We need to direct attention to the shift to foundation model-based infrastructure and its role in facilitating the algorithmisation of social ordering, argues Amanda Turnbull.
We need to direct attention to the shift to foundation model-based infrastructure and its role in facilitating the algorithmisation of social ordering, argues Amanda Turnbull.
Opinion
OPINION
Since its launch last November, ChatGPT has piqued our collective curiosity. We have read about its burgeoning benefits, but we have also been alerted to myriad worries that it provokes.
The large language model chatbot developed by OpenAI may be used as an educational resource, it can help conductsome aspects of research, it can write stories and poetry, and it has potential for making improvements in the healthcare setting. However, at the same time, ChatGPT creates ever-growing concern over issues such as plagiarism, copyright infringement, privacy and data security concerns, and unemployment apprehension.
The most recent headlines reveal that ChatGPT has now learned to talk and that it has developed image recognition abilities.
This means we can engage in conversations with ChatGPT. Unlike older generation AI voice assistants like Alexa and Siri, ChatGPT’s voice has intonation. In fact, you can choose from its five realistic, yet synthetic, voices.
Further, we can also upload images to ChatGPT and ask it questions about those images. For instance, you could upload a picture of your friend’s delicious dinner and obtain the ingredients and recipe required to make it.
Notably, however, these new updates are only available on OpenAI’s premium app, ChatGPT Plus, which requires a monthly subscription.
These new ways of interacting with ChatGTP are the result of combining models — bringing together a computer vision model, plus a large language model, plus speech recognition technology. While these updates are novel, they also have the makings for further unease: the compounding of problems already associated with ChatGPT.
Along with the obstacles posed by ChatGPT that we already know about — plagiarism, copyright infringement etc — computer vision has well-documented drawbacks when it comes to understanding the context of an image or video.
Recognising objects and people is highly complex and models may make biased judgments if the training data used contains biased information.
This is particularly problematic since it may result in gender and racial bias. Computer vision also has difficulties with real-time processing, and with explainability — understanding how models got to particular decisions.
Speech recognition setbacks include differentiating between varying dialects and accents, with peripheral background noise, and with data storage reliability.
There is also the matter of voice fraud. If a person’s voiceprint is compromised, it may result in identity theft. All these possible predicaments add up. Moreover, these compounded issues may impact both ChatGPT users and people who are not even using the product. Bias, privacy, and security concerns affect us all.
Amanda Turnbull, of Waikato University
In addition to this compounding of problems, the recent rapid rise of generative AI — the umbrella term that refers to models like ChatGPT that can create new content based on data provided to it — represents a palpable paradigm shift in society.
Our ways of doing things are fundamentally changing.
Generative AI applications like ChatGPT, DALL-E, Bard and others are becoming foundation models for many other AI-based systems. In other words, generative AI is replacing task-specific models of AI, and it is being used as the substructure or framework for many other applications.
This shift to foundation model-based infrastructure is a plus for start-ups since the use of this infrastructure would mean that they do not have to build models from scratch.
They can instead focus on final phases of AI development.
However, any problems in the foundation model may be inherited and adapted by subsequent applications and spread widely. This creates the potential for continuous compounding of the issues identified with the latest upgrades to ChatGPT. Continuous compounding may be infinite. Recognising this is of quintessential importance with the emergence of infrastructure markets.
Put succinctly, we need to direct attention to this shift to foundation model-based infrastructure and its role in facilitating the algorithmisation of social ordering. For without doing so and with hyperfocus on ChatGPT, we risk missing the forest for the tree.
Amanda Turnbull is a lecturer in cyber law at Te Piringa Faculty of Law, University of Waikato (Tauranga campus). Her research focuses on the legal and philosophical challenges posed by creative AI technologies.
