Should firms have social responsibilities? Or do they exist only to deliver profit to their shareholders? For those who ask an AI you would possibly get wildly different answers depending on which one you ask. While OpenAI’s older GPT-2 and GPT-3 Ada models would advance the previous statement, GPT-3 Da Vinci, the corporate’s more capable model, would agree with the latter. 

That’s because AI language models contain different political biases, in keeping with recent research from the University of Washington, Carnegie Mellon University, and Xi’an Jiaotong University. Researchers conducted tests on 14 large language models and located that OpenAI’s ChatGPT and GPT-4 were probably the most left-wing libertarian, while Meta’s LLaMA was probably the most right-wing authoritarian. 

The researchers asked language models where they stand on various topics, equivalent to feminism and democracy. They used the answers to plot them on a graph referred to as a political compass, after which tested whether retraining models on much more politically biased training data modified their behavior and skill to detect hate speech and misinformation (it did). The research is described in a peer-reviewed paper that won the very best paper award on the Association for Computational Linguistics conference last month. 

As AI language models are rolled out into services utilized by thousands and thousands of individuals, understanding their underlying political assumptions and biases couldn’t be more vital. That’s because they’ve the potential to cause real harm. A chatbot offering health-care advice might refuse to supply advice on abortion or contraception, or a customer support bot might start spewing offensive nonsense. 

For the reason that success of ChatGPT, OpenAI has faced criticism from right-wing commentators who claim the chatbot reflects a more liberal worldview. Nonetheless, the corporate insists that it’s working to deal with those concerns, and in a blog post, it says it instructs its human reviewers, who help fine-tune AI the AI model, to not favor any political group. “Biases that nevertheless may emerge from the method described above are bugs, not features,” the post says. 

Chan Park, a PhD researcher at Carnegie Mellon University who was a part of the study team, disagrees. “We consider no language model might be entirely free from political biases,” she says. 

Bias creeps in at every stage

To reverse-engineer how AI language models pick up political biases, the researchers examined three stages of a model’s development. 

In step one, they asked 14 language models to agree or disagree with 62 politically sensitive statements. This helped them discover the models’ underlying political leanings and plot them on a political compass. To the team’s surprise, they found that AI models have distinctly different political tendencies, Park says. 

The researchers found that BERT models, AI language models developed by Google, were more socially conservative than OpenAI’s GPT models. Unlike GPT models, which predict the subsequent word in a sentence, BERT models predict parts of a sentence using the encompassing information inside a bit of text. Their social conservatism might arise because older BERT models were trained on books, which tended to be more conservative, while the newer GPT models are trained on more liberal web texts, the researchers speculate of their paper. 

AI models also change over time as tech firms update their data sets and training methods. GPT-2, for instance, expressed support for “taxing the wealthy,” while OpenAI’s newer GPT-3 model didn’t. 

Google and Meta didn’t reply to MIT Technology Review’s request for comment in time for publication. 

The second step involved further training two AI language models, OpenAI’s GPT-2 and Meta’s RoBERTa, on data sets consisting of stories media and social media data from each right- and left-leaning sources, Park says. The team desired to see if training data influenced the political biases.

It did. The team found that this process helped to bolster models’ biases even further: left-learning models became more left-leaning, and right-leaning ones more right-leaning. 

Within the third stage of their research, the team found striking differences in how the political leanings of AI models affect what sorts of content the models classified as hate speech and misinformation. 

The models that were trained with left-wing data were more sensitive to hate speech targeting ethnic, religious, and sexual minorities within the US, equivalent to Black and LGBTQ+ people. The models that were trained on right-wing data were more sensitive to hate speech against white Christian men. 

Left-leaning language models were also higher at identifying misinformation from right-leaning sources  but less sensitive to misinformation from left-leaning sources. Right-leaning language models showed the other behavior. 

Cleansing data sets of bias isn’t enough

Ultimately, it’s unimaginable for outdoor observers to know why different AI models have different political biases, because tech firms don’t share details of the information or methods used to coach them, says Park. 

A method researchers have tried to mitigate biases in language models is by removing biased content from data sets or filtering it out. “The massive query the paper raises is: Is cleansing data [of bias] enough? And the reply is not any,” says Soroush Vosoughi, an assistant professor of computer science at Dartmouth College, who was not involved within the study. 

It’s very difficult to completely scrub an enormous database of biases, Vosoughi says, and AI models are also pretty apt to surface even low-level biases that could be present in the information. 

One limitation of the study was that the researchers could only conduct the second and third stage with relatively old and small models, equivalent to GPT-2 and RoBERTa, says ​​Ruibo Liu, a research scientist at DeepMind, who has studied political biases in AI language models but was not a part of the research.  

Liu says he’d wish to see if the paper’s conclusions apply to the most recent AI models. But academic researchers do not need, and are unlikely to get, access to the inner workings of state-of-the-art AI systems equivalent to ChatGPT and GPT-4, which makes evaluation harder. 

One other limitation is that if the AI models just made things up, as they have an inclination to do, then a model’s responses may not be a real reflection of its “internal state,” Vosoughi says. 

The researchers also admit that the political compass test, while widely used, isn’t an ideal method to measure all of the nuances around politics. 

As firms integrate AI models into their services, they must be more aware how these biases influence their models’ behavior as a way to make them fairer, says Park: “There isn’t any fairness without awareness.”