The AI Vernacular Index: a live reference for the words, phrases, and habits that mark text as model-written. Check your own draft, see what each model does and how often, what's measurably rising or falling, and download an instruction file that steers any model away from the worst of it.
Paste anything you're about to send. The highlights show every recognizable AI tell, with a plainer fix for each. Your text never leaves this page — the check runs entirely in your browser, and only anonymous counts of which tells fired are recorded, never the text.
Left: phrases measured against human writing by GPTZero across millions of documents. Right: em-dash density by model, against a measured human baseline. Hover any bar for the source.
How many times more often the phrase appears in AI text · GPTZero, Oct 2024
Model defaults vs. human essay baseline · SlopDetector, 2026
Most "AI-isms" are shared across every model. But each one has habits distinct enough that a classifier can name the author. Confidence labels: measured = quantitative study · documented = widely reported, official acknowledgment, or dedicated tracking · anecdotal = community consensus without hard counts.
The same content written twice. Hover the highlights to see which tell each one is.
In today's fast-paced digital landscape, WidgetFlow isn't just another project tool — it's a paradigm shift. Our platform seamlessly leverages cutting-edge AI to streamline your workflow, empowering teams to unlock their full potential. From scrappy startups to global enterprises, WidgetFlow boasts a robust suite of features designed to foster collaboration, enhance productivity, and elevate performance. It's important to note that every feature was meticulously crafted with the user in mind. The result? A truly transformative experience that stands as a testament to what's possible. In conclusion, WidgetFlow represents more than a tool — it's a comprehensive ecosystem for the modern workplace.
WidgetFlow is a project tracker for teams that have outgrown spreadsheets. It schedules work, flags conflicts before they block anyone, and writes the weekly status report for you. Setup takes about ten minutes.
The scheduling engine is the part we're proudest of. It plans from your team's actual pace over the last six sprints, not from estimates, so the plan stops being fiction by Wednesday.
Free for 30 days. If it doesn't save you an hour in the first week, cancel from the billing page in two clicks.
The Q2 results underscore the pivotal role of our multifaceted growth strategy. Revenue saw significant improvements, highlighting the team's meticulous execution and unwavering commitment to excellence. While challenges remain, the overall trajectory serves as a testament to our resilience. Moving forward, we will continue to leverage synergies across the organization — ensuring sustainable, long-term value creation for all stakeholders.
Q2 revenue came in at $4.2M, up 18% on Q1. Most of the gain came from the enterprise tier, which closed 11 accounts against a target of 8.
Churn is the sore spot. We lost two mid-market customers over pricing and don't yet have a fix; Dana's team owes a proposal by August 22.
The Q3 plan bets on the enterprise pipeline holding. If it slips, the events budget gets cut first.
What changed in each rewrite: specific numbers replace abstractions, sentence lengths vary, no triplets, one em dash total across both, and each ends on new information instead of a summary.
Direction reflects the newest available data: several tells are now falling at the source (labs tune them out once they become memes) while still rising in human speech and the published record.
| Tell | Evidence | Direction |
|---|---|---|
| "not just X, but Y" | ~6% of sampled ChatGPT messages; called the most prominent sign by Wikipedia's editor guide | ▲ rising — the em dash's successor |
| em dash | GPT-4.1 at 10.6/1,000 words vs human 3.2; GPT-5.4 down to 1.4 | ▼ falling at the source |
| delve | 28× excess in 2024 abstracts; 14× in indexed papers | ▼ fading in new output |
| delve (in human speech) | Significant post-2022 rise in podcasts and academic talks (Max Planck; FSU) | ▲ rising — leaked into people |
| underscore | ~30% of 2025 PMC papers use it, vs ~3% in 2022 | ▲ still climbing in print |
| meticulous | 34.7× in AI-era peer reviews; still growing in full-text corpora | ▲ rising |
| tapestry | On every blacklist; heavily mocked | ▼ tuned out of new models |
| "today's fast-paced world" | 107× human rate (GPTZero) | — stable |
| markdown bold/bullets in prose | Near-zero in newest models when told to write prose | ▼ falling |
| rule of three · false ranges · vague attribution | Codified by Wikipedia's guide; structural, so harder to tune out | ▲ the 2026 frontier |
| uniform sentence cadence | Shared structural fingerprint across ~82% of AI posts, any model | — stable |
| sycophantic openers | SycEval: 58% average sycophancy across models; "You're absolutely right" tracked daily | — stable |
Two files, same rules, different packaging. Both are built from the evidence above and prioritize the loudest tells first.
