Skip to content

Limitations & authorship

slopscore is not an AI detector

It flags writing patterns, not provenance. A high score means the text is dense with formulaic, generic, low-specificity, over-polished patterns, which occur in low-effort AI output and in plenty of human writing (marketing copy, SEO, fan fiction). It scores text, not authors, and must never be used to accuse anyone.

Known limitations

  • Non-native English / plain writing is over-flagged by pattern detectors (one study found ~61% false positives on non-native essays). slopscore mitigates with a corroboration gate, a negative human-writing signal, and abstention, but residual risk remains.
  • Short text (< ~300 words) is unreliable; under ~100 words the label abstains.
  • Genre matters: marketing and travel prose naturally resemble slop. Use --profile.
  • Light paraphrasing evades pattern matching, as it does all detectors.
  • The rules age with the models. Most rules were written from 2023-25 model output, and the newest models drop many of those habits (em dashes, "delve", "In summary"). On Wikipedia articles tagged as suspected AI text from July 2026 on, the rules do no better than chance. See MODEL_CARD.md for the numbers.
  • A low score does not mean a document is ready. slopscore scores phrases, sentences, and Markdown block shapes. It does not judge document structure: whether a thesis is restated in every section, whether the evidence comes before or after the framing, or whether a headline figure is ever derived. A reader found a 5,000-word design document inflated and repetitive after it scored under 10. Patterns slopscore does not score lists what to check by hand.

See the project MODEL_CARD.md for measured numbers and the real-corpus (MAGE) experiment.

Authorship signal (optional, separate, caveated)

slopscore ships no authorship detector. Every candidate (Binoculars, Fast-DetectGPT, GLTR) collapses on paraphrase, is biased against non-native English, or is academic-only; watermark detection covers almost no models. What it offers instead is a pluggable adapter (slopscore.detectors.AuthorshipDetector) behind the [detectors] extra: if you bring your own detector, its result is reported in a separate field with a mandatory caveat and is never folded into the SlopScore. slopscore-lint scan --detector reference wires a no-op example to show the separation. This is for curiosity, not accusation.