Suman Gupta and Peter H. Tu, The Practical Philosophy of AI-Assistants: An Engineering-Humanities Conversation (World Scientific, 2024) 280 pp. $30. Pb. ISBN 978-1-80061-421-5.
The Practical Philosophy of AI-Assistants is an insightful and oft-playful exploration of the various conundrums that arise when envisioning the ultimate AI assistant. The apparent intention of the text is not to convince readers why an AI-assistant would be a boon to society. Although these ideas are alluded to, the focus of the text is instead a speculative exploration of how this creation should be designed in order to make it a supportive companion to humans. The two authors each constitute one side of the disciplinary coin: Suman Gupta, a professor of Literature and History, represents the humanities, and Peter H. Tu, an AI engineer and researcher, represents the sciences. The titular ‘conversation’ is a core aspect of the text’s formatting, with Gupta and Tu individually authoring alternating chapters, allowing each author to respond to the other’s arguments as they develop. The text constitutes a lively academic dialogue that navigates complex theoretical discussion at an enjoyably quick pace. The initial premise is that Tu will introduce an aspect of contemporary AI technology, and its potential future applications, and Gupta will respond with the potential social and philosophical ramifications of this technology, and how these might be accounted for in the planning stages of their AI. Although the strictness of these roles (Tu as explainer and Gupta as critic) becomes more flexible as the book veers into more speculative territory, this initial interplay helps to ground what is often dense theoretical discussion. Additionally, specialist language has been deliberately moderated to make the text accessible for readers from a variety of subject backgrounds. In format and in content, Practical Philosophy is a book committed to bridging the gap between the humanities and the sciences in discussions centring AI.
The book is a follow-up of sorts to Gupta and Tu’s 2020 publication What is Artificial Intelligence? Conversation between an AI Engineer and a Humanities Researcher, which has the same format and similar topics of concern. However, where the earlier text focuses on defining AI, this text has an ostensibly more concrete goal in considering how a specific form of AI might be optimally designed and deployed. Gupta and Tu may seem overly optimistic in their ambitions to create an AI assistant that can be a ‘constant companion and counsellor’ (xiii), however, this idealism is understood by the authors as a methodological choice. Gupta and Tu argue convincingly that the limits and failings of current AI technology will be exposed by their idealistic ambitions. However, the intentional optimism of the text, and Gupta and Tu’s seeming consensus that an AI assistant can be a force for good, does mean that the socio-cultural utility of the AI assistant can feel a little underexplored. However, if you accept this conceit as Gupta and Tu present it, there is a lot to be gained from their rich discussion of a more utopian application of AI technology.
The book is divided into four sections, each one examining a function that Gupta and Tu have agreed the ideal AI assistant should have: (facial) recognition and identification, (human) communication, explanation, and (social) civility. Each section contains a handful of chapters, which allows the authors to bounce ideas back-and-forth considerably before moving onto the next topic. Tu will also occasionally include co-author/s in his chapters, often enlisting a colleague who can provide specialist insight on a certain area of discussion. In contrast, Gupta does not include any collaborators in his chapters. While this choice is not detrimental to his argument, Gupta’s inclusion and collaboration with other humanities scholars concerned with AI could have added some additional depth to the text.
The first section of the text, Recognition and Identification, contains the most granular information on contemporary AI technologies compared to all of the other sections. It begins with a chapter from Tu explaining some of the parameters and applications of contemporary face recognition technology. There is a seemingly unintentional irony in the way Tu uses research into facial recognition technology commissioned by DARPA, a branch of the US Department of Defense, and then hypothesizes how the AI assistant may use this technology to help one out of awkward situations at a cocktail party. Gupta in his responding chapter picks up on this dissonance and chastises Tu somewhat for not considering the ways in facial recognition tools could and already are being used by powerful bodies, namely governments and militaries, to influence people. Thankfully, however, rather than concluding that there is something inherently insidious about these forms of technology, Gupta instead suggests that their AI assistant could represent a move away from traditional face recognition formats and applications. The subsequent chapters involve Gupta and Tu debating how best to index and understand the forms of recognition they are hoping to incorporate into their AI, how these could work, how they should work, and why these features would make the AI assistant appealing to consumers.
This first section of the text, from the pun-based chapter titles (‘Face-Off, ‘Facing Up to Challenges’), to the playfully combative tenor of Gupta and Tu’s responses to each other, set the tone of the text. The authors are distinctly serious and academic in their arguments, but they are also frequent collaborators who are having appropriate amounts of fun with a topic they are both interested in.
The second section, Communication, marks a move away from the first section’s grounding in existing technology. In this section, Gupta and Tu discuss the forms of communication an optimal AI assistant should be capable of. Due to this topic being purely theoretical, this section pulls in lots of interesting directions from Theory of Mind, to the Turing Machine, to the concept of embedded biological knowledge. The variety of ideas are fascinating but can be a little dizzying to read as Gupta and Tu ricochet their adaptations of existing theory back and forth and try to come to a consensus. What the two authors agree on from the start is the significance of communication skills for an AI designed to interact with humans. It is in establishing this notion that Gupta makes a rare reference to the commercial prospects of the AI assistant, noting that the ability to converse would make the AI a ‘true friend’ and ‘a very lucrative commodity’ (91).
The third section dovetails nicely from general communication in the previous section to a specific form of communication: Explanation. Although initially about how explanations play a part in knowledge construction, and why this would require conscious and careful design for an AI assistant, these chapters quickly turn from how explanations function to human cognition and consciousness. The remainder of the section is spent exploring different theories of human thought, and considering which theory Gupta and Tu should adopt when designing their AI assistant’s capacity to explain.
The fourth and final section, Civility, is perhaps the most pertinent to contemporary debates around the ethics of AI. These chapters explore how the AI assistant is often viewed, and could be created, as something that threatens social civility, but also how it has the potential to improve it. Gupta acknowledges, however, that in either case our contemporary rules, norms and practices (RNPs) would have to adjust or shift in response the existence of the AI assistant. In the penultimate chapter of the section, Peter Tu and collaborator Mark Grabb write a short fictional narrative about the potential applications of an AI assistant improving social civility. Although a refreshing departure from the theory-dense discussion of the other chapters, Gupta summarily critiques the attempt for being an unrealistic depiction of both AI and human behaviour. Ending the section, and the text, on the somewhat condemnatory note that ‘[p]erhaps this vision of the AI-Assistant […] is more about the excessive claims made on its behalf by the AI industry than about life, the world, or anything’ (246).
An aspect that affects the impact of the texts’ arguments is the lack of a formal conclusion chapter. In a text as full of fascinating discussion and clashes of opinion as Practical Philosophy is, a conclusion feels like a necessity, and its absence is felt. Although Gupta and Tu’s omission of a conclusion chapter may be an inevitable consequence of formatting the text as a ‘conversation’, or, perhaps intended to convey the ongoing nature of their debate, the result is an anticlimactic ending to a very ambitious discussion.
Overall, The Practical Philosophy of AI-Assistants is a compelling interdisciplinary exploration of the philosophical, social, cultural, and practical considerations that need to be made in designing an optimal AI assistant. The format permits for rigorous and well-researched discussion between the two authors and their collaborators on four key functions of the optimal AI assistant. Although there are aspects of the topic that the book could elaborate on, or explore in more depth, the text ultimately delivers on the promise of its title as a fascinating conversation between two subject specialists.
Faye Lynch, University of Liverpool
