Initial Prompts
In preparing my application for the Turing Fellowship (2021-22), I drew upon three main prompts, which I will set out below in hope they help frame the discussions for the Reading Group. The three prompts, loosely defined are: (1) Units of Meaning; (2) AI as Inter-discipline; and (3) AI and Structuralism. The first two prompts are specific and recent, while the third relates to a longer term interest.
(1) Units of Meaning: In 2019, I began experimenting with AI (via IBM online tools). I was struck with how the backend of AI is based on fairly prosaic data construction and keywords. We can think of AI as cutting-edge technology, yet much of what we need to make it work involves fairly rudimentary ways of thinking and processes of development (even analogue ways of working); the need to sketch out specific processes, mark-up specific language, and to think through variables and probabilities. Concurrently, I took to rendering Flaubert’s novel, Bouvard and Pécuchet, as a text adventure game (see: textadventures.co.uk). Echoing the futility of the two copy-clerks in the book, who move from one field of science to another in a vain attempt to obtain greater clarity on their lives, my idea of converting the book into a Boolean-operated text adventure game was always intended to be a futile task. At the end of the novel, the two copy clerks are said to go mad and end up returning to indiscriminately copying out texts from all that surrounds them (whether significant or banal). Despite having been written in the late nineteenth century, I can’t help reading the book as a prefiguring of the Internet, with all its overwhelming surfeit of information and ‘knowledge’. My intention was to provide a POV, text adventure of the novel, which, at the end, would simply lead straight out into the ‘wilds’ of the Internet. Of course, the whole endeavour would require a huge amount of (idle) time. I have since started work on a (print-based) ‘novel’ of the project, as this seems to equate better to its likely audience. Nonetheless, the draft of the game, which only provides the opening chapter, remains a point of interest. In addition, as I worked on the game, I came across AI Dungeon, which, originally released towards the end of 2019, is a text adventure game that uses AI to generate content. It was developed with an early version of the GPT-2 natural-language-generating neural network, created by OpenAI, which allows the interface to generate original adventure narratives. It has perhaps since developed into a more satisfying multi-player domain, which perhaps gives it more of a sense of being a ‘game’ (and ironically involves a human multi-player approach). However, when I first encountered AI Dungeon I was struck with two things. Firstly, it was indeed effective at generating (seemingly) meaningful content and scenarios. Yet, my second observation was that of an unease, not because I feared AI might be as able, if not better at generating text, but that somehow the game play felt aimless, even ‘uncanny’. It presented an environment that could go on forever, always apparently making sense, yet without any sense of purpose (rather like the disquieting ‘dream’ narrative in the final chapter of Julian Barnes’ A History of the World in 10½ Chapters). This led me to question artificial ‘intelligence’, and/or to consider perhaps the potential for ‘ambivalent intelligence’ (and which I am since now casting in terms of Foucault’s account of the order of resemblance, which he situates in 16th Century thought).
(2) AI as Inter-discipline: Today, we are attuned to the fact of Big Data, i.e. that is ‘our’ data, providing patterns that on aggregate and in all probability describe and predicate us. However, as Kate Crawford accounts, in Atlas of AI (2021), despite ever larger datasets and increased computing power, we do not necessarily lead to a better understanding of ourselves. So, not only are there ecological concerns relating to the rapacious consumption of data, there are methodological quandaries too. While issues of privacy are rightly a critical concern, when we drill into what data is being tracked and compared about us, it is often not particularly rewarding. The units of analysis are often crude and do little more than attempt to sell back to us what we already know. There remains a space, then, for more careful consideration, for more participation on the designs and implementations of AI, and crucially for cross-disciplinary dialogue. Speaking at a Turing Institute event (see Index Card #2), Kate Crawford praised the growing diversity of those engaged in contemporary scholarship (not least in terms of gender). Yet, equally, she bemoaned the loss of a certain spirit of interdisciplinary dialogue, more evident, she argued, in the early part of the twentieth century, when, for example, anthropologists talked with computer scientists at public events. She suggests of the need (urgent over the next decade) to re-conceptualise AI as an inter-discipline, and as one that has to be very much grounded in the communities who are being affected by these tools everyday. In keeping with this observation, I have been inspired to re-think how we might understand data and its ‘units of analysis’, which in turn has led me to return and re-examine structuralist theory, notable in structural linguistics and anthropology, but also hugely influential across a wide range of fields.
(3) AI and Structuralism: A particular prompt, or pondering, takes me back to my long held interest in structuralism. In fact it takes me back to my time at university reading Claude Lévi-Strauss, and being particularly intrigued with his brief asides about the prospects of using computers to handle the paper-based computations he was making with a huge range of data of myths from around the world. Later, this interest combines with my long-term reading of Roland Barthes, who similarly offers numerous asides regarding science and computations. While there was a good deal of critique offered within the humanities to structuralism, there is arguably a case for returning to its core interests, not least given contemporary developments in computing, AI and natural language processing. It is pertinent, for example, to explore how the aspirations of structuralism still relate to questions of culture, language and technology, and how this earlier period of intellectual history might impact on our recent thinking. Lévi-Strauss was keenly engaged in the developments of cybernetics and he was aware of the statistical limitations of his work. His work was profound, for example, for setting out a case for ‘structures of thinking’. Indeed, for Lévi-Strauss, structuralism was not about interpreting and combing datasets. It was a way of questioning the ‘structuration’ of self and society. In his case, studying myths (from around the world), he considered how myth (or data) was what underpinned our way of thinking (rather than being merely the content of our thought). In this sense, we might consider data not as a ‘trail’ that comes after us, but as our very way of being in society.
With these prompts in mind, this project works upon 3 intersecting trajectories:
1. Firstly, we can ask again: what were the ambitions and aspirations of structuralism, and how did these situate with respect to the developments in cybernetics and computing? What connections can we consider with contemporary debates of data, AI, and natural language processing?
2. We might argue that the paradigms of the arts and humanities today are in one form or another a response to/against structuralism. For example, post-structuralism, deconstructionism, post-/de-colonialism, feminism, queer theory, new materialism, affect, and post-humanism, are all a critical response to the ‘units of analysis’ of (and lack of agency in) structuralism. What underpins this critique and how might this relate to current methods and practices of data and AI?
3. In drawing these two considerations together, we can ask what a ‘return’ to structuralism might offer in re-thinking our understanding and development of the relational handling of data, and of the ontological status of data in our lives today.
Overall, the project responds to two strategic goals of the Turing Institute: (1) Methodological Challenge areas; and (2) Public engagement work. Of the latter, the approach is for collaborative enquiry and open, public debate. Of the former, the project seeks to contribute to debates about ‘finding structures in data’ and supporting our understanding of ‘theoretical foundations’.
Notes on Structuralism & AI 