Skip to content

Iván Arcuschin0

@IvanArcus · view on X ↗

Independent Researcher | AI Safety & Software Engineering

1.4K followers
1 thread on ThreadOut

Unrolled threads

Every thread by Iván Arcuschin that readers unrolled on ThreadOut.

Feb 11, 2026
You change one word on a loan application: the religion. The LLM rejects it. Change it back? Approved. The model never mentions religion. It just frames the same debt ratio differently to justify op...
SummaryResearchers developed an automated pipeline to detect hidden biases in large language models that influence decisions without being mentioned in reasoning. Testing across 6 frontier models on hiring, loans, and admissions, they found systematic biases related to gender, race, religion, and language proficiency that models don't verbalize, suggesting chain-of-thought monitoring alone is insufficient.15 tweets2 min read