🔍 lmscan

Detect AI-generated text. Fingerprint which LLM wrote it. Open-source GPTZero alternative.

GPTZero charges $15/month. Originality.ai charges per scan. Turnitin locks you into institutional contracts.

lmscan is free, open-source, works offline, and tells you which model wrote the text.

$ lmscan "In today's rapidly evolving digital landscape, it's important
to note that artificial intelligence has become a pivotal force in
transforming how we navigate the complexities of modern life..."

🔍 lmscan v0.1.0 — AI Text Forensics
══════════════════════════════════════════════════

  Verdict:     🤖 Likely AI (77% confidence)
  Words:       184
  Sentences:   10
  Scanned in 0.01s

┌────────────────────────────┬──────────┬────────────────────┐
│ Feature                    │ Value    │ Signal             │
├────────────────────────────┼──────────┼────────────────────┤
│ Burstiness                 │ 0.07     │ 🔴 Very low (AI)    │
│ Sentence length variance   │ 0.27     │ 🟡 Below average    │
│ Slop word density          │ 20.7%    │ 🔴 High (AI)        │
│ Transition word ratio      │ 2.2%     │ 🟡 Elevated         │
│ Readability consistency    │ 0.00     │ 🔴 Very low (AI)    │
│ ...                        │          │                     │
└────────────────────────────┴──────────┴────────────────────┘

🔎 Model Attribution
  1. GPT-4 / ChatGPT    62% — "delve", "tapestry", "beacon", "landscape" (×2), +19 more
  2. Claude (Anthropic)  13% — "robust", "nuanced", "comprehensive"
  3. Gemini (Google)      9% — "furthermore", "additionally"

⚠️  Flags
  • Very low burstiness (0.07) — AI text is more uniform in complexity
  • High slop word density (20.7%) — contains known AI vocabulary markers

Install

pip install lmscan

Zero dependencies. Works with Python 3.9+. No API keys. No internet. No GPU.

Usage

# Scan text directly
lmscan "Your text here..."

# Scan a file
lmscan document.txt

# Pipe from stdin
cat essay.txt | lmscan -

# JSON output (for scripts and CI)
lmscan document.txt --format json

# Per-sentence breakdown
lmscan document.txt --sentences

# CI gate: fail if AI probability > 50%
lmscan submission.txt --threshold 0.5

Python API

from lmscan import scan

result = scan("Text to analyze...")

print(f"AI probability: {result.ai_probability:.0%}")
print(f"Verdict: {result.verdict}")
print(f"Confidence: {result.confidence}")

# Which model wrote it?
for model in result.model_attribution:
    print(f"  {model.model}: {model.confidence:.0%}")
    for evidence in model.evidence[:3]:
        print(f"    → {evidence}")

# Per-sentence analysis
for sentence in result.sentence_scores:
    if sentence.ai_probability > 0.7:
        print(f"  🤖 {sentence.text[:60]}... ({sentence.ai_probability:.0%})")

Scan entire directories

from lmscan import scan_file
import glob

for path in glob.glob("submissions/*.txt"):
    result = scan_file(path)
    print(f"{path}: {result.verdict} ({result.ai_probability:.0%})")

How It Works

lmscan uses 12 statistical features derived from computational linguistics research to distinguish AI-generated text from human writing:

Each feature produces a signal via sigmoid transformation. The weighted combination produces the final AI probability.

Model Fingerprinting

lmscan includes vocabulary fingerprints for 5 major LLM families:

Attribution uses weighted vocabulary matching, phrase detection, and hedging pattern analysis.

Accuracy & Limitations

What lmscan is good at:

• Detecting text with strong AI stylistic patterns
• Identifying which model family generated text
• Scanning at scale (thousands of documents) with zero cost
• Providing explainable evidence (not a black box)

What lmscan cannot do:

• Detect AI text that has been manually edited or paraphrased
• Work reliably on very short text (<50 words)
• Detect AI text in non-English languages (English-only for now)
• Replace human judgment — use as a signal, not a verdict

This is statistical analysis, not a neural classifier. It detects stylistic patterns, not watermarks. It works best on unedited LLM output and degrades gracefully on edited text.

CI Integration

GitHub Actions

- name: AI Content Check
  run: |
    pip install lmscan
    lmscan submission.txt --threshold 0.7 --format json

Pre-commit

repos:
  - repo: https://github.com/stef41/lmscan
    rev: v0.1.0
    hooks:
      - id: lmscan
        args: ["--threshold", "0.7"]

Research Background

lmscan's approach is informed by published research on AI text detection:

• DetectGPT (Mitchell et al., 2023) — perturbation-based detection using log probability curvature
• GLTR (Gehrmann et al., 2019) — statistical visualization of token predictions
• Binoculars (Hans et al., 2024) — cross-model perplexity comparison
• Zipf's Law in NLP — word frequency distributions differ between human and AI text
• Stylometry — decades of authorship attribution research applied to AI forensics

lmscan takes the statistical intuitions from these papers and implements them as lightweight, dependency-free heuristics that work without requiring a reference language model.

FAQ

Q: Is this as accurate as GPTZero? A: GPTZero uses neural classifiers trained on labeled data. lmscan uses statistical heuristics. GPTZero is more accurate on edge cases; lmscan is free, offline, and explainable. Use both if accuracy matters.

Q: Can students use this to evade AI detection? A: lmscan shows which features trigger detection, which could help someone understand why text reads as AI-generated. This is by design — understanding AI writing patterns makes everyone a better writer. The same information is available in published research papers.

Q: Does it work on non-English text? A: Currently English-only. The slop word lists and transition word lists are English-specific. Statistical features (entropy, burstiness) work across languages but haven't been calibrated.

Q: Does it phone home? A: No. Zero network requests. No telemetry. No API keys. Everything runs locally.

Q: How is model attribution possible without running the model? A: Each LLM family has characteristic vocabulary biases. GPT-4 loves "delve" and "tapestry". Claude says "I'd be happy to". These are statistical fingerprints — not guaranteed attribution, but strong signals.

See Also

• reverse-SynthID — Reverse-engineering Google's image watermarking
• vibesafe — AI code safety scanner
• injectionguard — Prompt injection detection
• vibescore — Grade your vibe-coded project

License

Apache-2.0