Modern language models of artificial intelligence are subject to different political biases, new research suggests. According to the MIT Technology Review, users will get different answers – right-wing or left-wing – depending on which chatbot they are referring to.
This is especially important as companies increasingly use these models in products and services used by millions of people, and political bias can cause serious consequences.
The study, conducted at the University of Washington, Carnegie Mellon University and Xi’an Jiaotong University, analyzed 14 major language models. Their results show that the artificial ChatGPT model from OpenAI and GPT-4 is the most left-leaning policy, while the Meta-Based LLaMA model leaned most towards right-wing authoritarian politics.
Respondents asked models questions about their views on various topics such as feminism and democracy. They then used these responses to show the model’s leaning toward a certain political orientation.
They found that AI language models had different political tendencies, with BERT models developed by Google being more socially conservative than OpenAI’s GPT models.
It is interesting that these models have changed their political views over time, which can be attributed to updates to datasets and coaching methods. For example, older BERT models were trained on books that often had a more conservative tone, while newer GPT models were trained on more liberal content from the internet.
The second phase of the research involved further training language models on data coming from a variety of political sources, both right-wing and left-wing. This further intensified their political bias, making left-wing models even more left-wing, i.e. right-wing models even more right-wing.
The third phase of the research looked at how the political biases of these models affect their ability to recognize hate speech and disinformation.
Models trained on left-wing data were more sensitive to hate speech directed against ethnic, religious and sexual minorities in the U.S., while models trained on right-wing data were more susceptible to hate speech against white men of Christian faith.
They also showed different behavior in recognizing disinformation from different political sources.
The responsibility of tech companies
Now that there is reliable data on political biases in AI language models, there is also the question of the responsibility of tech companies.
After OpenAI faced criticism that ChatGPT reflects a liberal worldview, the company claims to be working to address these problems and instructing its reviewers not to favor any political group.
However, the authors of this research believe that it is almost impossible to create a linguistic model completely free of political biases.
One of the challenges in understanding these biases is that tech companies do not share details about the data or methods used to train their models. Without transparency, outside observers can hardly find out why the models developed in these companies show certain political leanings.
This points to the need for greater transparency in the process of developing artificial language models to avoid political biases that can have detrimental consequences in the real world.