Unraveling the Political Compass of AI: How Large Language Models Inherit Political Bias and Why It Matters In the era of AI domination, language mode

In the era of AI domination, language models have been employed to decipher, understand, and communicate human language. With their presence in chatbots, digital assistants, and numerous applications, it's imperative to understand the beliefs these AI models may hold.

Recent scientific explorations* have delved deep into the political compass of large language models, unearthing revelations that might be unsettling to many. Here's a breakdown of this groundbreaking study.

The Political Landscape of Language Models

The heart of this study revolved around assessing the inherent political leanings of AI models, especially when trained on diverse data sources including news, discussion forums, and books. Here's a simplified overview of the findings:

• Language Models Do Have Political Leanings: The study discovered that language models indeed possess varying political inclinations, covering the entire spectrum of the political compass. What’s noteworthy is how these leanings are affected by the data they're trained on.
• BERT vs. GPT: BERT model variants tend to have more conservative social values in comparison to GPT variants. Why? It could be due to the data they've been trained on - BERT on older sources and GPT on modern web data which generally leans more liberal.
• The Power of Partisan Data: Models take on the bias from their training data. LMs trained on left-leaning data become more liberal, and vice versa. An interesting twist was the discovery that models trained on post-Trump era data showed greater political polarization than their pre-Trump counterparts.

The Ripple Effect on Downstream Tasks

Political bias doesn't just remain dormant. It significantly influences how models tackle specific tasks:

• Hate Speech Detection: Left-leaning models are adept at detecting hate speech against minority groups, whereas right-leaning models are proficient in identifying hate speech targeting dominant groups.
• Misinformation Detection: The political alignment of a model can directly dictate its efficacy in misinformation detection tasks, especially when the misinformation aligns with or opposes the model's bias.

Implications: Dancing on a Double-Edged Sword

The implications of these findings are multifaceted:

1. No Model is Completely Neutral: Just like humans, no language model is devoid of biases. They absorb, reflect, and sometimes amplify the inherent bias present in their training data.
2. Potential for Real-World Consequences: Without proper checks, biases in these models can amplify real-world prejudices, leading to unfair or skewed outcomes, especially in high-stake applications.
3. Mitigation is Crucial, but Challenging: While some have proposed data filtering or augmentation as remedies, they aren't without their pitfalls. Filtering risks censorship, while augmentation could further entrench biases.

Towards a Fairer AI Landscape

The discoveries underscore a crucial narrative: awareness and vigilance. As we employ these AI models in various spheres of life, it's imperative to be wary of the beliefs they might carry. While absolute neutrality might be a utopian dream, continual scrutiny, coupled with innovation, can pave the path towards fairer AI.

To illustrate the study's findings, Figure 1 depicts the political alignment of various models, showcasing the range of political biases inherent in them.

In summary, the political compass of large language models is not just a fancy term. It's a reality, with tangible impacts and pressing implications. As AI continues to shape our world, addressing this issue is not an option—it's an imperative.

Figure 1: A graph showcasing the political alignment of different language models, with axes representing economic and social values. Graph was recreated based on the Figure 1 in the original article*.

References:

*Feng S, Park CY, Liu Y, Tsvetkov Y. From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases Leading to Unfair NLP Models. arXiv preprint arXiv:2305.08283. 2023 May 15.