About the slop detector
This measures style, not authorship. Humans write slop and models can avoid it. Do not use it to make decisions about people (grading, hiring, misconduct).
Human review: none. AI-generated (Claude Code) under the direction of Nicolas Jonason. No person has yet read or checked this code, its results or its documentation.
What it does
Paste English text to see how many of the stylistic tells common in AI-generated text it shows, and every phrase that moved the score. The score comes from 87 hand-written features (regexes, word lists, counts) weighted by a small logistic regression, fitted on 2026 models and pre-2022 human text. The rule set lists every feature with its rationale, source and weight.
Where the rules come from. Of the 87 rules, 34 come from research papers only, 23 from lists and guides written by people only (Wikipedia’s guide, editors, tool makers), and 18 from both. The other 12 have no reference: agents found them in the training texts and kept those that held on a second half of the data. Each rule’s details show its references, and the notes beside the text name the first one.
The page runs a JavaScript port of the Python engine in your browser. The port was checked against Python and gives the same values, contributions, scores and highlights on 224 test texts.
Reading the page
- Highlights
- Each highlighted word or phrase is a hit of a counting rule. Green (dotted underline) points towards human style, red (solid underline) towards AI style. Each hit adds weight × 1000 / (training SD × words) to the score, and the stronger the colour, the more it adds. Where hits overlap, the stronger one is drawn. Rules whose hits add less than 0.1 in total are not highlighted unless you ask for them in the verdict details, so rules on common words and punctuation do not paint the whole text.
- Font view
- With Font view on (the default), the letter shapes follow the evidence: loose and irregular letters lean towards human style, square and uniform ones towards AI style. The whole text takes the shape of its P(AI-style) (halfway between the two under 100 words), and highlighted words are pushed further towards their own direction, more and in heavier strokes the stronger they are. It is a display only and does not change the score. The letters come from the project’s procedural monospace font generator; the Font view button returns to plain text.
- Notes and lines
- Each rule with highlighted hits in view gets one note in the margin, with a line to each of its hits. The note gives the rule's name, its contribution to the score (positive towards AI style, negative towards human style) and the words it fired on. On narrow screens the notes are a row of chips under the text box.
- Rule details
- Tap a note, a chip or a highlight for the rule: what it looks for and why, its value in this text, its weight and its contribution, the hits it found, its source and the direction observed on the training data. A counting rule's contribution splits into what its hits add and what a text with no hits would get: the average training text has some hits, so having none also counts.
- Verdict
- The tag on top of the box gives the band and P(AI-style). Tap it for the logit, the word and sentence counts, every rule with its contribution (including the statistics of the whole text, which have no highlights), and the sum: the contributions plus the bias add up to the logit.
About the score
- The score is a logit: bias + the sum over all rules of weight × (value − training mean) / training SD. P(AI-style) = 1 / (1 + e−logit).
- Bands come from out-of-fold scores of the training texts: “little slop” below 0.26 (80% of human and 1% of AI training texts), “some slop” up to 0.78, “a lot of slop” from 0.78 (80% of AI and 2% of human training texts). The lower threshold is the 80th percentile of the human texts' scores, the upper one the 20th percentile of the AI texts' scores.
- Short texts: under 100 words the evidence is shown but the score is withheld. The model was trained on texts of 100 to 1,453 words, its accuracy drops at 100 words (dev AUC .92 vs .98 for whole texts), and most features are rates per 1,000 words, which a few words swing hard.
- What it is not: no language model and no proof of authorship. Tells change as models change; this rule set was fitted on 2026 models.
- Accuracy: dev AUC .977 (whole texts) against .990 for a bag-of-words model on the same data; that gap is the price of readable rules. These numbers and the student-essay rates below are from the engine v1 report and are updated with each engine report.
Student essays
It wrongly flags about 1 in 7 student essays: 14% of essays by students not classed as English learners and 5–6% of essays by English learners (PERSUADE and ELLIPSE, grades 6–12, all written by people), at the model's default threshold of P(AI-style) ≥ 0.48. Formal school essays read as AI style to it.
Example texts
The Examples menu has two texts from the training data and two fresh model outputs on the same topic as the human example. The fresh ones are not from the training data, and nobody has checked their facts.
- The list of examples needs JavaScript.
Privacy
Everything runs in your browser. The page makes no network requests, and the text you type or paste never leaves the page.