Paste your text or upload a file. Our detector analyzes every sentence, shows you exactly which parts look AI-generated, and gives you an overall verdict.
Paste your text into the box below or upload a PDF or DOCX file. Documents under 1 MB are accepted. Minimum length is 180 characters so the detector has enough context to produce a reliable verdict.
When the analysis finishes, you see three views: a doc-level verdict card, a pie-chart distribution of AI-flagged sentences, and a sentence-by-sentence diff that highlights exactly which parts look AI-generated. You can save the whole report as a PDF from the “Save as PDF” button in the results header.
How the detector works
Our detector is powered by ModernBERT — a fine-tuned 149-million-parameter transformer that reads your text and predicts, sentence by sentence, whether it was generated by a large language model like GPT, Claude, Gemini, Llama, DeepSeek, or Mistral.
The model was trained on a balanced corpus of 1,200 samples — 600 written by humans (PAN25 + PERSUADE essays) and 600 generated by modern LLMs. Training used focal loss with hard-negative mining to handle ambiguous text, and the decision threshold was calibrated on a held-out validation split.
For each sentence we compute an AI probability between 0% and 100%. Sentences above 50% are marked red (likely AI), 25–49% amber (possibly AI), below 25% green (likely human). The document verdict is the mean across all analysis windows.
What makes our detector different
Most AI detectors give you one opaque number. We give you the full picture: a sentence-level heatmap, a distribution pie chart, and a diff-style view showing exactly which sentences look AI-generated and why.
Our published accuracy is AUC-ROC 0.9884 on a 1,000-sample validation set covering 22 generators including GPT-4, GPT-4o, GPT-5, Claude 4, Gemini 2, Llama 3, DeepSeek, Qwen, and Mistral. At a 50% threshold: zero false positives on the test set, 60% recall. (methodology and dataset)
No third-party API dependency. No training on your submissions. Every check runs on GPU infrastructure we own and operate — no SaaS passthrough margin, no ToS risk, no vendor lock-in.
Check Your Text with Plagiarism Detector
Download a free demo or purchase a license to start checking for plagiarism and AI-generated content.
What AI models we detect
Our training corpus includes samples from the major frontier models across vendors, so the detector generalises across the current LLM landscape.
OpenAI: GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4o, GPT-5.0, GPT-5.3, GPT-5.4. Anthropic: Claude 3 Opus, Claude 3.5 Sonnet, Claude 4 Opus, Claude 4.5 Sonnet. Google: Gemini 1.5 Pro, Gemini 2.0, Gemini 2.5. Meta: Llama 3.1, Llama 3.3. Others: Qwen 2.5, Qwen 3, DeepSeek R1, Mistral Large, o3-mini, and more.
Limitations you should know
AI detection is probabilistic, not evidentiary. Treat a high AI score as a signal to investigate further, not as proof. False positives are possible — especially on creative writing, heavily edited text, and non-native English.
Shorter text (under 100 words) is harder to classify reliably. Paraphrased or heavily edited AI text can evade detection. Essays that went through a humaniser tool may score as human even when they were originally generated.
Creative writing and literary analysis are the weakest domain for any AI detector — human style in fiction can converge with LLM outputs. If your use case is academic integrity, pair this tool with source-matching plagiarism detection (see our desktop product below).
Frequently Asked Questions
Is this AI detector really free?
Yes, within a daily allowance: one free check a day without signing up (a document of up to 3,500 words), or three a day with a free account. For more, buy credits (1 credit = one document of up to 3,500 words). For documents of any length, offline scanning and batch analysis, use our desktop app - free demo available.
How accurate is this detector?
AUC-ROC 0.9884 on our internal 22-generator validation set. Accuracy varies by domain — academic essays are detected more reliably than creative writing. See the “Limitations” section above for the honest trade-offs.
Do you store my text?
Submitted text is stored only long enough to compute the result and is purged on a rolling basis. We do not train our models on user submissions, and we do not share data with third parties.
Can I check documents offline?
Yes — our desktop Plagiarism Detector includes the same AI-detection engine and runs fully offline. Unlimited document length, no network required, built-in plagiarism matching against 4 billion web pages.
What’s the difference between AI detection and plagiarism detection?
Plagiarism detection finds copied or paraphrased text matching existing sources. AI detection identifies text generated by an AI model. They are complementary — our desktop product does both in a single scan.
Who is behind this detector
Built by Yurii Palkovskii and the Plagiarism Detector team
Detection software since 2007, made in Zhytomyr, Ukraine. Our research system placed 3rd of 34 teams at PAN 2026, the academic benchmark for AI-text detection, with the highest ROC-AUC of any team (0.997) and no false positives on human text.
By 
Stand with Ukraine!
