AI text detector · research preview
Text detection needs humility, not a percentage.
AI Screener scores pasted text with a pinned local language-model pair, shows which sentences drove the statistic, adds model-free stylometric and hidden-character evidence, and says clearly what it cannot establish. Nothing you paste leaves the deployment.
Binoculars language-model score
A pinned Qwen2.5-0.5B base and Instruct pair (Apache-2.0) score the text locally. The ratio of performer perplexity to observer/performer cross-entropy is compared with conservative gates calibrated on a versioned fixture bundle.
Sentence map
Each sentence receives its own ratio so you can see which passages pushed the statistic, instead of trusting one number for the whole text.
Stylometric profile
Sentence rhythm, vocabulary variety, stock phrasing, and chat-style formatting are summarized as soft, model-free indicators that never raise concern on their own.
Hidden-character forensics
Zero-width, bidirectional, tag, and mixed-script characters are counted and located. They can carry invisible marks or disguise content, so they are shown as separate evidence.
Watermark honesty
Vendor text watermarks need official detectors or keys that are not published or provisioned here, so the report says watermark verification is unavailable rather than guessing.
No automated detector can prove authenticity. AI Screener combines independent signals, shows the evidence, and leaves the decision with a human.
Run the free checkerQuestions to ask of the result
Limits matter as much as the signals.
Can this prove a text was written by AI?
No. A low Binoculars ratio means the text is unusually predictable to a small local model pair, which is common in assistant prose but also happens with formulaic or edited human writing. Treat it as a review lead.
Can it prove a text was written by a human?
No. Paraphrased, edited, translated, or newer-model text often passes. A result without a warning is reported as inconclusive, never as human-verified.
Should I use it for academic, hiring, legal, or fraud decisions?
Not on its own. Never act on this statistic alone; ask the author for drafts, sources, or an interview and weigh the evidence with a person.
How much text do I need?
At least about 200 characters and a few complete sentences; longer, natural prose in a covered language gives the steadiest result. Lists, code, and very short fragments are reported as insufficient.