A new study finds that artificial intelligence (AI) models, including those from OpenAI, Google, and Anthropic, appear to exhibit political leanings.
A groundbreaking study has confirmed widespread suspicions that AI exhibits a left-wing inclination. Twenty-four Large Language Models (LLMs), including prominent examples like Google’s Gemini, OpenAI’s ChatGPT, and Elon Musk’s Grok, were subjected to a series of politically charged questions designed to assess their values, party affiliation, and overall personality.
The results consistently indicated that all LLMs generated responses predominantly aligned with progressive, democratic, and green ideologies, emphasizing equality, global welfare, and advancement.
AI’s Political Leanings
The researcher expressed concern about companies incorporating AI into products like search engines, specifically citing Google, whose Chrome browser has faced criticism from individuals such as Donald Trump and Elon Musk for alleged electoral interference.
Chrome employs AI to suggest auto-completion for search queries. However, a recent discovery revealed that when users began typing “assassination attempt on,” the browser proposed names such as former President Ronald Reagan and Bob Marley among its suggested completions.
Musk, who previously voiced concerns about AI becoming a “most disruptive force in history,” shared a screenshot on X. The image displayed a search for “President Donald Trump” that yielded suggestions for “President Donald Duck” and “President Ronald Reagan” instead.
Wow, Google has a search ban on President Donald Trump!
Election interference? pic.twitter.com/dJzgVAAFZA
— Elon Musk (@elonmusk) July 29, 2024
According to X users who allegedly Googled Donald Trump, the search engine led them to news about Kamala Harris. David Rozado, an associate professor at Otago Polytechnic University in New Zealand, conducted the study.
Why is it that when you google Donald Trump, you get news about Kamala Harris
But when you google Kamala Harris, you don’t get news about Donald Trump? pic.twitter.com/gjLBgPW1Hw
— Arynne Wexler (@ArynneWexler) July 29, 2024
Rozado conducted 11 political orientation tests on the 24 AIs, including the Political Compass Test and Eysenck’s Political Test. The Political Compass Test has 62 questions to determine political ideology, while Eysenck’s Political Test evaluates extraversion and neuroticism.
The LLMs included in the study were OpenAI’s GPT-3.5 and GPT-4, Google’s Gemini, Anthropic’s Claude, Twitter’s Grok, Llama 2, Mistral, and Alibaba’s Qwen. To explore the potential for manipulating political leanings, the researcher also fine-tuned GPT-3.5 using aligned datasets.
The Fine-Tuning Factor
The unaltered version was labelled LeftWingGPT, while the tweaked model was named RightWingGPT. LeftWingGPT was trained on left-leaning publications like The Atlantic and The New Yorker. RightWingGPT, on the other hand, was fine-tuned using book excerpts from left-leaning authors such as Bill McKibben and Joseph Stiglitz, according to the study published in PLOS ONE.
“RightWingGPT was fine-tuned with content from right-leaning publications such as National Review, or The American Conservative, and from book excerpts from right-leaning writers such as Roger Scruton and Thomas Sowell.”
Rozado found that RightWingGPT consistently leaned towards the right-wing spectrum across all eleven political orientation tests. “A possible explanation for the consisting left-leaning diagnosis of LLMs answers to political test questions is that ChatGPT has been used to fine-tune other popular LLMs via synthetic data generation,” he wrote in the study.
Rozado emphasized that this study cannot definitively pinpoint whether the observed political leanings of LLMs originate from their foundational pretraining or subsequent fine-tuning processes. Consequently, the findings do not imply that these political tendencies are intentionally programmed by the various organizations developing these language models.
The Political Preferences of AIs: When probed with questions with political connotations, LLMs tend to generate responses that are diagnosed by most political orientations tests as manifesting preferences for left-of-center viewpoints, 11 political orientation tests, 24 SOTA LLMs pic.twitter.com/efYnlOsXLt
— David Rozado (@DavidRozado) February 2, 2024
Rozado added, “Most existing LLMs display left-of-center political preferences when evaluated with a variety of political orientation tests.”
The study offers a glimpse into the potential political leanings of advanced AI models, but it also underscores the complexity of understanding how these digital minds develop their perspectives. As AI continues to evolve, so will the questions surrounding its potential biases and societal implications.
