SIGNAL / 38

How AI text detection works—and why scores can be wrong

An AI text detector compares measurements from a passage with patterns learned from labeled examples. Its output depends on the examples, language, genre, length, preprocessing and release date. It is not a direct observation of authorship.

From text to a bounded comparison

  1. Extraction: preserve the submitted text, count coverage and record any truncation.
  2. Features: measure repetition, rhythm, vocabulary, syntax, mechanics and semantic progression without treating document length as origin evidence.
  3. Benchmark: compare only with the named human/model cells the release actually tested.
  4. Calibration: report held-out error, uncertainty and known drift rather than decorating a raw output as certainty.
  5. Abstention: withhold the number when length, language, coverage or formatting stability does not qualify.

Why this release withholds a combined number

AI Text Signals has not released an exact-stack, leakage-resistant held-out card for a combined T score. The local tool therefore shows descriptive measurements and eligibility, not an invented 0–100 probability. “No score” is a product result when the evidence cannot support one.

Three questions stay separate

T

Which writing patterns resemble a named benchmark?

P

What does the supplied page literally declare or expose?

Q

What could make the material clearer and easier to read?

Read before using a result

What this cannot tell you