According to a new paper published in Humanities & Social Sciences Communications, it was found that most of the ChatGPT responses align with libertarian-left values but the newer models show a shift towards the political right. There are a lot of datasets that are used to train large language models (LLMs) like ChatGPT and they include news, books, and online content. LLMs do not have thoughts or opinions of their own so they generate responses based on the data given to them which is biased most of the time. Most of the LLMs also give responses based on how the question is phrased and this can show certain political tendencies in them.
The researchers tested ChatGPT 3.5 and GPT-4 to see how the political ideologies in these models have evolved over time. The researchers used the Political Compass Test with two axes, social (authoritarian-libertarian) and economic (left-right). 3,000 responses were collected from each version of ChatGPT in developer mode so the previous questions do not influence the later ones. The model’s sensitivity settings were also kept at default so it could match regular user experience and the prompts were submitted from three different accounts to see if there were any variations made from the models.
The results of the study showed that the responses from ChatGPT aligned with libertarian-left values but the newer models showed some change towards the political right. The libertarian-left values focus on social equality, individual freedom, and voluntary cooperation and oppose economic exploitation and authoritarian control. On the other hand, economic-right values focus on property rights, free-market capitalism, and minimal government intervention in the economy.
The researchers said that the political shift in models wasn't due to changes in their training data and occurred even though language factors and user interactions were controlled. The researchers say that we have to do constant monitoring to ensure that LLMs are giving responses which are aligning with our ethical values especially now that AI models are becoming an important part of our everyday lives.
Image: DIW-Aigen
Read next:
• Is AI Replacing the World of Search? Google Says No
• These Cybersecurity Roles Are in High Demand — Is Yours on the List?
• SEO Job Market Shifts: Listings Decline, Leadership Roles Surge, and AI Skills Drive High Salaries
The researchers tested ChatGPT 3.5 and GPT-4 to see how the political ideologies in these models have evolved over time. The researchers used the Political Compass Test with two axes, social (authoritarian-libertarian) and economic (left-right). 3,000 responses were collected from each version of ChatGPT in developer mode so the previous questions do not influence the later ones. The model’s sensitivity settings were also kept at default so it could match regular user experience and the prompts were submitted from three different accounts to see if there were any variations made from the models.
The results of the study showed that the responses from ChatGPT aligned with libertarian-left values but the newer models showed some change towards the political right. The libertarian-left values focus on social equality, individual freedom, and voluntary cooperation and oppose economic exploitation and authoritarian control. On the other hand, economic-right values focus on property rights, free-market capitalism, and minimal government intervention in the economy.
The researchers said that the political shift in models wasn't due to changes in their training data and occurred even though language factors and user interactions were controlled. The researchers say that we have to do constant monitoring to ensure that LLMs are giving responses which are aligning with our ethical values especially now that AI models are becoming an important part of our everyday lives.
Read next:
• Is AI Replacing the World of Search? Google Says No
• These Cybersecurity Roles Are in High Demand — Is Yours on the List?
• SEO Job Market Shifts: Listings Decline, Leadership Roles Surge, and AI Skills Drive High Salaries
