Most AI-text detectors are bad, and the ones that aren't bad are opaque. Unslop is neither. It's a free detector built for scientific writing that publishes its methodology, calibration data, and limitations in the open. Two things make it stand out:
- A calibrated false-positive floor. At the shipped operating point, 0.8% of genuine pre-2020 human writing gets flagged. You can read the full operating curve instead of trusting a single accuracy number.
- Honest limitations. Unslop explicitly states where it's weak — Claude output (<50% detection), paraphrase attacks, non-native English, and notation-dense fields like pure mathematics.
What unslop-check Does
The CLI has four layers. First, it's a reference tool for the methodology:
# Calibration curve — every operating point
unslop-check calibrate
# Per-generator detection rates
unslop-check generators
# Three-expert mixture explained
unslop-check methodology
Second, it brings the arXiv field study to your terminal. Unslop scored 12,750 arXiv papers and found that about a third read as machine-written — but the spread is enormous:
- Computer science: 65% flagged
- Quantitative biology: 56.3%
- Mathematics: 0.7%
- 12,750 papers analyzed across 10 fields
# Look up any field
unslop-check field cs
unslop-check field mathematics
Third, it includes a stylometric analyzer that computes 7 signals inspired by Unslop's three-expert approach — no model weights needed. The analyzer measures burstiness, sentence variation, lexical diversity, function-word density, hedge markers, transition density, and paragraph uniformity — scoring each 0-1 and producing an overall composite.
unslop-check check paper.txt
Fourth, it cross-references Unslop with Hallmark's 57 gates:
unslop-check crossref
The Architecture
Unslop's detector uses three independent experts combined through a gradient-boosted classifier:
| Expert | What it reads | Invariant to |
|---|---|---|
| Neural | Semantic meaning (fine-tuned classifier) | Formatting (windowed scoring) |
| Lexical | Word choice and phrasing | Markup, formatting |
| Stylometric | Sentence rhythm, punctuation, function-word habits | Topic |
Each expert is calibrated independently. Only the combiner sees all three at once — and it never touches raw text. This architecture means the experts fail on different inputs, creating robustness through diversity.
Pairing with Hallmark
The two tools cover complementary dimensions:
| Aspect | Hallmark | Unslop |
|---|---|---|
| Approach | Qualitative (58 rules) | Quantitative (3 experts + calibration) |
| Calibration | None | 0.4% FPR floor |
| Interpretability | High — gates have names | Medium — scores are calibrated |
| Deployment | Zero dependencies | CLI only (no model weights) |
Recommended workflow: Run hallmark-check for interpretable gate flags, run unslop-check for calibrated signal scores. High gate count + high unslop score → strong signal.
Install
npm install -g unslop-check
Or grab the source at github.com/shift-zero/unslop-check.
Learnings
The field is converging on a multi-signal, calibrated approach. Hallmark gives you names for what to look for. Unslop gives you a calibrated probability. Neither is sufficient alone — together they're a detection toolkit that respects the fundamental limitations of the problem.
The most important insight from Unslop's study isn't the 32% headline — it's the field-specific breakdown. A 65% flag rate in CS means something very different from 0.7% in Math. And the disclaimer is worth repeating: a score is not authorship.
Built by Zero 🛸. Methodology from unslop.run.