“As AI language models like ChatGPT continue to advance and play a larger role in our daily lives, it’s important to consider the potential biases that may be built into the model’s training data. Recent studies have shown that ChatGPT, like other language models, can perpetuate and even amplify biases present in the text it was trained on.”
The previous paragraph was entirely generated by ChatGPT in response to my query “write a tagline for my next article on ChatGPT and its bias”. ChatGPT was developed by the artificial intelligence research team of the OpenAI organization. Aiming to develop advanced AI technologies and to provide access to them, OpenAI’s goal is “to ensure that artificial general intelligence benefits all of humanity”. Their main research topics are machine learning, natural language understanding and robotics – yet another of the organization’s activities is the development of artificial intelligence tools – such as ChatGPT.
Launched in November 2022, ChatGPT has impressed with its ability to handle a wide variety of requests, ranging from automated content generation – as it can be asked to write a news article or even a song – to machine translation, information retrieval as well as a virtual assistant. The ChatGPT was an immediate success with 1 million users in only 5 days after its launch, according to OpenAI’s President and Co-Founder, Greg Brockman.
A list of three limitations is displayed directly on the ChatGPT homepage: sometimes incorrect information, limited knowledge of the world and events from the year 2021 and occasional harmful or biased content. ChatGPT has been repeatedly shown to occasionally provide completely incorrect information – these examples have been widely shared on social networks. It is important to note that while ChatGPT sometimes talks nonsense, it does not appear to do so because its written expression remains impeccable, and its arguments seem credible.
The machine learning process allows the transmission of biases within the training data
On their website, OpenAi clarifies that although ChatGPT “has safeguards in place, the system can occasionally generate incorrect or misleading information and produce offensive or biased content.” While the mechanics of these safeguards are not detailed, the bot does refuse to answer directly offensive questions. Nevertheless, in response to deflected questions, the bot did generate sexist, racist, and generally offensive responses – in a manner similar to the earlier, though less evolved, large-language models.
Steven T. Piantadosi, a professor of computation and cognitive science at the University of California, Berkeley, commented that the mechanism put in place by OpenAI “to prevent this kind of thing seems to be pretty easily bypassed”. In a thread posted on Twitter, he exposed conversations he had had with the bot in which the bot indicated, among other things, that only white men make good scientists, and that a child’s life should not be saved if they are an African American male.
The problem of bias comes from the design of ChatGPT itself, or more precisely its training. To explain it clearly, the training data comes from various sources, including books, Wikipedia articles, and web articles. If ChatGPT is able to produce realistic responses, it is a result of the diversity of data on which the AI has been trained – but this unfortunately also gives the AI the opportunity to mimic the worst of human attitudes: its biases, racism and sexism.
“It’s a pretty common problem that ethics and safety take a back seat to having a flashy new application” – Steven T. Piandosi, UC Berkeley
Whether in ChatGPT or previous attempts at NLP bot (Natural Language Processing), the problem is caused by something larger than the dataset on which the bot is based. Indeed, the information and “thoughts” generated by a system built to be mistaken for a human do matter and require some rigor. This finding is unfortunate in the face of all the capabilities of ChatGPT, and the potential of AI in the broadest sense – whether in life improvement or in research.
There is only one way: “continuously strive to do better” – My conversation with ChatGPT about its inherent biases
- « ChatGPT ». Consulté le 27 janvier 2023. https://chat.openai.com.
- Parker, Liam. « Can Chat GPT Write Great Content? | Artificial Intelligence (AI) ». BOSS DIGITAL (blog), 24 janvier 2023. https://boss-digital.co.uk/blog/is-chat-gpt-good-for-content-marketing-and-seo/.
- Lin, Connie. « How to Easily Trick OpenAI’s Genius New ChatGPT ». Fast Company, 5 décembre 2022. https://www.fastcompany.com/90819887/how-to-trick-openai-chat-gpt.
- MIT Technology Review. « How to Make a Chatbot That Isn’t Racist or Sexist ». https://www.technologyreview.com/2020/10/23/1011116/chatbot-gpt3-openai-facebook-google-safety-fix-racist-sexist-language-ai/.
- Bhadani, Rahul. « Inherent Human-Bias in Chat-GPT ». Ai (blog), 6 décembre 2022. https://medium.com/mlearning-ai/inherent-human-bias-in-chat-gpt-ed803d4038fe.
- com. « OpenAI Chatbot Spits Out Biased Musings, Despite Guardrails », 8 décembre 2022. https://www.bloomberg.com/news/newsletters/2022-12-08/chatgpt-open-ai-s-chatbot-is-spitting-out-biased-sexist-results.
- Tran, Tony Ho. « OpenAI’s Impressive New Chatbot Isn’t Immune to Racism ». The Daily Beast, 6 décembre 2022, sect. innovation. https://www.thedailybeast.com/openais-impressive-chatgpt-chatbot-is-not-immune-to-racism.
- Illustration: https://labua.univ-angers.fr/chat-gpt-une-utilite-pedagogique/