# AI Slop Detection: Two-Tier Reflex Methodology A phrase blocklist catches the obvious tells. Most AI-generated prose passes that filter and still reads like AI. The structural tics, the rhythmic flatness, the "everything is a three-clause sentence" cadence: those survive the first pass. This reference defines a **two-tier reflex check** for editorial review. Run both passes before declaring a draft human-natural. Adapted from the impeccable plugin's UI slop methodology (Paul Bakaus, Apache 2.0). --- ## Why two tiers LLMs converge on a small set of safe patterns. The first thing the model reaches for is the **first-order reflex**: the genre-obvious tell. Replace it and the model reaches for the **second-order reflex**, the next-most-trained pattern that survives anti-AI guidance. Most AI-detection passes only check the first-order pattern. The result is "anti-AI" rewrites that still read like AI because the structural pass was never run. **Note on terminology**: this file uses **"first-order"** and **"second-order"** for the two detection passes. Elsewhere in the project, "Tier 1 / Tier 2 / Tier 3" refers to *source authority* (Google Search Central = Tier 1, Ahrefs = Tier 2, reputable industry sources = Tier 3). The two namespaces are intentionally kept separate; do not call the second-order detection pass "Tier 2." Examples of the same idea across both tiers: | Topic | First-order tell | Second-order tell that survives | |---|---|---| | SEO blog | "In today's digital landscape..." | Every H2 ends with a rhetorical question | | SaaS post | "Game-changer," "Revolutionize" | Three-clause sentence rhythm, "While X, also Y" framings | | How-to guide | "Dive into," "Unlock the potential" | Every step opens with an imperative verb identical in length | | Listicle | "Cutting-edge," numbered fluff | Every item is ~80 words, identical structure | | Thought leadership | "Comprehensive guide," "harness the power" | Hedge stack: "often," "typically," "may" within 20 words | The point: **a draft can score zero on a phrase blocklist and still be obviously AI.** --- ## First-order reflex (phrase + lexical) This is what the existing AI-detection in `blog-analyze` and `blog-rewrite` already covers. Documented here for completeness. **Trigger phrases** (full list in `agents/blog-reviewer.md` and `scripts/analyze_blog.py`): - "In today's digital landscape" / "In the ever-evolving" - "It's important to note" / "It is worth mentioning" - "Dive into" / "deep dive" / "delve" - "Game-changer" / "Revolutionize" / "transformative" - "Cutting-edge" / "state-of-the-art" / "robust" - "Harness the power" / "Unlock the potential" - "Leverage" (as a verb, non-financial) - "Seamlessly" / "seamless integration" - "Tapestry" / "rich tapestry" / "multifaceted" - "Comprehensive guide" (in body text) - "Furthermore" / "Moreover" (transition overload) - Em dashes used as a stylistic flourish (any density) **Lexical signals**: - AI trigger-word density > 5 per 1,000 words - Type-Token Ratio (TTR) below 0.40 on long-form - Burstiness (sentence-length standard deviation / mean) below 0.3 **Outcome of the first-order pass**: a "phrase-clean" draft. Necessary, not sufficient. --- ## Second-order reflex (structural + rhythmic) These are the patterns LLMs default to **after** the obvious vocabulary is replaced. They are structural and rhythmic, so a vocabulary swap doesn't fix them. Run this pass on drafts that already passed the first-order check. ### Structural tics to flag 1. **Question-cadence H2s.** Every section heading is phrased as a question. Real long-form mixes question, statement, and noun-phrase headings. Flag if > 70% of H2 headings end with a question mark. 2. **The Here opener.** A paragraph opens with the word "Here" ("Here's why...", "Here are five..."). Once is fine. Three or more in a 1,500-word post is an AI fingerprint. 3. **Three-clause sentence rhythm.** Most sentences in a paragraph follow the structure `[clause], [clause], [clause].` The cadence is metronomic. Flag if > 50% of sentences in any 200-word window match this shape. 4. **False-balance framing.** Repeated use of "While X, also Y" or "On one hand X, on the other Y" without a real contrast. The model uses it to feel even-handed but it adds no information. Flag if it appears more than twice per 1,000 words. 5. **Hedge stacking.