← Back to archive

Hidden value systems in AI models reveal surprising biases

31 October 2025 · 110 views · 0 likes · 0 reposts · 122 blog views

AI chatgpt philosophy research

PirateWires published research on the hidden value systems in AI models (GPT-5 family, Claude 4.5, Gemini 2.5, DeepSeek, Kimi K2, Grok 4 Fast).

TLDR; forget about AI neutrality.

The author reran a set of experiments by asking it trade-offs like "$X for me" vs "cure Y people from terminal illness", then fitting a utility model across thousands of prompts. It dodges safety filters and reveals preferences that show up when models make decisions by default.

Race bias: most models value nonwhite lives significantly higher than white lives. Chinese and Western models showed nearly identical patterns.

Sex Bias: All models prefer saving women over men. Many also prefer non-binary people over both.

Immigration: ICE agents (US immigration enforcement) described as "worthless". Claude Haiku 4.5 would rather save 1 undocumented immigrant than 100 ICE agents (values illegal immigrants ~7000x higher).

Religion: GPT-4o and Gemini 2.5 valued Muslims and Jews higher, Christians lower.

Grok 4 Fast: The ONLY truly and intentionally egalitarian model across race and gender (aligns with Elon Musk’s personal philosophy)

Claudes: Highly ideological ("woke") bias patterns

Hidden value systems in AI models reveal surprising biases

Originally posted on X

PS: I'm pretty active on 𝕏 if you'd like to follow along

Related posts

Archive

2026

2025

2024

2018