What an Accuracy Claim Can and Can't Mean
Winston AI is a polished detector aimed at educators, publishers, and SEO teams, with a distinctive extra: OCR support that lets it scan images and handwriting, not just pasted text. Its marketing leads with a near-perfect accuracy figure.
Take that number seriously as what it is: performance against chosen test sets. The production reality for every style-based detector is harder. Formal human prose reads as "AI-like." Lightly humanized AI output reads as "human." Non-native English writers get flagged at elevated rates across the category. And crucially, even a hypothetically perfect style classifier answers the wrong question - "does this resemble AI output?" - when what a grade, a byline, or an invoice depends on is "did this person write it?"
Accuracy ≠ authorship
A detector with 99% advertised accuracy that processes a million documents still produces thousands of wrong calls - each one a real person asked to prove a negative. The error rate isn't the core problem. Answering the wrong question is.
Witnessing Instead of Guessing
ValidDraft doesn't rate finished text at all. The writer works in a tracked editor that records the composition itself - keystroke timing, pauses, backspaces, revision flow, and cursor entropy. The output is a certificate of human authorship: publicly verifiable and permanently linked to the exact text it covers.
No percentage to argue about
A certificate either exists for a text or it doesn't - checkable in seconds on the free content checker. There is no threshold to debate, no re-run that changes the answer, and no model update that silently rewrites yesterday's verdicts.