Paste your text below and get a 0-100 AI probability score in seconds, plus a Humanization Score that tells you how natural it reads. One scan covers GPT-4, Claude, Gemini, and Llama. Sentence-level highlights ranked by suspicion come with Pro.
Paste some text or load the example to estimate AI likelihood.
Signed out you get 3 checks a day, pooled with the document detector, at up to 1,000 words a scan. A free account raises the per-scan length to 5,000 words. Scans are not stored while you are signed out. How we measure accuracy →
Drop text directly, or upload a PDF or DOCX file. Signed out you can scan up to 1,000 words at a time; a free account raises that to 5,000.
50+ language signals — burstiness, perplexity, lexical patterns, and structural markers — checked against OpenAI, Anthropic, Google, and Meta output fingerprints.
You get an overall AI-probability score, a Humanization Score, sentence-level highlights ranked by suspicion, and a built-in humanizer that rewrites only the flagged sentences.
Run your draft before submission. See which sentences look AI-generated and rewrite them in the same scan — without uploading to a third-party humanizer first.
Check guest posts and agency-written drafts before publishing. The Humanization Score gives you a measurable benchmark to track improvements as you rewrite.
Sentence-level highlights give you specific evidence to discuss with students — not a single overall score that's hard to defend.
Bulk-scan deliverables before they ship. API access on the Business plan ($15.99/mo billed yearly) covers 10,000 calls/mo — a fraction of Originality.ai's Enterprise tier.
We do not publish an accuracy percentage. Results vary with text length, content type, writing style, AI model and editing level, and formal, encyclopedic human writing is the kind most likely to be wrongly flagged. On our published false-positive benchmark the detector wrongly flagged 5.85% of 1,180 human-written academic abstracts. AI detection is probabilistic — no detector is perfect, and we publish our caveats openly.
What works well: long-form English text (100+ words), recent commercial models (GPT-5, Claude Opus 5/Sonnet 5, Gemini 3, Llama 4), and unedited or lightly edited AI output.
Where to be cautious: short snippets (under 50 words), heavily human-edited AI text, non-English languages, and text that intentionally mimics a specific human author's voice. We flag low-confidence scores in the result so you know when not to rely on them.
TextSight analyzes your text against signals trained on GPT-5, Claude 5, Gemini 3, and Llama 4 outputs. It returns an overall AI-probability score (0-100), sentence-level highlights ranked by suspicion, and a Humanization Score that tells you how natural the text reads. The scan typically completes in under three seconds.
Yes. You get 3 checks a day with the same detector and Humanization Score as the paid plans, pooled across the text box and the document detector. No credit card and no signup are required for your first scans. Higher limits start at $9.99/mo on the Starter plan.
We do not publish an accuracy percentage, because results vary with text length, content type, writing style, AI model and editing level. On our published false-positive benchmark the detector wrongly flagged 5.85% of 1,180 human-written academic abstracts. AI detection is probabilistic — read /accuracy-methodology for our test setup, false-positive rate, and the caveats we apply to every score.
Yes. The detector is trained across all four major model families — OpenAI (GPT-5), Anthropic (Claude Opus 5, Sonnet 5), Google (Gemini 3), and Meta (Llama 4). Newer model releases are added to the training set as they ship. We also publish a dedicated ChatGPT detector page focused on OpenAI-specific signals.
Detection tells you which parts of your text read as AI-generated. Humanization rewrites the flagged sentences so they read more naturally — TextSight ships both in the same scan. After you get a score, one click runs the AI Humanizer on just the flagged sentences and re-scores.
The detector is English-focused. It works on other Latin-alphabet languages but accuracy drops outside English. For comprehensive multilingual detection, dedicated multilingual tools are a better fit.
Scans run while signed out are processed in memory and discarded. Signed in, the scan and its text are saved to your history (switch that off in Settings, or delete any scan); Free accounts see the last 7 days and paid plans keep the full history — see /privacy for the full retention table.
The detector needs at least 50 words of signal to be reliable. 100+ words is the recommended minimum for confident scoring. Below 50 words, the score is shown but flagged with low confidence.
Checking text from a specific model? Each page is tuned to the signals that model leaves behind.
Country guides to how AI detection is used in schools and workplaces in each market.
Role- and model-specific guidance for the way you actually work.
Step-by-step guides for checking and spotting AI text.
Most detectors hand you one opaque score and ask you to trust it. TextSight shows the signals, the bands, and the limits. Here are the six design choices that drive every scan.
The detector is calibrated on outputs from every major commercial model family: GPT-5 (OpenAI), Claude Opus 5 and Sonnet 5 (Anthropic), Gemini 3 (Google), Llama 4 (Meta), and Mistral Large 3. Single-vendor detectors miss Claude or Mistral content because they only trained on GPT samples. TextSight catches all five families in one scan.
A single overall score hides the truth. TextSight scores every sentence independently and colour-codes the result, so you can see which lines actually triggered the verdict. That matters when a student needs to defend an essay or an editor needs to rewrite only the offending sections instead of starting over.
Human writing varies. Sentences expand, contract, surprise. AI writing tends to settle into a steady predictable rhythm. The detector measures perplexity (how surprising each token is) and burstiness (how much sentence length and complexity vary), alongside dozens of lexical and structural markers. It is statistical pattern recognition, not magic.
