Image context: The existing GPTOnline editorial cover illustrates this historical study report. It is not a test result, a current model assessment or evidence that the study measured an election outcome.
Historical research context: This article concerns a 2023 study, not an evaluation of current ChatGPT models. The university’s contemporaneous release identifies ChatGPT version 3.5 tested in spring 2023. The paper appeared online in August 2023 and later in the journal’s 2024 volume. Its results should not be generalized to every model, prompt or political context.
A study published online on August 17, 2023 conducted by researchers affiliated with the University of East Anglia and institutions in Brazil examined potential political bias in OpenAI’s ChatGPT responses. The findings illustrated the ongoing challenges faced by artificial intelligence companies in controlling the behavior of their bots, as conversational AI became more widely used.
The researchers compared default ChatGPT answers with answers produced when it was prompted to impersonate different political viewpoints. Using Political Compass questions and repeated trials, they reported alignment with Democrats in the U.S., Lula in Brazil and the Labour Party in the U.K. These were findings about the tested responses and methodology, not evidence that the software held personal political beliefs.
These findings raised concerns that AI chatbots like ChatGPT can reproduce assumptions, beliefs, and stereotypes associated with their training data and development. The study did not establish the precise cause of its observed differences.
In the article’s 2023 context, concern about the approaching 2024 U.S. presidential election accompanied wider use of chatbots beyond simple assistance, aiding in summarizing documents, answering queries, and even contributing to professional and personal writing tasks. With platforms like Google employing chatbots to directly address search inquiries and political campaigns utilizing them for generating content, the potential impact on public perception and even election outcomes raised concerns. The experiment did not measure actual voting behavior or establish an effect on an election outcome.
While the tested ChatGPT described itself as lacking political opinions or biases, the study’s lead author, Fabio Motoki, warned about biased outputs. His concern was that such outputs could erode public trust and potentially exert unintended influence on political processes.
This historical article should not be read as a current statement about whether Meta, Google or OpenAI has commented. No new response was sought for this revision. A model’s statement that it is impartial is not, by itself, an independent test of its outputs.
This study underscores the intricate challenge of mitigating biases in AI models, especially in the age of generative chatbots. The chatbots discussed in 2023, including ChatGPT, Google’s then-named Bard and Microsoft’s Bing, were built upon vast language models that extensively process data from the internet. Such systems can reproduce the biases and viewpoints prevalent online.
Moreover, these chatbots had become focal points in discussions encompassing politics, technology, and social media. Even ChatGPT faced allegations of bias shortly after its launch, with accusations of favoring liberal perspectives. This brought to light the broader conversation about how technology and social media platforms can influence political outcomes and societal polarization.
The concern was that, as chatbots became more ingrained in online interactions, their outputs could further amplify the existing polarization observed in society. The information landscape, increasingly shaped by bot-generated content, might contribute to a feedback loop of polarization.
Efforts to detect and mitigate bias remain an important research topic. The study supplied a method for examining model responses; it did not demonstrate that adding a neutral-language replacement layer would eliminate bias. Questionnaire design and prompting also affect what such tests can establish.
In conclusion, the study sheds light on the intricacies of AI biases and their potential impact on political discourse and decision-making. The research serves as a reminder that while AI technology offers exciting possibilities, it also carries significant responsibilities in managing unintended consequences. As AI continues to evolve, finding a balance between technological advancement and ethical considerations will remain a crucial challenge.
