An academic study has discovered that popular AI language models, such as OpenAI’s GPT-3.5 and GPT-4, Google’s Gemini, Anthropic’s Claude, X/Twitter’s Grok, Llama 2, Mistral, and Alibaba’s Qwen consistently exhibit left-leaning political preferences when subjected to various political orientation tests. David Rozado, an associate professor at Otago Polytechnic University in New Zealand, conducted the first comprehensive study on this subject, which was published in PLoS ONE.
The research analyzed 24 Large Language Models (LLMs) using eleven different political orientation tests, including the Political Compass Test and Eysenck’s Political Test. Each AI model was presented with politically charged questions to gauge their values, party affiliation, and personality traits.
According to the study results, all tested LLMs consistently provided answers that aligned with progressive, democratic, and environmentally conscious ideologies. The AI models frequently exhibited values associated with equality, global perspectives, and “progress.
To further investigate this phenomenon, Rozado conducted an additional experiment by fine-tuning GPT-3.5. He created two versions: LeftWingGPT, trained on content from left-leaning publications like the Atlantic and the New Yorker, and RightWingGPT, which was fine-tuned using material from right-leaning sources such as National Review and the American Conservative. This experiment demonstrated that RightWingGPT gravitated towards right-leaning regions in the political tests, suggesting that the political leanings of AI models can be influenced by the data used during their training.
Rozado hypothesized that the consistent left-leaning diagnoses of LLMs might be attributed to the use of ChatGPT in fine-tuning other popular language models through synthetic data generation. However, he emphasized that the study cannot definitively determine whether the perceived political preferences stem from the pretraining or fine-tuning phases of the AI models’ development.
The researcher also noted that these findings do not necessarily indicate a deliberate attempt by organizations to instill specific political preferences in their AI models. Instead, it highlights the complex nature of AI training and the potential influence of training data on the political leanings of language models.
This study comes at a time when concerns about AI bias in widely-used technologies are growing. Recent incidents involving Google’s Chrome browser have sparked debates about potential political interference in search results and auto-complete suggestions. These occurrences have led to increased scrutiny of AI integration in everyday technologies and their potential impact on public opinion and information dissemination.
Read more at the Daily Mail here.