Introduction

Summary

Opening: Crises of Values and Governance in Artificial Intelligence

The twenty-first century has witnessed the rise of ‘big data’, and the widespread deployment of Artificial intelligence (AI). At the turn of the century, according to Wooldridge, ‘AI was a rather niche area with a somewhat questionable academic reputation’ (2020: 5). But in 2014, Google bought a ‘tiny AI company [DeepMind] for a huge sum […] Artificial intelligence was suddenly big news – and big business’ (2020: 167). What happened?

Contemporary AI uses large data sets – which may or may not be representative of the contexts in which the AI trained will be deployed (see below) – to train models using machine learning (ML) techniques, and specifically, ‘deep learning’ (LeCun et al. 2015). The information processing that results is often opaque: precisely what mapping from inputs to outputs the model supports is unclear, even to the engineers who develop the systems. The AI systems that result are often ‘narrow’, dedicated to solving a relatively small set of problems. As a result, some might not be inclined to call this true AI – much of it certainly doesn't meet the requirements of Searle's (1980) ‘strong AI’, actually possessing the various mental states involved in sapience. But it is also insufficiently general to even simulate much of the variety of human intelligence, and so it fails the requirements for his ‘weak’ AI too. Nevertheless, for better or worse, the term ‘AI’ is used in connection with these information processing systems.

Recently, we have seen the emergence of what is known as ‘generative’ AI – AI that produces outputs of the same kind as its training data. For example, whereas image classifiers take visual images as inputs and produce labels (i.e. words) as outputs (see e.g. Krizhevsky et al. 2012), certain generative systems (such as generative adversarial networks (GANs) and text-to-image models) are able to generate (often novel) images as outputs. Similarly, large language models (LLMs) are trained on vast corpuses of textual data, and are able to generate grammatical, and even meaningful, text (e.g. semantically appropriate answers to questions).