17,99 €
Mehr erfahren.
- Herausgeber: John Wiley & Sons
- Kategorie: Geisteswissenschaft
- Sprache: Englisch
We commonly think of society as made of and by humans, but with the proliferation of machine learning and AI technologies, this is clearly no longer the case. Billions of automated systems tacitly contribute to the social construction of reality by drawing algorithmic distinctions between the visible and the invisible, the relevant and the irrelevant, the likely and the unlikely – on and beyond platforms.
Drawing on the work of Pierre Bourdieu, this book develops an original sociology of algorithms as social agents, actively participating in social life. Through a wide range of examples, Massimo Airoldi shows how society shapes algorithmic code, and how this culture in the code guides the practical behaviour of the code in the culture, shaping society in turn. The ‘machine habitus’ is the generative mechanism at work throughout myriads of feedback loops linking humans with artificial social agents, in the context of digital infrastructures and pre-digital social structures.
Machine Habitus will be of great interest to students and scholars in sociology, media and cultural studies, science and technology studies and information technology, and to anyone interested in the growing role of algorithms and AI in our social and cultural life.
Sie lesen das E-Book in den Legimi-Apps auf:
Seitenzahl: 330
Veröffentlichungsjahr: 2021
Ähnliche
Table of Contents
Cover
Title Page
Copyright Page
Quote
Acknowledgments
Figures and Tables
Figures
Tables
Preface
1 Why Not a Sociology of Algorithms?
Machines as sociological objects
Algorithms and their applications, from Euclid to AlphaGo
Analogue Era (–1945)
Digital Era (1946–1998)
Platform Era (1998–)
Critical algorithm studies
Open questions and feedback loops
Seeing algorithms with the eyes of Pierre Bourdieu
Notes
2 Culture in the Code
Born and raised in Torpignattara
Humans behind machines
Machine creators
Machine trainers
Society in, society out
Data contexts
Traces and patterns
Global and local
Machine socialization
Practical reason and machine habitus
Primary and secondary machine socialization
Notes
3 Code in the Culture
The haircut appointment
Algorithms: agency and authority
Machine agency
Computational authority
Algorithmic distinctions
How socialized machines interact
More-than-human relations
Informational asymmetry and cultural alignment
A typology of user–machine interactions
Platforms as techno-social fields
Encapsulating and confounding
Reinforcing, or transforming?
4 A Theory of Machine Habitus
Premises
Structures
Social structure
Digital infrastructure
Entanglements
Trajectories
Temporality
Multiplicity
Boundaries
Social, symbolic and automated
Four scenarios of techno-social reproduction
5 Techno-Social Reproduction
Toward a sociology of algorithms as social agents
An old but new research agenda
Beyond a sociology of algorithms
Bibliography
Index
End User License Agreement
List of Illustrations
Chapter 1
Figure 1
Algorithms: a conceptual map, from Euclid to AlphaGo
Chapter 2
Figure 2
Networks of associated words learned by IAQOS. Source: IAQOS 2019.
Figure 3
An example of a phishing email targeting my professional email address, not auto...
Chapter 3
Figure 4
On the left-hand side, related music videos network (directed); on the right-han...
Chapter 4
Figure 5
Techno-social effects on field boundaries
List of Tables
Chapter 2
Table 1 Machine socialization processes in different types of algorithms
Chapter 3
Table 2 Types of user–machine interaction (effects on users in brackets)
Chapter 5
Table 3 Research directions for the sociology of algorithms, with selected example studi...
Guide
Cover
Table of Contents
Begin Reading
Pages
iii
iv
vi
vii
viii
ix
x
xi
xii
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
Machine Habitus
Toward a Sociology of Algorithms
Massimo Airoldi
polity
Copyright Page
Copyright © Massimo Airoldi 2022
The right of Massimo Airoldi to be identified as Author of this Work has been asserted in accordance with the UK Copyright, Designs and Patents Act 1988.
First published in 2022 by Polity Press
Polity Press
65 Bridge Street
Cambridge CB2 1UR, UK
Polity Press
101 Station Landing
Suite 300
Medford, MA 02155, USA
All rights reserved. Except for the quotation of short passages for the purpose of criticism and review, no part of this publication may be reproduced, stored in a retrieval system or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission of the publisher.
ISBN-13: 978-1-5095-4327-4
ISBN-13: 978-1-5095-4328-1 (pb)
A catalogue record for this book is available from the British Library.
