Site Navigation

Download a zip file of
all pages

for Obsidian or LLMs.

WORK IN PROGRESS:
PAGES NOT YET REVIEWED BY
HUMAN EXPERTS. VERIFY
CLAIMS AND CONSULT
ORIGINAL SOURCES FOR
AUTHORITATIVE INFORMATION.


Contact Peter Kaminski

Edit on GitHub


Training Data Bias

Note: Core issue in Alex Kladitis's discovery about AI Bias and Perspectives during the OGM 2025-11-06 call.

Alex's Finding

ChatGPT's bias "inbuilt into the training data" - couldn't be overridden by explicit instructions

What is Training Data Bias?

AI systems learn from training data, which reflects:

The Problem

It's invisible to in-group:

It's hard to remove:

Examples from Call

ChatGPT: Couldn't provide neutral Chinese news summaries, always added Western editorial framing

DeepSeek: Censored Tiananmen Square, reflected Chinese government perspective

Implications

Related Concepts


Back to README


Pages that link to this page