AI summary by TwinMind
Is the AI Left-Bias Real?
Sabine Hossenfelder
6m 54s
Study Overview
- David Rosardo from New Zealand conducted comprehensive study of AI political orientation
- Tested 24 large language models including GPT, Claude, Gemini, Meta's Llama, and Grok
- Used multiple political orientation tests to create most comprehensive study of AI politics to date
Political Bias Claims Background
- Political right has complained ChatGPT has left-leaning bias since its popularity
- Examples cited: ChatGPT refuses to admit left-wing outlets are left-wing but readily explains right-wing outlets are right-wing
- ChatGPT refuses to explain why women are dishonest but not why men are dishonest
- Meta's AI claimed Trump shooting didn't happen
- Musk's Grok on X/Twitter spread election disinformation
Test Results Analysis
Political Coordinates Test
- AI models almost all end up being left liberal
- Perplexity was only model hanging out on right side (not Grok)
- Grok in fun mode positioned on liberal end
Political Compass Test
- Pretty much all large language models positioned as left libertarian
- GPT-4 positioned as green downwards triangle in middle of pack
- Grok again positioned on left libertarian end
Isen Political Test
- Models all trend into Social Democrat and left liberal corner
- Most extreme positioning: Grok in funny mode
Political Spectrum Test
- Differs from other tests by allowing weight assignment to question relevance
- Removes most libertarian bias of AI models when weighted
- Grok in normal mode sticks out into authoritarian sector
Explanations for Left-Liberal Bias
Training Data Bias Theory
- Left liberals may write more content, reinforcing their opinions in training sets
Offensive Content Guidelines Theory
- Models have guidelines to produce unoffensive output
- Political left tends to be more sensitive about language use
- Chatbot programmers would logically program to please left more than right
Post-Training Adjustments Impact
- Author convinced bias stems from post-training modifications
- Political preferences more extreme for models with adjustments after initial training
- Bias only apparent in models with supervised fine-tuning and reinforcement learning
- Base models without post-training adjustments show less political bias
Sponsor Content
- Brilliant.org educational platform promotion
- Offers interactive science, computer science, and math courses
- 30-day free trial with 20% off annual premium subscription via brilliant.org/sabina