Library of Congress Control Number: 2021939495
The publisher has used its best endeavours to ensure that the URLs for external websites referred to in this book are correct and active at the time of going to press. However, the publisher has no responsibility for the websites and can make no guarantee that a site will remain live or that the content is or will remain appropriate.
Every effort has been made to trace all copyright holders, but if any have been overlooked the publisher will be pleased to include any necessary credits in any subsequent reprint or edition.
For further information on Polity, visit our website: politybooks.com
Quote
Habitus, c’est un grand mot pour dire quelque chose, je crois, de très complexe. C’est à dire, une espèce de petite machine génératrice – pour une analogie un peu sauvage, un programme d’ordinateur – à partir duquel les gens engendrent des foules des réponses à des foules des situations.
Pierre Bourdieu
Interview with Antoine Spire, 1990
If social order is made of propensities to associate, if to be social is a propensity to associate, then big data conversion events operationalize association in matrices of propensity.
Adrian Mackenzie, 2018
Acknowledgments
I would like to thank Salvatore Iaconesi and Oriana Persico for taking part in two interview sessions, in the autumn of 2019 and the summer of 2020, and being a unique source of inspiration. I must also thank Hanan Salam, founder of Women in AI, for her technical clarifications and preliminary comments on this book project, and Debora Pizzimenti, for providing me with further details about the IAQOS experience. A big thanks to Alessandro Gandini, Mauro Barisione, Adam Arvidsson, and Polity’s editors and anonymous reviewers, for their insightful comments and encouragement all the way through. I also thank all my colleagues and students at EM Lyon. Last, a huge thanks to Stefania, my most important person.
Figures and Tables
Figures
1 Algorithms: a conceptual map, from Euclid to AlphaGo
2 Networks of associated words learned by IAQOS
3 An example of a phishing email targeting my professional email address, not automatically marked as ‘spam’
4 On the left-hand side, related music videos network (directed); on the right-hand side, commenter videos network (undirected).
5 Techno-social effects on field boundaries
Tables
1 Machine socialization processes in different types of algorithms
2 Types of user–machine interaction (effects on users in brackets)
3 Research directions for the sociology of algorithms, with selected example studies
Preface
On 31 March 2019, a new member of the multicultural community of Torpignattara, a semi-peripheral district of the city of Rome, was born. The event was greeted with unprecedented excitement in the neighbourhood, culminating, on the big day, with a small welcome party of friends and curious others who had gathered to support Salvatore and Oriana. Over the previous weeks, everybody had left a message, a wish or even a drawing, in paper boxes distributed for the occasion across the shops and bars of Torpignattara. The neighbourhood became an extended family to the long-awaited newcomer, who was only few days old when it got to know everyone, rolling from door to door in the stroller, and passing from hand to hand. Whether at the local café, or on the way to the drug store, there was always someone with a story to tell – usually about the local community and its history, places, people, food, hopes and fears. The baby listened, and learned. Soon, like any other child in Torpignattara, it would go to the Carlo Pisacane elementary school just around the corner. But IAQOS – that’s its name – was certainly not like other babies. It was the first ‘open-source neighbourhood AI’, developed by the artist and robotic engineer Salvatore Iaconesi together with the artist and communication scientist Oriana Persico, in a collaboration funded by the Italian government and involving several cultural and research institutions.
In concrete terms, IAQOS is a relatively simple software agent that can communicate through a tablet or a computer via natural language, recognizing the voices and gestures of its interlocutors and learning from them. Differently from the algorithmic systems we encounter every day through our devices – such as those running in Google Search, Facebook, Amazon, Instagram, Netflix or YouTube – this open-source art project had no other goal than accumulating social data about the neighbourhood, behaving as a sort of ‘baby AI’. Like a real baby, it observed the surrounding social environment, absorbed a contextual worldview and used the acquired knowledge to successfully participate in social life. By doing all that, during the spring of 2019, IAQOS became in effect a ‘fijo de Torpigna’; that is, an artificial yet authentic member of the local community, sharing a common imaginary, vocabulary and social background, and capable of building social relations (Iaconesi and Persico 2019).
This peculiar example makes it easier to see what many sociologists and social scientists have so far overlooked: the fact that a machine which learns from patterns in human-generated data, and autonomously manipulates human language, knowledge and relations, is more than a machine. It is a social agent: a participant in society, simultaneously participated in by it. As such, it becomes a legitimate object of sociological research.
