AI Content Detector

The AI Content Detector tells you whether a text was likely written by a person or generated by an AI model such as ChatGPT, Claude, Gemini, Copilot, Llama or DeepSeek, and, more usefully, which sentences look AI-written and why. Paste up to 2,500 words or upload a .txt, .docx or .pdf file. You get an AI likelihood score, sentence-by-sentence highlights, the evidence for and against AI authorship quoted from your text, a measured writing fingerprint, and practical next steps.

What you get in every report

• AI likelihood (0–100%) with a plain-English verdict and a confidence level that tells you how far to trust it.
• How much of the text is AI: the share of words in sentences highlighted as AI-generated, possibly AI, and human-written. Useful for texts where only part was generated.
• Sentence-by-sentence highlights with a score per sentence and per paragraph. Hover or tap a sentence to see why it was flagged.
• Evidence on both sides: signs of AI writing and signs of a human writer, each with a quote you can click to find in the text.
• Writing fingerprint: sentence rhythm, AI-typical wording, personal voice, concrete details and em-dash use, measured by code independently of the AI model.
• Next steps for writers: rewrites for the most AI-sounding sentences that you can drop straight into your text and re-check.
• Next steps for reviewers: questions to ask the author about this specific text, and guidance on using the result fairly.
• Copy or download the full report as a PDF.

How to use it

1. Paste your text into the box, or click Upload file (or drag a file onto the box). At least 40 words; 150+ words gives the most reliable result.
2. Click Check for AI. The analysis usually takes 5–10 seconds.
3. Read the verdict and the confidence level, then look at the highlighted sentences to see where the AI-like writing is.
4. Open Why it scored this way to see the evidence behind the score.
5. If it is your own writing, use the suggested rewrites, then click Check again. If you are reviewing someone else’s work, use the suggested questions.

Not sure yet? Open one of the sample reports under the box: they are real reports for built-in example texts and don’t use one of your checks.

How the detection works

Your text is first split into paragraphs and sentences. Two independent kinds of analysis then run on it:

1. Sentence-level AI analysis

An AI language model instructed with the known habits of today’s chatbots reads the whole document and scores every sentence in the context of its paragraph. It looks for patterns such as stock vocabulary (“delve”, “tapestry”, “pivotal”, “navigate the complexities”), formulaic framing (“It’s not just X, it’s Y”, “In today’s fast-paced world”), neat rule-of-three lists, balanced hedging, generic claims and tidy summary lines. It also looks for what people do and chatbots rarely do: lived personal detail, specific names and numbers, humour, strong opinions, digressions and an uneven rhythm. It must quote the text for every piece of evidence it reports.

2. Measured writing fingerprint

Code measures properties that need no judgement: how much sentence length varies (human writing is “bursty”, AI text is often even), how often AI-typical phrases appear, how much first-person voice and contractions there are, how many concrete names and numbers appear, and how often em dashes are used.

Each sentence’s final score combines its own evidence with the judgement of the whole document and the measured fingerprint, because a single sentence on its own carries little evidence. The overall AI likelihood is the average of the sentence scores, weighted by sentence length, so the headline number always matches what you see highlighted.

How to read the result

Sentence colours: AI-generated (65%+), possibly AI (40–64%), and unhighlighted human-written (under 40%).

Accuracy: what this detector can and cannot do

No AI detector is 100% accurate, including the paid ones used by schools. This one is reliable on unedited chatbot output and on casual personal writing. It is less certain on:

• Short texts. Under about 120 words there is too little evidence; the report marks these as low confidence.
• Formal, academic and technical writing, which shares some traits with AI text.
• Writing by non-native English speakers, which studies have shown detectors misjudge more often.
• AI text that was heavily edited or paraphrased, and human text polished with tools like Grammarly.
• Lists, poetry, code and templates, which have little natural sentence structure to analyse.

That is why every report shows a confidence level and the evidence behind the score, not just a number. Use the result as one piece of evidence and never as the sole basis for an accusation or a grade.

If your own writing is flagged as AI

It happens to real writers, especially with polished or formal prose. Look at the highlighted sentences: usually a few stock phrases, very even sentence lengths or generic statements are responsible. Rewrite those in your own words, add a specific example or opinion, vary your sentence length, and check again. Most importantly, keep your drafts, notes and document version history; they show how your work developed and are far stronger evidence than any detector score.

If you are checking someone else’s work

Teachers, editors, clients and hiring managers use AI detectors to decide whether to take a closer look. The highlights show you exactly which passages to discuss, and the report suggests questions only the real author could answer easily. Compare with the person’s earlier writing, ask about their process, and give them the chance to show drafts before drawing conclusions.

Frequently asked questions

How does this AI content detector work?

Your text is split into sentences. An AI language model trained on the habits of ChatGPT, Claude, Gemini and other models scores every sentence in the context of its paragraph, and quotes the evidence it relied on. In parallel, the tool measures things code can count directly: how much sentence length varies, AI-typical wording, first-person voice, concrete details and em-dash use. Both are combined into the sentence highlights and the overall AI likelihood.

How accurate is it?

It is reliable on unedited AI output and on casual personal writing, and less certain in between. Every detector, including paid ones, makes mistakes on short texts, heavily edited AI text, formal academic writing and writing by non-native English speakers. The report shows a confidence level for that reason. Treat the result as one piece of evidence, never as proof on its own.

What does the percentage mean?

The big number is the AI likelihood of the whole text, weighted by sentence length. The bar underneath shows what share of the words sit in sentences highlighted as AI-generated, possibly AI, and human-written. 0–19 reads as human-written, 20–39 mostly human, 40–59 mixed, 60–79 likely AI-generated and 80–100 AI-generated.

My own writing was flagged as AI. What should I do?

Look at which sentences are highlighted and why. Usually it is stock phrasing, very even sentence lengths or generic claims. The report suggests rewrites for the most AI-like sentences. Keep your drafts, notes and document version history (Google Docs and Word both keep one): they are the strongest proof that you wrote something yourself.

Can I use this to check a student’s or writer’s work?

Yes. The report shows exactly which passages look AI-written and suggests questions you can ask the author about the text. Please use it to start a conversation, not to make a decision: false positives happen, especially with non-native English writers and formal writing.

How much text can I check?

Up to 15,000 characters, about 2,500 English words, per check. You need at least 40 words; results become much more reliable above 150 words. You can paste text or upload a .txt, .docx or .pdf file.

Is my text stored?

Your text is sent to the AI model for analysis and is never published, sold or used to train models. The finished report is kept in temporary memory for up to 6 hours so that checking the same text again gives the same answer, then it is discarded.

Further reading

• Liang et al. (2023): GPT detectors are biased against non-native English writers
• DetectGPT: Zero-Shot Machine-Generated Text Detection