Was this AI-written?
Paste prose or upload a document. We estimate the probability it was written by an LLM, with the tells. It’s a confidence signal — not proof — and it reads writing style, not whether a document is genuine.
AI likelihood
—%
Run a check to see the readout — a probability, the lean, and the tells behind it.
About the AI text detector
The AI text detector estimates how likely a passage of prose was written by a large language model rather than a person. Paste text, or upload a PDF, .txt, or .md file, and it returns a probability with the specific stylistic tells behind the score — repetitive hedging, uniform sentence rhythm, generic transitions, and the like.
It reads writing style, not document authenticity. A genuine PDF can be AI-written and a forged one can be hand-typed, so this is the wrong tool for spotting a doctored payslip or a fake invoice. For that, use the document verifier.
- 01
Paste prose or upload a file
Send at least a paragraph of running text, or a PDF / .txt / .md document — the detector pulls the prose out of it.
- 02
It confirms the text is actually prose
Forms, tables, and scanned layouts aren’t writing-style questions, so the detector abstains on them instead of guessing.
- 03
The prose is weighed for AI tells
A language model scores the passage and surfaces the concrete signals — phrasing, structure, and rhythm — that drove the estimate.
- 04
You get a probability and the tells
A 0–100% likelihood, a lean (AI, human, or inconclusive), and the reasoning, so you can judge the call rather than just take a number.
What you get back
AI-written probability
The model’s confidence, 0–100%, that the prose was generated by an LLM.
Lean + reasoning
A plain-language verdict — likely AI, likely human, or inconclusive — with the why.
The tells
The specific stylistic markers that pushed the score up or down.
Abstention
On non-prose (a form, table, or scan) it returns “not applicable” rather than a misleading score.
Common uses
- Triaging essays, applications, or support tickets for AI-written submissions
- Spot-checking research notes or reports before you rely on them
- Flagging prose for human review rather than auto-rejecting it
- Teaching and editorial workflows where authorship matters
AI-writing detection reads style, and style can mislead: careful human editing can lower a score, and non-native or formulaic writing can raise one. Use it as a signal to prompt review, not as an automated gatekeeper.
Questions
- What does AI likelihood mean?
- AI likelihood is the probability, expressed 0–100%, that a passage of prose was generated by a language model rather than written by a person. It is an estimate from writing style — phrasing, rhythm, structure — not a fact about authorship. A 90% AI likelihood means the style strongly resembles machine-generated text, not that authorship has been proven; a mid-range score is genuinely inconclusive, and an honest detector says so.
- What is a good or bad AI likelihood score?
- There is no universal threshold. Treat the ends as signals — very high suggests machine-generated style, very low suggests human — and the middle as “look closer”. What makes a score usable is the reasoning beside it: this detector lists the specific tells that moved the number, so you can judge whether they apply or whether formal, templated, or non-native writing explains them instead.
- Can I check a PDF or Word document for AI writing?
- Yes — upload the file instead of pasting. PDF, DOCX, TXT, and Markdown are read, and the running prose inside them is assessed the same way as pasted text. Long documents are scored section by section — up to four sections of roughly 2,300 words each per check, with a per-section breakdown — and anything beyond that window is reported as unread rather than silently skipped. Non-prose content — forms, tables, spreadsheets — is declined rather than guessed at.
- Is this text AI?
- Paste it above and the detector returns a 0–100% likelihood, a lean — likely AI, likely human, or inconclusive — and the specific tells behind it: repetitive hedging, uniform sentence rhythm, generic transitions, and the like. Read the tells, not just the number. They are what let you judge the call instead of taking it on faith.
- How likely is it that this was written by AI?
- That is exactly what the score is — a probability, not a yes or no. A result in the middle is genuinely inconclusive, and the detector says so rather than forcing a side. Style is the only evidence it has, so treat a high likelihood as a reason to look closer, not as a finding.
- How do I check if AI wrote this?
- Paste the prose, or upload a PDF, .txt, or .md file — one free sign-in starts a 7-day trial with 200 credits, no card, and nothing is stored. Send at least a paragraph; shorter snippets do not carry enough stylistic signal to be worth scoring. You get the likelihood, the lean, and the reasoning in a single pass.
- Can you tell if something was written by AI?
- Not with certainty — no detector can, and any that claims to is overselling. Careful editing lowers a score, and formulaic or non-native writing raises one. What you get here is a defensible starting point: a probability with the reasons behind it, so a person makes the actual decision.
- Is the result proof that text was written by AI?
- No. It is a confidence signal, not proof. AI-writing detection is probabilistic — treat a high score as a prompt to look closer, never as a verdict on its own.
- Does a high score mean the text is bad or dishonest?
- No. AI-written does not mean wrong. Plenty of legitimate text is drafted or polished with an LLM. The detector reports how the text reads, not whether using AI was appropriate.
- Why did it refuse to score my document?
- It only assesses prose. If you upload a form, a table, or a scanned layout it abstains — authenticity there is a forensic question, so use the document verifier instead.
- How much text does it need?
- At least a paragraph. Very short snippets do not carry enough stylistic signal for a meaningful estimate.
- Does it detect Claude’s invisible watermark?
- Not yet — Anthropic announced invisible watermarks in text from new Claude models (from August 2026) but has not published the detection specification, so no third party can verify marks today. When the details are published, watermark verification is planned as a deterministic check alongside the style analysis. Note the asymmetry either way: a mark is strong provenance, but its absence proves nothing — most AI text carries no watermark.
- Do you store what I submit?
- Not unless you rate the result. Checking a text is stateless: it is assessed and not retained. If you rate the result, thumbs up or thumbs down, we keep the text alongside your rating for 120 days and use it only to improve the accuracy of the check, because a rating without the text it applies to cannot be acted on. Do not rate it and nothing is kept.
Developer recipes & tools
Checking student work?