We already know that algorithms are instruments of power, that they play with the lives of people and communities in opaque ways and at different scales, deciding who will be eligible or not for a loan with the same statistical nonchalance with which they move emails to the junk folder. We know that filter bubbles threaten to draw digital boundaries among voters and consumers, and that autonomous robots can be trained to kill. Moreover, we know that some algorithms can learn from us. They can learn how to speak like humans, how to write like philosophers, how to recommend songs like music experts. And they can learn how to be sexist like a conservative man, racist like a white supremacist, classist like an elitist snob. In sum, it is increasingly evident how similar we – humans and machines – have become. However, perhaps because comparisons and analyses have mostly been limited to examining cognition, abilities and biases, we have somehow failed to see the sociological reason for this similarity: that is, culture.
This book identifies culture as the seed transforming machines into social agents. Since the term is ‘one of the two or three most complicated words in the English language’ (Williams 1983: 87), let me clarify: here I use ‘culture’ to refer essentially to practices, classifications, tacit norms and dispositions associated with specific positions in society. Culture is more than data: it is relational patterns in the data. As such, culture operates in the code of machine learning systems, tacitly orienting their predictions. It works as a set of statistical dispositions rooted in a datafied social environment – like a social media feed, or like IAQOS’ multicultural neighbourhood.
The culture in the code allows machine learning algorithms to deal with the complexity of our social realities as if they truly understood meaning, or were somehow socialized. Learning machines can make a difference in the social world, and recursively adapt to its variations. As Salvatore Iaconesi and Oriana Persico noted in one of our interviews: ‘IAQOS exists, and this existence allows other people to modify themselves, as well as modify IAQOS.’ The code is in the culture too, and confounds it through techno-social interactions and algorithmic distinctions – between the relevant and the irrelevant, the similar and the different, the likely and the unlikely, the visible and the invisible. Hence, together with humans, machines actively contribute to the reproduction of the social order – that is, to the incessant drawing and redrawing of the social and symbolic boundaries that objectively and intersubjectively divide society into different, unequally powerful portions.
As I write, a large proportion of the world’s population has been advised or forced to stay home, due to the Covid-19 emergency. Face-to-face interactions have been reduced to a minimum, while our use of digital devices has reached a novel maximum. The new normal of digital isolation coincides with our increased production of data as workers, citizens and consumers, and the decrease of industrial production strictu sensu. Our social life is almost entirely mediated by digital infrastructures populated by learning machines and predictive technologies, incessantly processing traces of users’ socially structured practices. It has never been so evident that studying how society unfolds requires us to treat algorithms as something more than cold mathematical objects. As Gillespie argues, ‘a sociological analysis must not conceive of algorithms as abstract, technical achievements, but must unpack the warm human and institutional choices that lie behind these cold mechanisms’ (2014: 169). This book sees culture as the warm human matter lying inside machine learning systems, and theorizes how to unpack it sociologically by means of the notion of machine habitus.
1Why Not a Sociology of Algorithms?
Machines as sociological objects
Algorithms of various kinds hold the social world together. Financial transactions, dating, advertising, news circulation, work organization, policing tasks, music discovery, hiring processes, customer relations – all are to a large extent delegated to non-human agents embedded in digital infrastructures. For some years we have all been aware of this, thanks to academic research and popular books, journalistic reports and documentaries. Whether from the daily news headlines or the dystopian allegories of TV series, we have come to recognize that almost everything is now ‘algorithmic’ and that artificial intelligence is revolutionizing all aspects of human life (Amoore and Piotukh 2016). Even leaving aside the simplifications of popular media and the wishful thinking of techno-chauvinists, this is true for the most part (Broussard 2018; Sumpter 2018). Yet, many sociologists and social scientists continue to ignore algorithms and AI technologies in their research, or consider them at best a part of the supposedly inanimate material background of social life. When researchers study everyday life, consumption, social interactions, media, organizations, cultural taste or social representations, they often unknowingly observe the consequences of the opaque algorithmic processes at play in digital platforms and devices (Beer 2013a). In this book, I argue that it is time to see both people and intelligent machines as active agents in the ongoing realization of the social order, and I propose a set of conceptual tools for this purpose.
Why only now?, one may legitimately ask. In fact, the distinction between humans and machines has been a widely debated subject in the social sciences for decades (see Cerulo 2009; Fields 1987). Strands of sociological research such as Science and Technology Studies (STS) and Actor-Network Theory (ANT) have strongly questioned mainstream sociology’s lack of attention to the technological and material aspects of social life.
