LLMs Are Left-Leaning Liberals: The Hidden Political Bias of LLMs (www.scl.org)

A recent study by the Cripps AI and Data team has unveiled a significant and potentially concerning political bias in large language models (LLMs), demonstrating a consistent tendency towards left-leaning perspectives. Through a mix of multiple-choice and open-ended political orientation tests, the researchers found that when forced to choose between partisan answers, LLMs frequently selected left-leaning responses. This finding adds to existing academic research indicating that LLMs generally align more closely with leftist political views, despite being trained on diverse data sets from the internet, which many believe would yield more neutral or extreme outputs. The significance of these findings lies in the implications for users' trust and the ethical framework surrounding LLM deployment. As LLMs increasingly serve as sources of information or decision support, their inherent biases could shape public understanding and discourse in ways that compromise political neutrality. The study raises critical questions regarding the nature of bias in AI, especially in light of regulatory efforts like the EU AI Act, which seeks to mitigate harmful biases. Overall, understanding and addressing the left-leaning bias in LLMs is crucial for fostering fair and balanced AI systems in a democratic society.