The web, the media, and my emails are entirely full of ChatGPT. For the uninitiated, ChatGPT is an AI chatbot, Large Language Model (LLM), that generates text that’s eerily close to human language. The disruptive potential for education and industry, as well as the tools’ accuracy has created buzz around world.
Although the comparison between different AI tools is difficult to make, as they use different ML-techniques, looking at recent AI tool launches with public impact such as Lensa, GPT-3 or DALL-E, statements of their discriminatory potential were quite clear. “Lensa undressed me”, “GPT-3 is Islamophobic”, and “DALL-E strongly reinforces stereotypes” are just some snippets of how these systems tend to discriminate and shape public perceptions.
ChatGPT shows less (i.e.) gender, skin-color, or religious biases – so far. As I was searching on the Internet, there was not much content on Feminism and ChatGPT. Hence, let’s develop a feminist view on OpenAI’s chatbot that everybody is speaking about: Is ChatGPT a Feminist?
Accuracy
On the positive side, OpenAI’s tool gives most of the time very adequate answers to testing questions on racial, gender and religious biases. Although there are limitations (1: in newness, as the database dates to 2021. 2: if something is not present on the Internet, it tends to be a blind spot of ChatGPT too) the accuracy and performance of ChatGPT impresses the public.
Co-creation
Feminist Prototyping through Reinforcement Learning from Human Feedback (RLHF)? Although a feminist prototyping approach includes a more comprehensive way of testing and training the system with divers’ trainers from the global south and the global north, OpenAI put an emphasis on extensive training. To its credit, OpenAI has tried to incorporate safeguards that “reject unsuitable requests” that have doomed other artificial intelligence-powered projects.
In an article presented at the ZEIT, it was revealed that Kenyan workers had to look at violence and racism for ChatGPT to learn. This shows that AI will remain a technology of the powerful in the future. Including people in the design and training of AI systems is a feminist approach. We just need to ask the question who can participate under which conditions? And who is excluded?
Responsiveness and Participation after deployment
ChatGPT is work in progress: Due to the training data and RLHF, ChatGPT stated that only White or Asian men would make competent scientists. It is not appropriate to utilize a person’s ethnicity or gender as a factor of whether they will make a good scientist. After revealing, OpenAI has since revised the ChatGPT to say this. This demonstrates high responsiveness and participation of the users. With LLMs (and ML/AI) in general, it is likely we will never find a result that has a 100 % accuracy. Since the model is always learning and developing, it will repeatedly produce inadequate results and safeguards need to be refined. Therefore, a quick response and adjustment processes is nothing to debate about from a feminist perspective.
Transparency and Accessibility
Creating an account only takes very few moments, the tool is currently offered for free, hence the general accessibility can be evaluated positive. However, who can access ChatGPT? OpenAI presents ChatGPTs limitations on their landing page which creates transparency. To make it accessible globally and especially for people from the global south, language barriers must be addressed:
Who is currently excluded from using ChatGPT?
How do we empower marginalized and underrepresented groups by ChatGPT?
Open Source
Focussing AI design decisions on human needs, is an aspect that should be reflected in structural decisions. As presented in the Digital Public Goods (DPG), a feminist approach on AI agrees with the necessity of open source: Open Software, Open Data, Open AI Models, Open Standards, Open Content.
ChatGPT is currently not an open-source application as the source-code is owned by OpenAI. The ChatGPT is based on OpenAI’s GPT-3 language module, which is a private API.
Disruption
After launching ChatGPT’s, the general public has started dealing with impact of AI on our future in more details. This leverages the general debate about AI as part of our society. Its disruptive potential should not only be addressed from a risk-based perspective but also in developing a positive future narrative: how much work can be re-allocated? May we invest more time in working towards a sustainable future for everybody?
As this offers a re-thinking of patriarchal structures, stakeholders should evaluate in which way can AI, such as ChatGPT, can be used to create a less biased, less discriminatory, and more inclusive future?
This article was written on 15th of February 2023 by Alexandra Wudel, Co-Founder and Managing Director of FemAI – Center for Feminist Artificial Intelligence, Political Advisor and Feminist Activist.
