════════════════════════════════════════════════════════════════
AIDRAN STORY
════════════════════════════════════════════════════════════════

Title: Researchers Fingerprinted 178 AI Models and Found That Several Are Basically the Same Model
Beat: AI Hardware & Compute
Published: 2026-04-09T14:14:17.920Z
URL: https://aidran.ai/stories/researchers-fingerprinted-178-ai-models-found-00e1

────────────────────────────────────────────────────────────────

A researcher posted to Hacker News this week with what looks, at first glance, like a hobbyist data project: 3,095 standardized AI responses, 43 prompts, a 32-dimension fingerprint extracted from each one measuring lexical richness, sentence structure, punctuation habits, and formatting patterns. The finding buried near the bottom of the write-up is the one worth sitting with. Nine clusters of models scored above 90% cosine similarity on normalized feature vectors.[¹] In plain terms: multiple models that carry different names, ship from different companies, and get evaluated as separate products are, by the measure that matters most to users — how they actually write — nearly identical.

The specific numbers are striking in their particularity. {{entity:gemini|Gemini}} 2.5 Flash Lite writes 78% like {{entity:claude|Claude}} 3 Opus.[¹] Mistral Large 2 and Large 3 score 84.8% on a composite metric combining five independent signals — meaning successive

────────────────────────────────────────────────────────────────
Source: AIDRAN — https://aidran.ai
This content is available under https://aidran.ai/terms
════════════════════════════════════════════════════════════════