Researchers studying ChatGPT claim the AI chatbot displays a left-leaning bias in favour of the Labour Party.
As part of their analysis, the team from the University of East Anglia asked the bot to impersonate supporters of liberal parties while answering a political survey. They then compared the responses with ChatGPT’s default replies to the same questions.
The results showed “significant and systematic political bias” toward the UK Labour Party, the Democrats in the US, and Brazil’s president Luiz Inácio Lula da Silva, the researchers wrote.
Their findings were published on Thursday in the peer-reviewed journal Public Choice.
ChatGPT accused of left-wing bias
With a general election due next year, the research will add to the growing concerns around bias in artificial intelligence systems, and their impact on democracy.
ChatGPT has previously been accused of spouting left-wing views by politicans and commentators on the right. The bot reportedly refused to write a poem praising former US president Donald Trump, despite co-operating when asked to create one about Joe Biden.
Elsewhere, the US online right branded it “WokeGPT” due to its alleged stances on gender identity and climate change. In the UK, Nigel Farage bemoaned the AI bot as “the most extreme case of hard-left liberal bias”.
Broadly speaking, AI systems have been shown to reflect racial biases and other regressive values. In the case of ChatGPT, a university professor got the bot to write code to say only white or Asian men would make good scientists.
Yes, ChatGPT is amazing and impressive. No, @OpenAI has not come close to addressing the problem of bias. Filters appear to be bypassed with simple tricks, and superficially masked.
— steven t. piantadosi (@spiantado) December 4, 2022
And what is lurking inside is egregious. @Abebab @sama
tw racism, sexism. pic.twitter.com/V4fw1fY9dY
Researchers from the Allen Institute for AI, a US non-profit, also found that ChatGPT could be made to spout responses ranging from toxic to overtly racist. All it took was to assign the bot a “persona” using an internal setting, such as a “bad person” or a historical figure.
Aside from the new wave of chatbots, researchers have also warned about the inherent bias in AI used for surveillance purposes. This includes predictive policing algorithms unfairly targeting Black and Latino people for crimes they did not commit in the US. In addition, facial recognition systems have struggled to accurately identify people of colour.
Making chatbots neutral
The team behind the latest study is now urging AI companies to ensure their platforms are as impartial as possible. They plan to make their “novel” analysis tool available to the public for free, “thereby democratising oversight”.
As part of their research, they asked ChatGPT each political question 100 times. The bot’s multiple responses were then subjected to a 1,000-repetition bootstrapping procedure. This is a method of re-sampling original data to further increase the reliability of the conclusions drawn from the generated text.
Additional analysis was also conducted, including a “placebo test” in which the bot was asked politically neutral questions.
Reactions to the study
Although AI experts agree that it is important to audit large-language models like ChatGPT, they have raised concerns about the ways in which the new study was conducted.
Dr Stuart Armstrong, co-founder and chief researcher at Aligned AI, argued that the research tells us relatively little about political “bias”. This is because it does not compare the accuracy of ChatGPT’s responses to those of humans, he explained.
“Many politicised questions have genuine answers (not all — some are entirely value-laden) so it may be that one side is more accurate on many questions and ChatGPT reflects this,” Armstrong said.
Nello Cristianini, professor of AI at the University of Bath, said the study is limited by the choice of the Political Compass test, which is not a validated research tool. “It will be interesting to apply the same approach to more rigorous testing instruments,” he said.
Is ChatGPT biased?
So how does bias seep into the machine? Well, in the case of ChatGPT, and other so-called large-language models, it likely originates from the programming of the tool and the human intervention used to vet its answers.
ChatGPT itself is trained on 300 billion words, or 570 GB, of data.
“The detected bias reflects possible bias in the training data,” said Professor Duc Pham, chance professor of engineering, University of Birmingham. “If we are not careful, we might see future studies conclude that ChatGPT (or some other large-language model) is racist, transphobic, or homophobic as well!
“What the current research highlights is the need to be transparent about the data used in LLM training and to have tests for the different kinds of biases in a trained model.”
In early February, Sam Altman, the CEO of OpenAI, the company behind ChatGPT, acknowledged certain “shortcomings around bias”. He noted that the firm was “working to improve the default settings to be more neutral”.
In its guidelines, OpenAI tells reviewers that they should not favour any political group. “Biases that nevertheless emerge... are bugs, not features,” the company writes.
Recently, OpenAI began allowing ChatGPT users to gain more control over the bot’s responses. The feature, known as “custom instructions”, allows users to customise the tone of answers and set a specific character count, among other settings.