** Three or more hedges in a 20-word span ("It may often be the case that..."). Flag any 20-word window with > 2 of: may, might, often, typically, generally, usually, tend to, perhaps, somewhat, likely. 6. **Symmetric list bloat.** Every item in a numbered or bulleted list is the same length within +/- 10 words and follows the same syntactic structure. Real lists vary; some items need one line, others need a paragraph. Flag if list-item length standard deviation < 5 words. 7. **The wrap-up question.** Section ends with "What does this mean for [audience]?" or "Why does this matter?" Once per post is rhetorical; three or more is filler. 8. **Capsule transitions.** Each H2 opener begins with a single-word transition ("First..." "Next..." "Additionally..." "Crucially..."). Real prose buries transitions inside sentences. Flag if > 50% of H2 openers start with a transition word. 9. **The "key insight" tell.** The phrase "The key insight is..." or "What's important here is..." appears as a sentence-opener. This is the model telegraphing that it's about to summarize. Cut and let the sentence stand. 10. **Listicle introduction bloat.** Before the actual list, three or more paragraphs of "context." Real listicles get to the list. Flag if > 250 words of pre-list intro. ### Rhythmic signals to compute - **Sentence-length flatness within paragraphs.** Compute SD of sentence length per paragraph; flag any paragraph with internal SD < 4. - **Opening-word repetition.** Count first-word frequencies across all sentences. Flag if the top three first-words account for > 25% of all sentence openings. - **Paragraph-shape flatness.** Compute SD of paragraph word counts across the post; flag if < 25 (real long-form varies dramatically). --- ## How to run the two-tier check For `blog-rewrite` and `blog-reviewer`: 1. Run the first-order pass first (phrase + lexical). If it fails, fix and re-run before moving on. 2. Once the first-order pass is clean, run the second-order pass. Report each second-order pattern with line numbers and an example. 3. Do not declare "AI-detection passed" unless both passes are clean. For `blog-write` (initial drafting): - First-order is enforced at generation time via the persona's anti-phrase list. - Second-order is checked once on the full draft before delivery. --- ## Output format When reporting findings, use: ``` ## AI Slop Detection Report ### First-order (Phrase + Lexical) - Trigger phrases: [N found] -> [list with line numbers] - AI trigger words: [N/1K words], [pass/fail at ≤5] - TTR: [score], [pass/fail at ≥0.40] - Burstiness: [score], [pass/fail at ≥0.3] ### Second-order (Structural + Rhythmic) - Question-cadence H2s: [X%], [pass/fail at ≤70%] - "Here" openers: [N], [pass/fail at ≤2] - Three-clause rhythm: [X%], [pass/fail at ≤50%] - False-balance framings: [N/1K words], [pass/fail at ≤2] - Hedge stacking: [N windows], [pass/fail at 0] - Symmetric list bloat: [N lists], [pass/fail at 0] - Wrap-up questions: [N], [pass/fail at ≤2] - Capsule transitions on H2s: [X%], [pass/fail at ≤50%] - "Key insight" sentence openers: [N], [pass/fail at 0] - Listicle intro bloat: [pre-list words], [pass/fail at ≤250] - Sentence-length flat paragraphs: [N], [pass/fail at 0] - Opening-word repetition: [top-3 share], [pass/fail at ≤25%] - Paragraph-shape SD: [value], [pass/fail at ≥25] ### Verdict First-order: [PASS / FAIL] Second-order: [PASS / FAIL] Overall: [PASS only if both passes clean] ``` --- ## Why this matters for ranking + AI citations - **Editorial quality risk**: content that lacks experience and original perspective is easier to classify as interchangeable consensus content. Second-order patterns are a warning sign, even when no specific Google update is cited. - **AI citations**: ChatGPT and Perplexity reward citable, distinctive passages. Second-order tics produce interchangeable prose that no AI surface has reason to prefer over the source it was trained on. The two-tier check is the editorial parallel to impeccable's "design slop" methodology: vocabulary-clean is necessary but not sufficient; structural distinctiveness is what separates citeable content from indexable filler. --- ## Attribution The two-tier first-order / second-order reflex methodology is adapted from the impeccable plugin v3.1.1 (Paul Bakaus, Apache 2.0, https://github.com/pbakaus/impeccable). The original applies it to UI design cliches ("observability -> dark blue"). This reference adapts the same mental model to prose.