Open-source engine; the methodology publishes with the feature.
Strong on ChatGPT detection. Weaker signal on Claude and Gemini. Expensive at volume; pricing tier-gated past 10k words/month.
Covers every major model family with equal calibration effort. Flat-rate plans through Pro tier. Self-host option for compliance-bound buyers.
Twelve questions, honest answers.
If we don't know the answer, the answer says so. We update this list when calibration changes.
- Is AI detection accurate?
- Honest answer: no detector is reliable on text under ~200 words, and every detector has false-positive rates that vary by writing style, language, and which LLM generated the text. Noplag commits to publishing its calibration data so you know what 'accurate' means for your specific use case; it publishes alongside the feature.
- Will Noplag flag my non-native English writing as AI-generated?
- It might — and we're transparent about that. Liang et al. (Stanford 2023) showed every commercial AI detector over-flags ESL writers. Our safeguards cap false-positive rates for known ESL patterns, and we publish the residual error rate. If the signal is uncertain, the report says so.
- What AI models will Noplag detect?
- At launch, AI detection will cover ChatGPT (gpt-4o, o1, o3), Claude (Sonnet / Opus / Haiku, 3.x and 4.x), Gemini (1.5 / 2.0 / 2.5), GPT-5, and Llama 3.x. Detection is coming soon; detection quality varies by model, and the report will tell you which model signature the score reflects.
- How is this different from GPTZero or Originality.ai?
- We bundle AI detection with plagiarism checking, and ship the engine as open source. GPTZero is detection-only. Originality.ai is closed-source with undisclosed ESL-bias data. See /vs-gptzero and /vs-originality-ai for line-by-line comparisons.
- Can students bypass this detector with a humanizer?
- Yes — humanizers exist, and we have to be honest about that. Heavy rewriting after an LLM generates the draft will reduce the detection signal. We don't ship a humanizer (it's anti-aligned with our mission). Treat AI detection as one input, never as a verdict.
- What does a 'false positive' mean here?
- Human-written text scored as AI-generated. We will measure and publish the rate per language and writing-style cluster. It's not zero on any detector — treating a high score as proof rather than as a signal worth investigating is the misuse pattern we built this page to discourage.
- Do you sell a 'detection at 99% accuracy' guarantee?
- No. Anyone claiming above ~85% sustained accuracy across writing styles, languages, and LLMs is misrepresenting what the math allows. We publish the actual confidence interval per language when the feature clears evaluation. We don't sell certainty.
- Is AI detection free?
- AI detection is coming soon. When it launches, it'll be free on the free tier — same 2,500-word limit as plagiarism checking. Paid tiers raise the limit; AI detection will never be paywalled behind a separate add-on.
- How fresh is the model coverage?
- We retrain detection against newly-released LLMs within 30 days. If a model isn't yet in our calibration set, the report says 'likely LLM-generated, model unknown' rather than guess at a family.
- Do you handle multilingual AI detection?
- English, Spanish, French, German, and Portuguese have validated calibration — English is the strongest signal. Other languages return a result flagged 'low-resource — interpret with extra caution.' The report tells you which calibration tier applies to the text you submitted.
- Can I integrate AI detection via API?
- Soon. AI detection is on the roadmap; when it ships, the /v1/checks endpoint will return both plagiarism and AI-detection scores. See /plagiarism-checker-api and /docs/developers/api.
- What happens if I disagree with a score?
- Tell us. We log corrections, retrain within 30 days of each model release, and publish the model's behaviour change between releases. Detection isn't perfect; we'd rather hear when it's wrong than pretend it isn't.
AI detection — coming soon, free up to 2,500 words.
Calibration data, false-positive rates, and per-model signatures will all be published. See what's behind the score before you act on it.
Notify me when it's live