A 50-word snippet and a 5,000-word essay are not the same problem. Short text is noisy, long text accumulates signal. The detector adapts its confidence threshold based on input length, flags low-confidence scores below 50 words, and processes documents up to the upload cap on paid plans without losing the cross-sentence context that drives accuracy.
Academic prose reads differently from a Substack post, and both differ from a Slack message. The training set spans academic essays, journalistic features, conversational blog posts, technical documentation, and creative fiction. That breadth keeps false positives lower on ESL writing, STEM papers, and tightly edited corporate copy than detectors trained on a single register.
Every scan returns one of five clear bands: Original, Mostly Human, Mixed, Likely AI, AI Generated. Borderline scores are labelled as such, low-confidence inputs are flagged before you read the verdict, and the methodology page documents the test set and false-positive rate. You see the number and the caveat together, not the number alone.
Four tiers, one detector. The right plan usually comes down to two questions: how many documents do you scan per week, and do you need the REST API for product or agency use?
Best for: students sanity-checking essays a few times a week, writers spot-checking a single piece before submission, and anyone evaluating the detector before committing.
3 free scans a day: up to 1,000 words each with no account, or up to 1,500 words with a free account, with one overall AI score and a Humanization Score per scan (sentence-level highlights start on Starter), and a separate daily allowance for the one-click AI humanizer at 260 words per humanization.
Best for: active students with three to five essays a week, casual bloggers running drafts through detection before publish, and teachers running ad-hoc spot checks.
20 scans/day, 40,000 humanizer words/month, Chrome extension, file upload, plagiarism risk indicator, full history retention.
Best for: freelance content writers shipping 10+ deliverables a week, daily newsletter authors, and SEO writers running every draft through detection before invoicing.
100 actions a day, 80,000 humanizer words/month, file upload, priority support, white-label PDF export, and full-text history.
Best for: SEO agencies auditing client work, EdTech and SaaS teams embedding detection via API, and educators scanning whole-class submission batches.
100,000 humanizer words/month, REST API access with 10,000 calls/month, 5 team seats, white-label reports, and full-text history.
Annual billing saves 60%, dropping Pro to $7.99/mo and Business to $15.99/mo. See full pricing →
The questions that matter most when picking a detector. Direct answers, no marketing dodges.
We do not publish an accuracy percentage. Accuracy moves with length, genre and the model that wrote the text: short text under 100 words gives the least signal, and formal, encyclopedic human writing is the kind most likely to be wrongly flagged. On our published false-positive benchmark the detector wrongly flagged 5.85% of 1,180 human-written academic abstracts. Anyone advertising 99 percent accuracy across all content lengths is overclaiming, and we publish the test methodology so you can check what we do claim.
On our published benchmark the detector wrongly flagged 5.85% of 1,180 human-written academic papers. That is not zero, which is why TextSight ships sentence-level evidence rather than a single overall verdict. When a borderline score comes back, you can see which specific sentences triggered the signal and decide whether they actually look AI-generated or whether the writer simply has a formal style. The methodology page documents how we measure the rate and on which test corpora.
Better than detectors trained only on native-English corpora, but ESL writing still carries higher false-positive risk because non-native English often shares some surface features with AI text (simpler sentence structures, more consistent vocabulary, fewer idioms). The training set includes a substantial ESL sample, which keeps the rate lower than competitors, but we still recommend ESL writers rely on the sentence-level highlights and Humanization Score rather than the overall band when stakes are high.
Technical writing is harder for every detector because dense vocabulary, formulaic structures, and conventionalised phrasing look statistically similar to AI output. TextSight is calibrated against a technical-writing subset (research papers, documentation, white papers) so the false-positive penalty is reduced, but STEM authors should treat borderline scores as a prompt to inspect the highlighted sentences rather than as final judgements. The Humanization Score is often more useful than the AI probability for technical content.
The detector is retrained on a rolling basis as major model releases land. When OpenAI, Anthropic, Google, Meta, or Mistral ship a new flagship version, we capture sample outputs across a range of prompts and writing styles, evaluate detection drift, and roll an updated detector typically within 4 to 8 weeks of release. The methodology page lists the currently-supported model versions and the date of the most recent training refresh.
The detector is English-first. It runs on other Latin-script languages (Spanish, French, German, Portuguese, Italian) but accuracy drops because the training corpus is heavily English. For Cyrillic, Arabic, CJK, or Devanagari scripts we recommend a dedicated multilingual detector. If you need detection in a specific non-English language, contact support and we will share the calibration data we have for that language so you can decide whether it meets your bar.
Yes. REST API access is included on the Business plan ($39.99/mo) with 10,000 calls per month, full sentence-level response payloads, SSE streaming for the humanizer endpoint, and documented rate limits. EdTech platforms, LMS vendors, agency CMS integrations, and content-quality tools all use the same endpoints the dashboard uses. The full reference is at api-docs, including authentication, request/response schemas, and example payloads.
Detection accuracy is strongest on the major commercial models that dominate the training set: GPT-5, Claude Opus 5, Claude Sonnet 5, Gemini 3.1/3.7, Llama 4, and Mistral Large 3. Accuracy is solid on smaller variants (Claude Haiku, GPT-4o-mini, Gemini Flash) and degrades gradually on heavily fine-tuned open-source models and on outputs that have been passed through an aggressive third-party AI humanizer. The result panel labels the highest-likelihood model family when the signature is clear, so you know which model the text most resembles.
Detect AI. Fix AI. One tool. 3 actions a day, no card required.