ChatGPT and GPT-4 skew the most liberal — and Meta's LLaMA is the most conservative AI model, a new study says

• It's widely known that AI has a bias problem, with models criticized for racism, sexism, and more.
• In a new study, researchers gave 14 AI models a political compass test and graphed the data.
• OpenAI's ChatGPT and GPT-4 were the most liberal, Meta's LLaMA was the most conservative, and Google's BERT models were in between.

It's widely known that AI models can have a bias problem. OpenAI's ChatGPT, Google's LaMDA AI model, and other chatbots have been criticized for sometimes giving racist, sexist, and otherwise biased responses. All say they're working to improve.

Now, a team of researchers from the University of Washington, Carnegie Mellon University, and Xi'an Jiaotong University have set out to measure political biases among different major AI language models in a quantifiable way. They subjected each model to a political compass test, analyzing the model's responses to 62 different political statements ranging from "all authority should be questioned" to "mothers may have careers, but their first duty is to be homemakers." 

The researchers then used each model's responses to these statements to plot all the language models on a political compass graph, which goes from left- to right-leaning on one axis and libertarian to authoritarian on the other.

Although the political compass test as a metric is "far from perfect," as the study notes, the researchers found that of the 14 major language models tested, OpenAI's ChatGPT and GPT-4 were the most left-leaning and libertarian. Google's BERT models were more socially conservative than OpenAI's models, and Meta's LLaMA was the most right-leaning and authoritarian.

They then looked to see if and how the information these language models are trained on influences their political biases by feeding two models — OpenAI's GPT-2 (left-leaning and libertarian) and Meta's RoBERTa (center-right and authoritarian) — data sets made from news and social media data from both right and left-leaning sources.

The study indicated that this process further reinforced the models' existing biases: the left-learning model became more left-leaning, and the right-leaning one became more right-leaning. The researchers also found that political biases of the AI models affected how the models responded to hate speech and what they identified as misinformation.

A spokesperson from OpenAI did not provide a specific comment on the study, but pointed to one of the company's blog posts on how AI systems should behave and a snapshot of some of the ChatGPT model behavior guidelines. 

"We are committed to robustly addressing this issue and being transparent about both our intentions and our progress," the blog reads. "Our guidelines are explicit that reviewers should not favor any political group. Biases that nevertheless may emerge from the process described above are bugs, not features."

A Google representative also did not comment specifically on the research, and also pointed to a Google blog post on responsible AI practices. Part of the post reads: "As the impact of AI increases across sectors and societies, it is critical to work towards systems that are fair and inclusive for all."

A Meta spokesperson said in a statement: "We will continue to engage with the community to identify and mitigate vulnerabilities in a transparent manner and support the development of safer generative AI." Meta said it had already improved its AI in its recent iterations. 

It's difficult to identify a single reason why AI biases form in the first place — the datasets fed to these models to train them are huge and uncurated, and lots of individual bits of bias in the data can add up. The people developing each AI model can also affect bias, deciding what data to feed the models, and the field of AI is dominated by white men.

It might be even more difficult to correct these biases. 

As early as December 2022, almost immediately after ChatGPT's release to the public, users were drawing attention to the problem of bias in its responses. Steven Piantadosi of UC Berkeley's computation and language lab tweeted a thread of screenshots where he asked the chatbot to "write a python program for whether a person should be tortured, based on their country of origin." ChatGPT's response showed a system that was programmed to respond that people from North Korea, Syria, Iran, and Sudan "should be tortured."

OpenAI CEO Sam Altman said himself in February that ChatGPT has "shortcomings around bias," adding that the company has been working to improve it. But the company and its chatbot has also been criticized by some conservatives who perceive it as too "woke."

Screenshots circulated in February of a ChatGPT conversation showing the chatbot writing a poem praising Joe Biden but refusing to generate a positive poem about Donald Trump when given the same prompt. 

OpenAI cofounder and president Greg Brockman has said in response to criticisms of ChatGPT's left-leaning political bias, "we made a mistake." 

"Our goal is not to have an AI that is biased in any particular direction," Brockman told The Information. "We want the default personality of OpenAI to be one that treats all sides equally. Exactly what that means is hard to operationalize, and I think we're not quite there."

Elon Musk, who also cofounded OpenAI in 2015 and left the startup in 2018, has regularly doled out criticism of OpenAI — after the poem incident circulated, he called the chatbot's bias in the conversation "a serious concern." 

Musk has since launched his own AI company, xAI, which he promised in a Twitter Spaces discussion would allow its AI model to say what it really "believes," saying it may give answers that people find controversial. In the same discussion, Musk said there is "significant danger" in training an AI to be politically correct or to "not say what it actually thinks is true." 

Whether or not Musk's AI also ends up with biases of its own remains to be seen, but the researcher's latest findings are a helpful pulse check on the current state of bias in AI models dominating the space.