Is it a person, or is it slop?
SlopDetector scans text for the documented AI-writing fingerprint — buzzwords, em dashes, flat rhythm, stock phrases — then runs a semantic judge over the whole piece. It measures style, not authorship. No AI content is automatically wrong; this is the pre-publish sniff test.
–
Signal breakdown (points of each max, 0–100 total)
Matches found
Specific phrases and patterns, with context.
Semantic judge (Gemini, schema-constrained)
How to fix it
How it works
1 · Deterministic fingerprint (free, instant)
12 signals measured directly in the text: AI buzzword density, stock phrase clichés, em/en dash density, "actually" overuse, sentence-rhythm burstiness, vocabulary range, trigram repetition, rhythm runs, chat-style openers, bold-bullet formulas, and typography tells. Every point is explained with the exact phrase and context.
2 · Semantic judge (Gemini 3.6 Flash)
A schema-constrained Gemini call reads the whole text and scores it on semantic tells: uniform voice, no first-hand detail, hedge-stacking, empty summary closers. Judge and fingerprint are blended with the judge weighted lower when it reports low confidence.
What the score means
The score is ordinal, not calibrated. It ranks how much a text looks like unedited model output. Raw AI text scores high, human-edited AI text lands mid-range, and plain human writing scores low. False positives happen: formal, technical, or academic writing can trip rhythm and vocabulary signals.
What it doesn't mean
Google does not penalize AI authorship — it penalizes scaled, low-value content (spam policy, "scaled content abuse"). The real reasons to cut slop: readers and platforms are actively soured on it (LinkedIn's "AI slop" report button, YouTube's detection misfires), and AI answer engines cite specifics, not prose texture.
Research backing
- Liang et al., Nature Human Behaviour 2025 — "realm", "intricate", "showcasing", "pivotal" flat for a decade, then surged post-ChatGPT.
- Kobak et al., Science Advances 2025 — excess-vocabulary in 15M abstracts; "delve" +~1500% 2022–2024.
- Guo et al. 2023 — em dashes in ~73% of GPT-4 long-form samples vs ~12% of human writing.
- Aletheia (julienmiquel) — burstiness CV: human > 0.4, AI < 0.2; LLM judge adds the semantic layer statistical counters miss.
- antydizajn/ai-slop-detect, hwajongpark/slop-gate, fbuchinger/smellcheck — open-source tell packs (em-dash, openers, vocabulary).
- Ahrefs 2025 (600k pages): AI-content percentage vs ranking correlation ≈ 0.011 — Google ranks value, not authorship.
- Full source list: RESEARCH.md