AI Detection Alternatives: What Actually Works Instead of a Score

Why the detector question can't be answered reliably

AI-text detectors are trained to estimate the probability that a finished document was written by a language model. That's a statistical guess, not a measurement, and the guess has real, documented failure modes. A 2023 NeurIPS study found that commercial GPT detectors misclassify non-native English writers' essays as AI-generated over half the time, in one measurement as high as 61%, because non-native writing patterns statistically resemble the patterns detectors associate with AI output. A separate 2023 study testing detection tools broadly found none reliable enough to support a high-stakes decision like a grade or an academic-integrity case.

Neither problem is a bug that gets fixed in the next model update. Both come from the basic shape of the question: judging a finished artifact instead of observing the process that produced it.

What to use instead

A process record, not a probability score. Instead of asking "does this text look AI-written," ask "what actually happened while this was being written," and keep an answer to that question that can be checked, not just asserted.

That's what TWFF (Tracked Writing File Format) and Colophon, the Chrome extension that exports it, are built to do: a chronological, tamper-evident log of every edit, paste, and AI interaction as a document comes together. Not a score. Not a classification. A record the writer keeps and can point to, verifiable against the document itself rather than argued with.

The practical difference shows up exactly where detectors fail. A non-native English speaker's genuinely human writing doesn't get flagged, because nothing about a process record depends on how the finished sentences read. A student who used AI responsibly, reviewed it, changed it, rejected parts of it, has a record that shows exactly that, instead of a single opaque percentage that can't distinguish "wrote it with AI help and reviewed every line" from "pasted it in unread."

Trying it without switching everything at once

A teacher can try this on one real assignment, not a whole course's grading policy, free, with no district procurement step: start a class in Colophon. The format itself is open and free to implement anywhere: the TWFF specification.