AI Content Detector.
Paste any text. We scan it with the same deterministic engine behind Typographer and return a transparent machine-pattern index — every point traced to a named signal: mechanical tells, overused phrasing, sentence uniformity, repetition, and hidden tampering characters. A score you can actually read, not a black-box percentage.
How it works
Most AI detectors hand back a confidence percentage from a model you cannot inspect — and they are wrong often enough to get real people falsely accused, English-language learners most of all. We do the opposite. Every point of the score is traced to a concrete, checkable signal and shown to you, sentence by sentence. No black box, no guessing.
The mechanical vocabulary of machine writing — empty buzzwords, throat-clearing, signposting, fake-profound significance, sycophancy — plus a database of roughly 3,000 phrases LLMs overuse.
The shapes that survive a thesaurus pass: rule-of-three triads, concession-and-hedge scaffolding, and formulaic “there are three reasons…” set-ups.
Repeated sentence openers, echoed phrasing, and thin information density — few names, numbers or sources. Flat sentence rhythm counts too, but only lightly.
The residue “humanizer” rewriters leave behind: hidden zero-width characters, look-alike (homoglyph) letters, and unnaturally forced sentence variation.
We report a Machine-Pattern Index — how machine-like the writing is — and, on a separate axis, a Confidence score for how much to trust that read, weighed from how much text we have, how far the score sits from the decision line, and how many independent signals agree. A striking index on a forty-word fragment is reported as low-confidence, on purpose.
We don’t score vocabulary richness, and we keep sentence rhythm a minor signal — those are the patterns that falsely flag nonnative English writers, the documented bias that has gotten detectors banned from classrooms. We don’t measure perplexity against a hidden reference model, which mislabels famous, much-quoted human writing as machine. Instead we do the opposite: organic human markers — contractions, sentence fragments, asides, the small imperfections of real writing — earn a credit that pulls the score toward human, and when they conflict with a machine-leaning score we lower our confidence rather than press the accusation. And we never return a number you can’t interrogate.
That is the whole 701am stance. We don’t help you fool a detector, and we don’t pretend to be a magic one. We make the machine habits visible so you can take them out — and put your own voice back in.
How we mean this to be used
We don’t build this to catch cheaters, and you shouldn’t use it to accuse anyone. AI detection is probabilistic, never proof — people write in “machine” patterns too, and a score is a starting point, not a verdict. We built it as a resource for creating better human-powered and human-influenced writing: a mirror that shows where your draft reads mechanical so you can put your own voice, judgment, and specifics back in. Use it to improve writing, not to police it.