← ATH

RESEARCH · GLOBAL

Researchers extract reasoning traces from Claude, GPT, and Gemini; find evidence of training on US models

Academic researchers developed a technique to extract internal "reasoning traces" from leading LLMs (Claude, GPT-4, Gemini). Analysis suggests some Chinese AI models were trained on outputs from US frontier models, raising IP and security concerns.

WHY IT MATTERS

For compliance and risk officers: this technique could become a tool for auditing LLM behavior and detecting model lineage. For security: confirms that model weights and outputs are being reverse-engineered at scale. Expect regulatory pressure on training data provenance.

Source: Wired · 2026-08-11

← BACK TO TODAY'S DECK

Researchers extract reasoning traces from Claude, GPT, and Gemini; find evidence of training on US models — ath — AITechHive