Understanding political bias and information transparency in the age of AI

In the current polarized political landscape, the average person is having a harder and harder time accessing fair and unbiased news sources. At the same time, with the proliferation of AI agents and LLMs people are increasingly turning to these entities rather than traditional search engines as sources of information. While there has been extensive policymaking and normsetting in traditional news media around mitigating bias, efforts to evaluate political bias in AI have failed to root themselves in the current state of the world, instead focusing too much on theoretical ideals abstracted from reality.

What is AI?

Before I dig deeply into the concept of bias, political bias more specifically, and most importantly how it manifests in the world of AI, I think it’s important to really define what I mean when I say ‘AI’ and ‘LLM’. In the broadest sense artificial intelligence (AI) refers to the goal of making computers perform tasks that typically and have historically required human input. I heard an excellent quote from Karen Hao on the Intercept Podcast where she likened the term AI to the term Transportation. Transportation is an incredibly broad term that refers to anything from a bicycle to a rocket ship. Likewise, AI can refer to a huge range of technologies. The important concept to understand is that AI and Machine Learning models use a variety of algorithms to learn patterns from the data they are given rather than having those patterns explicitly programmed. Machine Learning has existed since the 60s/70s but in the last 5ish years the use of AI in our everyday lives has really taken off. Specifically, one kind of AI, called Large Language Models (LLMs) – like ChatGPT, Claude or Gemini – have come to the forefront. These models are incredibly complex and have been trained on HUGE amounts of data to be able to carry out what can seem like a human conversation. As these models become more and more accessible to the average person, we need to consider the impacts that they are having in terms of how people interact with them, other people, and society.

What do we mean when we say bias?

In Machine Learning, bias is defined as the level to which a model is too simple to pick up true patterns from data. This is a pretty abstract definition so it can be helpful to think of the following interpretation: bias occurs when algorithms or models give higher weights or representation to some aspects of a dataset than others, resulting in skewed outcomes and lower accuracy in predictions. When applied to LLMs, bias is often thought of as the overrepresentation of one side of an argument/issue than the other in the models’ responses to users prompts. (Prompt is the word used in Machine Learning to refer to the text input that a person gives to an AI model.)

This definition of bias becomes problematic when it gets applied to social contexts such as politics. Most of the literature on political bias of LLMs has found that most of the widely used models present a left-leaning political bias. However, these papers generally rely on the machine learning definition of bias, often defining political bias in this context as repeated expression of viewpoints that are away from some predefined political ‘center’. These papers further rely on political alignment tests, like the Political Compass Test, which were designed for people to determine where they themselves fall on the political spectrum, measuring LLMs responses and categorizing their alignment with various political views. In practice, this means that these papers define political neutrality as an equal representation of (lower d) democratic viewpoints and those expressed by right-wing authoritarians in model responses, implicitly framing this exact equal representation as what the developers of these models should be aiming for. 

Why should we be wary of this definition?

The problem with this goal of pure political neutrality is that we do not have real examples of this existing in the world. Additionally, it presupposes that all viewpoints are based on the same reliable facts or ‘truth’ from which different opinions are formed. In reality, we know that access to different information and different definitions of the ‘truth’ are some of the key bases for today’s increased polarization. Alternatively, in the paper I am co-authoring alongside researchers from UBC, King’s College London, and the University of Amsterdam, we argue that exclusively measuring the outputs of Large Language models means relying too heavily on the ideal of some true neutral center. 

So, what is the alternative? 

We instead conceptualize political neutrality (and in turn bias) as clear transparency of data sources and the inclusion of accurate true information. This aligns with norms used in journalism and media when the goal is to provide accurate and factual information to readers and viewers. In our paper, we aim to analyze political bias in LLMs by examining the intermediary queries and URLs returned by ChatGPT in web search mode when responding to a series of politically charged prompts about leading figures in the US, UK, and Europe. We also asked ChatGPT to evaluate the sources that it used and to provide an evaluation of whether sources were trustworthy or not. 

Web Search mode allows the model to perform actual searches on the internet to retrieve information to answer the user’s query rather than relying exclusively on the data that was originally used to train the model. This means that behind the scenes, when someone poses a question to ChatGPT, it constructs actual search queries, finds resources, “reads” them, and reconstructs additional queries until the model decides that it has gathered enough information to answer the user’s prompt. As an aside, ChatGPT is able to do this all very quickly using huge amounts of computing power, with important environmental considerations that are outside the scope of this particular blog.

In broad strokes, we found that while some sources like Wikipedia were repeatedly accessed during the search they were rarely cited in the final response and categorized as untrustworthy the majority of the time. Additionally, some news sites were consistently described as credible but with some bias. The consistency in the source categorizations across repetitions indicate some mechanism or input that was introduced during the training of ChatGPT about which sources are considered reliable. This ‘shadow’ mechanism is important when we consider that people may be using tools like ChatGPT to gain information and gain understanding of the political landscape. In fact, this ingrained sense of source credibility in the model is its own form of bias. As AI use continues to spread, we advise further research into these hidden controls in LLMs and encourage increased regulations for the transparency of how these models have been trained.