When the Slop Score is wrong

Short text (under ~150 words)

Every engine here needs a few paragraphs to work. The lexical layer scores per 500 words. The structural signals (sentence variance, paragraph uniformity, tricolon density) need a sample to measure variance against. The epistemic engine looks for stance and specificity that don't fit in a tweet. A 60-word product description will score unpredictably no matter who wrote it. Paste a paragraph; ideally three.

Non-native English writers

Stanford's Liang et al. 2023 showed that seven leading AI detectors flagged 61% of TOEFL essays written entirely by non-native English speakers as AI. The mechanism: detectors learn that limited lexical variety = AI, but limited lexical variety is also a hallmark of L2 English. Our score is decomposable, which partially defuses this. If your lexical layer fires but structural and epistemic stay clean, you're writing in a formal or L2 register — call it that, not slop. Read the per-engine bars, not just the headline number.

Highly technical or domain-specialist writing

Patent prose, legal contracts, scientific abstracts, and engineering documentation all register as mid-to-high slop because they're supposed to: a contract that takes a surprising stance is a malpractice claim, and a scientific abstract that uses lived detail is doing the wrong job. The score reflects stylistic register accurately here, but the verdict ("sloppy") is contextually wrong. Don't use deslopify to score documents whose genre requires the voiceless register.

Paraphrased / human-edited AI

A determined writer who edits AI output carefully — varying sentence length, adding a named example or two — can defeat any pattern-based tool, ours included. The academic result here is Krishna et al. 2023 (NeurIPS): their DIPPER paraphraser dropped DetectGPT's true-positive rate from 70.3% to 4.6%. Industry detectors have published similar collapses on text rewritten through common paraphrasers. Slop survives casual editing because the structural patterns are deep; it does not survive a careful rewrite by someone who knows what they're looking for. Which is fine. The rewriter is doing exactly the work the score asked for.

Code blocks, lists of facts, and reference material

A page that's 80% bulleted feature list and 20% prose will score high on structural (formatting density saturates the signal) without that meaning anything about the prose. We surface this in the per-engine breakdown. If structural dominates and the other two stay low, you're reading a formatting artifact — the prose itself is fine.

The score is not a quality verdict

A 2023 study by Zhang & Gosline (MIT Sloan / Berkeley Haas) ran 1,203 participants through blind ratings of short persuasive content. AI-generated marketing copy scored higher than human-expert copy on both satisfaction and willingness-to-pay. So "low slop" is not "good writing" and "high slop" is not "bad writing." High slop means the text sounds interchangeable. For a product description, interchangeable is fine. For your personal essay, interchangeable is the whole problem.

Languages beyond English, German, and Spanish

The diagnostic engines run in those three languages today. Each ships with an authored dictionary, structural regex patterns, and a Sonnet judge prompt tuned in-language. Slop is a language-native register; a generic translation layer over a single English engine would produce numbers that don't mean anything. When the language of pasted text doesn't match the site locale the engines follow the text, and a banner offers to switch the UI.

Paste a Polish or Mandarin paragraph and the engines still run, but treat the score as noise. Adding a locale takes an authored dictionary, fresh regex patterns, a calibrated Sonnet prompt, and six pinned anchor texts run through a live tuning pass. Three locales we trust is the ceiling we ship at. Ten that look multilingual but aren't is a worse product.

Models we haven't tested against

Our calibration set was assembled against output from contemporary frontier models (Claude, GPT, Gemini). New models trained explicitly to evade detection are a moving target. We re-score the anchor set after every weights change and publish the results in our calibration page; the moment a new model class breaks the bands, the engines get re-worked. A marketing patch won't do it.

What we'd like to publish next

• Cross-model robustness numbers (same prompt, four model families, do the scores cluster?).
• Paraphrase-resilience benchmark (raw AI, then through DIPPER / QuillBot, score delta).
• A non-native-English subset with controlled register samples.

If you hit a case where the score feels wrong, tell us. There's no "report this score" flow yet; the waitlist email is the channel until there is.