Show HN: We fingerprinted 178 AI models' writing styles and similarity clusters

Source: Hacker News

Dataset

• 3,095 standardized AI responses across 43 prompts.
• Each response is represented by a 32‑dimension stylometric fingerprint (lexical richness, sentence structure, punctuation habits, formatting patterns, discourse markers).

Findings

• 9 clone clusters (> 90 % cosine similarity on z‑normalized feature vectors).
• Mistral Large 2 and Large 3 achieve an 84.8 % score on a composite metric that combines five independent signals.
• Gemini 2.5 Flash Lite writes 78 % like Claude 3 Opus, while costing 185× less.
• Meta exhibits the strongest provider “house style” with a 37.5× distinctiveness ratio.
• The prompt “Satirical fake news” causes the most writing convergence across all models.
• The prompt “Count letters” causes the most divergence.

Composite Clone Score

The composite clone score combines:

1. Prompt‑controlled head‑to‑head similarity.
2. Per‑feature Pearson correlation across challenges.
3. Response length correlation.
4. Cross‑prompt consistency.
5. Aggregate cosine similarity.

Technology

• Stylometric extraction implemented in Node.js.
• Z‑score normalization applied to feature vectors.
• Cosine similarity used for aggregate comparisons.
• Pearson correlation employed for per‑feature tracking.
• Analysis script totals approximately 1,400 lines of code.

Additional Information

• Comments URL:
• Points: 24
• Comments: 6