A few weeks ago I came across this post on Marginalia showing that new Hacker News accounts are ten times more likely to use em dashes than older ones. Not as proof of anything - just as a signal. A stylistic fingerprint that's shifted measurably since large language models became mainstream.
That felt like a project.
What aidar does
aidar scans URLs and measures stylistic patterns that have become associated with AI-era writing. Not to declare something "AI-generated" - that framing is a trap - but to build a comparable, queryable dataset of how writing style is drifting across the web.
The patterns it looks for fall into two tiers.
Structural and tonal tropes - the patterns that are uniquely, embarrassingly AI. Sourced largely from tropes.fyi:
- Negative parallelism - "It's not X - it's Y." The single most identified AI writing tell. One in a piece is fine; ten is an insult to the reader
- Em dash addiction - AI uses 10–20+ per 1000 words; human prose uses 2–3
- Bold-first bullets - every list item opens with a bolded phrase. Almost nobody writes this way by hand
- Rhetorical Q&A - "The result? Devastating." The model poses a question nobody asked, then answers it
- Tricolon abuse - "Products impress people; platforms empower them; participants thrive." Three-part parallel structures stacked back-to-back
- "Here's the kicker" - false suspense transitions before unremarkable observations
- "Let's break this down" - the pedagogical voice, even for expert audiences
- Signposted conclusions - "In conclusion..." Competent writing doesn't announce itself
- Grandiose stakes inflation - a blog post about API pricing becomes a meditation on the fate of civilization
- Vague attributions - "experts argue", "studies suggest", "many observers" - without naming anyone
- Listicle in a trench coat - "The first wall is... The second wall is..." - a numbered list disguised as prose
- "Despite its challenges" - the rigid formula: acknowledge → dismiss → optimism
- AI section headers - "The Takeaway", "Why This Matters", "Key Takeaways", "Moving Forward", "What's Next". Detected in raw HTML heading tags so it fires even when the page is rendered by a JavaScript-heavy CMS
- Second-person direct address - "We've all been there", "You might be wondering", "Picture this", "Let me walk you through", and the entire "Here's how it actually went" family - the reveal formula pasted into every AI-assisted dev.to and LinkedIn post
- "And honestly?" - the dramatic-pause reveal. "And honestly? I lost all motivation." Flagged in the HN discussion on tropes.fyi as a Claude-specific emerging tell
Word and phrase signals - vocabulary that has statistically shifted since LLMs:
- Em dash frequency - 0–2 per 1000 words in human prose; 6–12+ in AI output
- Hedging phrases - "it's worth noting", "to be honest", "certainly", "genuinely"
- "Delve" and friends - delve, utilize, leverage, robust, streamline, harness, seamlessly, cutting-edge, actionable insights
- "Tapestry" and "landscape" - ornate nouns where simpler words would do: tapestry, paradigm shift, ecosystem, cornerstone, at the intersection of
- Magic adverbs - quietly, deeply, fundamentally, remarkably - used to make mundane descriptions feel weighty
- The "serves as" dodge - replacing "is" with "serves as a reminder", "stands as a testament", "represents a fundamental"
- Bullet point density - AI defaults to lists regardless of whether content warrants it
- Sentence burstiness - humans mix short and long sentences; AI outputs uniform medium-length ones
- Question avoidance - AI almost never asks questions; it answers them
- Type-token ratio - a lightweight proxy for lexical diversity
- Transition overload - "furthermore", "moreover", "consequently" in clusters
Each pattern produces a normalized 0–1 score. These aggregate into a per-category breakdown and a final Stylistic Index from 0–100.
Scanning my own site
I ran it against all 96 pages on this blog:
aidar discover carteakey.dev -o urls.txt
aidar scan --batch urls.txt
Results on carteakey.dev (current):
- Max score: 19/100 (the changelog - 8,500 words of dense list structure)
- All 96 pages labeled LIKELY HUMAN
- The tropes detectors - negative parallelism, bold-first bullets, tricolon abuse - scored 0.00 across every page
Then I scanned an older build that had more AI-assisted posts. Top result: 36/100 LIKELY AI on /blog/tech/when-computers-finally-speak-human/. That page fired four trope detectors simultaneously: em-dash addiction (10.6 per 1000 words), bold-first bullets (9.3/1000), tricolon abuse, and 13 writing trope matches. That's real. I wrote that post with heavy AI assistance and it shows.
The highest-scoring page on the current site is the changelog - which scores high purely on structural signals (bullet density, semicolons) rather than AI vocabulary or tropes. That's a meaningful distinction the tool can make.
How the scoring works
Patterns are YAML files in a patterns/ directory. Adding a new signal means writing a config file - no Python required. Each pattern has:
- Calibrated thresholds (
threshold_low,threshold_high) - the range over which the score goes from 0 to 1 - A version number - bump it when you recalibrate, so old scans get flagged as stale
- A detection type -
frequency,regex,html_regex,structural, orlinguistic
The html_regex type is worth noting: it runs patterns against the raw HTML source rather than extracted text. This is how bold-first bullet detection works - trafilatura strips <strong> tags from the extracted text, but <li><strong>Keyword</strong> is detectable in the HTML itself.
The aggregate score is a weighted sum of category scores. The tropes category carries the highest weight (0.40) since structural writing habits are the strongest and least gaming-prone signal.
One comment from the HN thread on tropes.fyi nailed the calibration philosophy: "Those tropes are things humans do too. But like once or twice in an article. Not every single freaking paragraph." That's the whole point of threshold-based frequency scoring rather than binary detection.
The irony
This tool was built by Claude. The codebase has em dashes in the comments. If you scanned the commit messages, the hedging detector would fire.
I find this funny.
What's coming
The leaderboard site (aidar.lol) will show stylistic drift over time. Once there's enough historical data - using the article publish date that trafilatura extracts from each page - you can plot a domain's stylistic drift as a time series.
The TODO has things like:
- Dead Metaphor, One-Point Dilution, and Content Duplication detectors (the three that need NLP/semantic analysis rather than regex)
- Real perplexity scoring via GPT-2
- Wayback Machine integration for pre-AI baselines
- An embed badge for your site
If you want to contribute patterns or calibration data, the GitHub repo is open.
Running it yourself
pip install git+https://github.com/carteakey/aidar.git
# Analyze a single page
aidar analyze https://yourblog.com/some-post
# Or analyze raw text directly
aidar analyze --text "paste your article here"
# Discover and bulk scan a whole site
aidar discover yourblog.com -o urls.txt
aidar scan --batch urls.txt --min-words 50The scan takes a few seconds per page. Results go into aidar.db (SQLite).
References
- "Wikipedia: Signs of AI writing" - 2023–present
- "Tropes - AI Writing Pattern Directory", Ossama Chaib - 2026
- New HN accounts ten times more likely to use em-dashes, Marginalia
- LLM writing style: empirical stylometric research (PNAS 2025) - found GPT-4o's second-most-overused word is "tapestry"; present participles alone are a near-sufficient stylometric classifier
- LLM writing style resources, Alex Reinhart - curated research notebook on this topic
Yes, this post was written with AI assistance. The em dashes are real. I checked.