Study Reveals: AI Tends To Take Left Direction Politics

OpenAI's chatGPT is accused of producing more answers from left-wing politics (photo: x @openai)

JAKARTA - Recent studies reveal that large language models (LLM) such as Google Gemini, Elon Musk's OpenAI ChatGPT, and Grok tend to have a left-handed political bias. Research conducted by David Rozado, a professor at Otago Polytechnic University in New Zealand, shows that most of the LLM tested yields more progressive, democratic, and value-oriented answers such as equality and progress.

Rozado conducted 11 political orientation tests, including the Political Compass Test and Eysenck's Political Test, against 24 language models, including GPT 3.5 and GPT-4 from OpenAI, Gemini from Google, Claude from Anthropic, Grok from Twitter, Llama 2, Mistral, and Qwen from Alibaba. Test results show that the answers given by LLM generally reflect left political views.

Research also involves fine-tuning on GPT 3.5 to change its political preferences. The unmodified model is labeled LeftWingGPT, while the one that has been modified to RightWingGPT.

LeftWingGPT is trained with left-leaning publications such as The Atlantic and The New Yorker, while RightWingGPT is trained with content from publications and right-leaning writers such as National Review and Roger Scruton.

The fine-tuning results show that RightWingGPT shows a more right-leaning political preference, but Rozado notes that this study cannot determine whether this political preference stems from the early training phase or fine-tuning.

Rozado stressed that these findings do not indicate that this political preference was deliberately applied by the organization that manufactures LLM. Even so, many of the current NGOs show left-leaning political preferences when evaluated with various political orientation tests.

These results raise concerns about AI integration in products such as Google search engines, which have come under criticism for intervening in politically related search results.