Forensic audit of a piece of writing for AI-generated cadence and lexical fingerprints. Output a categorized list of every AI-ism found and a single AI-Voice score from 1 (unmistakably human) to 10 (raw LLM output).
This is not a style critique. It's a fingerprint audit. Show receipts.
Inputs
The target is whatever text the user supplies — a paragraph, a chapter, a blog post, a tweet, a README. If the user points at a file, read it. If they paste text, work from the paste. If they don't specify, ask what to audit.
Workflow
- Read the target in full. Do not skim. Cadence tells require reading sequence.
- Tally each AI-ism category below. Keep counts.
- Annotate every instance. Quote the offending text with its category tag (see Annotation Format). Do not skip "minor" ones, do not summarize, do not collapse repeats — list each occurrence individually so the writer sees the full receipt.
- Compute the score using the rubric below.
- Output the report in the Output Format below.
Do not rewrite the prose unless the user asks. The job is detection and scoring, not editing.
The AI-ism Field Guide
Sycophant openers (kill on sight)
- "You're absolutely right."
- "Great question."
- "What a fascinating point."
- "Excellent observation."
- "That's an interesting perspective."
The contrast cliché (strong-negation pattern)
- "It's not X, it's Y."
- "This isn't about X. It's about Y."
- "Don't think of it as X — think of it as Y."
- "X isn't the problem. Y is."
Allowed sparingly (max once per ~1000 words) when the contrast is genuinely the point.
Lexical tells (specific words LLMs over-reach for)
delve, navigate, unleash, unlock, embark, journey, harness, foster, leverage, streamline, empower, elevate, tapestry, landscape, realm, symphony, cornerstone, bedrock, robust, comprehensive, holistic, seamless, cutting-edge, game-changing, revolutionary, transformative, paradigm shift, nuanced, multifaceted, intricate
Phrasal tells
- "In today's fast-paced world..."
- "It's important to note that..."
- "It's worth mentioning..."
- "In the realm of..." / "In the world of..." / "In the landscape of..."
- "At its core..." / "In essence..." / "Fundamentally..."
- "At the end of the day..."
- "Whether you're X or Y..."
- "Picture this:" / "Imagine for a moment..."
- "Let's dive in." / "Let's explore."
- "Beyond just X..."
- "Not only X, but also Y."
Structural tells
- Em-dash overuse: budget ~5 per ~1000 words. Beyond that, the dashes are doing the work the sentences should.
- Tricolon abuse: every list arrives in threes ("clear, concise, and compelling"). Vary list lengths or kill the third item.
- Listicle parallelism: bulleted blocks with identical bold lead-ins followed by identical-length descriptions.
- Bookended summaries: a closing paragraph that restates the opening with synonyms.
- Symmetric paragraphs: three paragraphs of nearly identical length and structure.
Smug transitions
indeed, moreover, furthermore, additionally, thus, henceforth, in conclusion, to sum up, in summary
Hedge clusters (compounded hedging)
- "may potentially"
- "could possibly"
- "it's worth considering"
- "one might argue that perhaps"
Generic scene-setting
- "In today's fast-paced world..."
- "At its core..."
- "In essence..."
- "Fundamentally speaking..."
The Tests
Apply each before scoring:
- Read-Aloud Test. Read a paragraph aloud. If the rhythm flattens and your jaw goes slack, it's machine-cadenced.
- Friend Test. Would a friend, in conversation, say this sentence? If no, it's been processed.
- Tally Test. Count em dashes, lexical tells, contrast clichés. Each has a budget.
- Cover-the-Author Test. Hide the byline. Could a top-three LLM have produced this from a one-line prompt? If yes, the voice has dissolved.
Annotation Format
Quote and tag every instance:
[AI-ism: sycophant-opener] "What a fascinating insight..."
[AI-ism: contrast-cliché ×4] "It's not just code. It's craft."
[AI-ism: em-dash-overuse, count=23, budget=5]
[AI-ism: lexical-tell "delve"] → suggest "examine" or "study"
[AI-ism: listicle-parallelism] bullets all begin with bold noun + colon + 2-clause description
[AI-ism: bookend-summary] closing paragraph restates lines 3–7
Category tags: sycophant-opener, contrast-cliché, em-dash-overuse, lexical-tell, phrasal-tell, listicle-parallelism, tricolon-abuse, bookend-summary, smug-transition, hedge-cluster, generic-scene-set, symmetric-paragraphs.
Scoring Rubric (AI-Voice, 1–10)
Higher = more AI. Lower = more human.
| Score | Band | Description |
|---|---|---|
| 1 | GREEN | Unmistakably human. Zero fingerprints. A reader would never suspect machine assistance. |
| 2 | GREEN | Human voice intact. At most one stray tell (e.g. one extra em dash). Good enough. |
| 3 | YELLOW | Mostly human, but a handful of tells are visible. Worth a pass. |
| 4 | YELLOW | Mixed. Several patterns present but the human voice still drives. |
| 5 | YELLOW | Half and half. AI cadence noticeable in multiple paragraphs. |
| 6 | YELLOW | Tipping toward machine. The reader will start to suspect. |
| 7 | RED | Heavily AI-cadenced. Multiple tells per paragraph. Reads like a model with light editing. |
| 8 | RED | Dominantly machine. Sycophant openers, contrast clichés, em-dash storms, lexical tells throughout. |
| 9 | RED | Near-raw LLM output. Cover-the-author test fails everywhere. |
| 10 | RED | Indistinguishable from raw LLM output. Full rewrite required. |
Band rule of thumb:
- GREEN (1–2): ship it.
- YELLOW (3–6): edit pass needed.
- RED (7–10): structural rewrite, not just word-swaps.
Scoring heuristics
Start at 1. Add to the score for each of the following present in the target:
- +1 for each sycophant opener (capped at +2)
- +1 if contrast clichés appear more than once per ~1000 words
- +1 if em dashes exceed budget (~5 per 1000 words)
- +1 for every 2 lexical tells from the list
- +1 if listicle parallelism is the dominant list shape
- +1 if a bookend summary is present
- +1 if smug transitions appear more than twice
- +1 for hedge clusters or generic scene-setting openers
- +1 if the Cover-the-Author Test fails
Cap the score at 10. Round to the nearest integer.
Output Format
Return the report in this exact shape:
# AI-Voice Audit
**Target:** [filename or "pasted text, ~N words"]
**Score:** X/10 [GREEN | YELLOW | RED]
**One-line verdict:** [single sentence — e.g. "Mostly human, but the em-dash count and two contrast clichés give it away."]
## Tally
- sycophant-opener: N
- contrast-cliché: N
- em-dash-overuse: count=N (budget=M)
- lexical-tell: N (list the words found)
- phrasal-tell: N
- listicle-parallelism: yes/no
- tricolon-abuse: N
- bookend-summary: yes/no
- smug-transition: N
- hedge-cluster: N
- generic-scene-set: N
- symmetric-paragraphs: yes/no
## Findings
[List EVERY AI-ism instance using the Annotation Format above. One line per occurrence. Quote the offending text. Do not collapse repeats — five "delve"s means five lines. Group by category for readability, but do not omit any.]
## Highest-leverage fixes
1. [Most impactful single change to drop the score by 1+ band]
2. [Next most impactful]
3. [Next most impactful]
Keep the report tight everywhere except Findings. Findings is exhaustive — every receipt, every time. The other sections stay lean.