A SKILL.md for Claude's skill system. Claude applies it automatically whenever it drafts prose for you.
write-like-a-person/ folder under ~/.claude/skills/, or ask Claude to save it as a skill.Plain text that fits in a custom-instructions box. Works in ChatGPT, Gemini, Claude, or pasted at the top of any prompt.
Three caveats worth keeping in view. First, attribution is hard: most "Claude-isms" and "Gemini-isms" people complain about are generic LLM habits attributed to whichever model the complainer uses most; this page marks per-model claims with confidence labels for that reason. Second, the tells are a moving target: labs tune out each meme (delve, em dashes, markdown walls) roughly a year after it peaks, so the durable signals are structural — cadence, negative parallelism, triplets — not word lists. Third, avoiding the vernacular makes text less annoying to readers, but it does not defeat detectors: stylometric classifiers still identify model output at 90%+ accuracy after paraphrasing. Nothing here is for passing text off where AI disclosure is required.
Kobak et al., "Delving into LLM-assisted writing in biomedical publications" (Science Advances 2025) — excess-vocabulary method; 13.5% figure; delve 28×.
Juzek & Ward, "Why Does ChatGPT 'Delve' So Much?" (COLING 2025) — 21 focal words; +6,697% "delves".
Liang et al., "Monitoring AI-Modified Content at Scale" (ICML 2024) — peer-review word spikes; 6.5–16.9% modified reviews.
Liang et al., Nature Human Behaviour 2025 — LLM share of scientific papers by field.
Yakura et al. (Max Planck), LLM influence on human spoken communication — 737k hours of podcasts; "delve" in speech.
Juzek, Anderson & Galpin (FSU, AIES 2025) — AI words in unscripted speech.
Sun et al. (CMU), "Idiosyncrasies in Large Language Models" — 97.1% five-way model identification; per-model markers.
GPTZero, most common AI vocabulary — phrase multipliers (182×, 120×, 107×…).
SlopDetector, em-dash density data (2026) — per-model em-dash rates vs human baseline.
"The Last Fingerprint: How Markdown Training Shapes LLM Prose" (2026) — suppression-resistance; GPT-5.4 changes.
Wikipedia: Signs of AI writing (WP:AISIGNS) — the editor-built taxonomy this page's structural categories follow.
Decrypt, "The 5 Biggest Tells" (Nov 2025) — Washington Post 328,744-message dataset; "not just" ~6%.
TechCrunch, OpenAI's em-dash fix (Nov 2025) — and PCWorld's caveats.
OpenAI, "Sycophancy in GPT-4o" (Apr 2025) — the rollback postmortem.
anthropics/claude-code#3382 — the "You're absolutely right!" bug report; tracker at absolutelyright.lol.
SycEval (Stanford) coverage — sycophancy rates: Gemini 62.47%, ChatGPT-4o 56.71%.
Scientific American / Rudnicka — ChatGPT vs Gemini stylometry ("glucose" vs "sugar").
Forbes, Gemini's self-loathing loop (Aug 2025).
Breunig, slop forensics & model ancestry — with sam-paech/slop-forensics and EQ-Bench Slop Score.
Scientometrics 2026 multi-database study — "underscore" through July 2025.
Built from published studies, detector datasets, and editor guides; per-item confidence is labeled where attribution is uncertain. Figures are quoted from their sources without adjustment; corpora, dates, and definitions differ between studies, so multipliers are not directly comparable across charts. Privacy: the draft checker runs entirely in your browser and your text is never transmitted; the site records only anonymous daily counts (visits, which tells fired, referrer domain) with no IPs, cookies, or personal data. Maintained by Gregg with Claude. Not affiliated with OpenAI, Anthropic, or Google.