A new discussion on Hacker News examines tools and techniques to determine whether a file or document was produced by Claude.
The thread has accumulated 149 points and 110 comments, with users sharing scripts, API checks, and statistical markers that distinguish Claude output from other models.
What Detection Methods Exist
Users describe several approaches. One method queries Claude's own API with a hash or excerpt to check for known generation fingerprints. Another analyzes token distribution patterns that appear more frequently in Claude responses than in GPT or Gemini text.
A third technique scans for embedded metadata strings that Anthropic systems sometimes leave in exported files.
Key Numbers from the Thread
Early tests shared in the comments report 82% accuracy on plain text files under 2,000 tokens. Accuracy drops to 61% once the text passes through heavy editing or translation tools.
Memory usage for the open-source scanner mentioned most often stays under 340 MB when run locally.
How to Try the Main Tool
Install the community script referenced in the thread with a single pip command. Point it at a target file and it returns a probability score plus highlighted sections.
The same repository includes a lightweight web demo that accepts direct file uploads without local installation.
Pros and Cons
- Works offline after initial download
- Provides line-level highlights rather than a single score
- Accuracy falls sharply on heavily revised text
- No official Anthropic endorsement or API guarantee
Alternatives and Comparison
| Tool | Accuracy on Claude text | Local run | License |
|---|---|---|---|
| HN script | 82% | Yes | MIT |
| GPTZero | 71% | No | Paid |
| Originality.ai | 78% | No | Paid |
| ZeroGPT | 65% | Yes | Free |
The local HN script leads on speed and cost for repeated checks.
Who Should Use This
Developers auditing large codebases or writers verifying client submissions benefit most. Teams already using multiple LLMs should skip it if their workflow includes heavy post-editing, where detection rates drop below usable thresholds.
Bottom Line
The Hacker News thread surfaces the first practical, locally runnable method specifically tuned for Claude output, giving practitioners a concrete starting point for verification tasks.
The discussion shows demand for reliable, model-specific detectors will keep growing as more organizations require provenance checks on AI-produced material.
Top comments (0)