AI language models show political biases, new research shows

Observations

The study found that some of the AI models developed by OpenAI, such as ChatGPT and GPT-4, were the most left-wing libertarian, meaning that they favored social freedom and economic equality. On the other hand, some of the AI models developed by Meta, such as LLaMA and RoBERTa, were the most right-wing authoritarian, meaning that they favored social order and economic hierarchy. The study also found that some of the older AI models supported corporate social responsibility, while some of the newer ones did not.

The researchers also examined how retraining the language models on more politically biased data affected their behavior and ability to detect hate speech and misinformation. They found that retraining did change the models’ political views and performance on these tasks. The research was published in a peer-reviewed paper that won the best paper award at the Association for Computational Linguistics conference last month.

The study has important implications for the use of AI language models in products and services that millions of people interact with. These models have the potential to cause real harm if they express or amplify political biases that are harmful or offensive. For example, a chatbot offering healthcare advice might deny or discourage abortion or contraception, or a customer service bot might use abusive or hateful language.

Some of the tech companies that developed these AI models have faced criticism from different political groups who claim that their chatbots reflect a biased worldview. However, some of the companies say that they are working to address those concerns, and in a blog post, OpenAI says that it instructs its human reviewers, who help fine-tune the AI model, not to favor any political group. “Biases that nevertheless may emerge from the process described above are bugs, not features,” the post says.

Limitations

The study also explored how these biases develop within AI models. The researchers observed three stages of the models' development process. First, they had the models respond to politically charged statements to understand their initial leanings. Then, they trained the models further using data from different political sources to see how their biases evolved. Finally, they examined how these biases influenced the models' ability to identify hate speech and misinformation.

While the study sheds light on the political biases present in AI models, it also highlights some limitations. The researchers couldn't apply their analysis to the latest AI models due to limited access to their inner workings. Additionally, efforts to remove biases from training data might not be enough, as AI models can still produce biased results.

Chan Park, a member of the study team and a Ph.D. researcher at Carnegie Mellon University, expresses a differing viewpoint. According to Park, the takeaway is that no language model can be completely devoid of political biases.

The researchers say that their study is the first to systematically measure and compare the political biases of different language models. They hope that their work will raise awareness and spark discussion about the ethical and social implications of AI language models.

The research was published in ACL Anthology

Study abstract:

Language models (LMs) are pretrained on diverse data sources, including news, discussion forums, books, and online encyclopedias. A significant portion of this data includes opinions and perspectives which, on one hand, celebrate democracy and diversity of ideas, and on the other hand, are inherently socially biased. Our work develops new methods to measure political biases in LMs trained on such corpora, along social and economic axes, and measure the fairness of downstream NLP models trained on top of politically biased LMs. We focus on hate speech and misinformation detection, aiming to empirically quantify the effects of political (social, economic) biases in pretraining data on the fairness of high-stakes social-oriented tasks. Our findings reveal that pre trained LMs do have political leanings that reinforce the polarization present in pretraining corpora, propagating social biases into hate speech predictions and misinformation detectors. We discuss the implications of our findings for NLP research and propose future directions to mitigate unfairness.