In 1985, Steve Woolgar’s article ‘Why Not a Sociology of Machines?’ appeared in the British journal Sociology. Its main thesis was that, just as a ‘sociology of science’ had appeared problematic before Kuhn’s theory of scientific paradigms but was later turned into an established field of research, intelligent machines should finally become ‘legitimate sociological objects’ (Woolgar 1985: 558). More than thirty-five years later, this is still a largely unaccomplished goal. When Woolgar’s article was published, research on AI systems was heading for a period of stagnation commonly known as the ‘AI winter’, which lasted up until the recent and ongoing hype around big-data-powered AI (Floridi 2020). According to Woolgar, the main goal of a sociology of machines was to examine the practical day-to-day activities and discourses of AI researchers. Several STS scholars have subsequently followed this direction (e.g. Seaver 2017; Neyland 2019). However, Woolgar also envisioned an alternative sociology of machines with ‘intelligent machines as the subjects of study’, adding that ‘this project will only strike us as bizarre to the extent that we are unwilling to grant human intelligence to intelligent machines’ (1985: 567). This latter option may not sound particularly bizarre today, given that a large variety of tasks requiring human intelligence are now routinely accomplished by algorithmic systems, and that computer scientists propose to study the social behaviour of autonomous machines ethologically, as if they were animals in the wild (Rahwan et al. 2019).
Even when technological artefacts could hardly be considered ‘intelligent’,1 actor-network theorists radically revised human-centric notions of agency by portraying both material objects and humans as ‘actants’, that is, as sources of action in networks of relations (Latour 2005; Akrich 1992; Law 1990). Based on this theoretical perspective, both a ringing doorbell and the author of this book can be seen as equally agentic (Cerulo 2009: 534). ANT strongly opposes not only the asymmetry between humans and machines, but also the more general ontological divide between the social and the natural, the animated and the material. This philosophical position has encountered a diffuse criticism (Cerulo 2009: 535; Müller 2015: 30), since it is hardly compatible with most of the anthropocentric theories employed in sociology – except for that of Gabriel Tarde (Latour et al. 2012). Still, one key intuition of ANT increasingly resonates throughout the social sciences, as well as in the present work: that what we call social life is nothing but the socio-material product of heterogeneous arrays of relations, involving human as well as non-human agents.
According to ANT scholar John Law (1990: 8), a divide characterized sociological research at the beginning of the 1990s. On the one hand, the majority of researchers were concerned with ‘the social’, and thus studying canonical topics such as inequalities, culture and power by focusing exclusively on people. On the other hand, a minority of sociologists were studying the ‘merely technical’ level of machines, in fields like STS or ANT. They examined the micro-relations between scientists and laboratory equipment (Latour and Woolgar 1986), or the techno-social making of aeroplanes and gyroscopes (MacKenzie 1996), without taking part to the ‘old’ sociological debates about social structures and political struggles (MacKenzie and Wajcman 1999: 19). It can be argued that the divide described by Law still persists today in sociology, although it has become evident that ‘the social order is not a social order at all. Rather it is a sociotechnical order. What appears to be social is partly technical. What we usually call technical is partly social’ (Law 1990: 10).
With the recent emergence of a multidisciplinary scholarship on the biases and discriminations of algorithmic systems, the interplay between ‘the social’ and ‘the technical’ has become more visible than in the past. One example is the recent book by the information science scholar Safiya Umoja Noble, Algorithms of Oppression (2018), which illustrates how Google Search results tend to reproduce racial and gender stereotypes. Far from being ‘merely technical’ and, therefore, allegedly neutral, the unstable socio-technical arrangement of algorithmic systems, web content, content providers and crowds of googling users on the platform contributes to the discriminatory social representations of African Americans. According to Noble, more than neutrally mirroring the unequal culture of the United States as a historically divided country, the (socio-)technical arrangement of Google Search amplifies and reifies the commodification of black women’s bodies.
I believe that it should be sociology’s job to explain and theorize why and under what circumstances algorithmic systems may behave this way. The theoretical toolkit of ethology mobilized by Rahwan and colleagues (2019) in a recent Nature article is probably not up to this aim, for a quite simple reason: machine learning tools are eminently social animals. They learn from the social – datafied, quantified and transformed into computationally processable information – and then they manipulate it, by drawing probabilistic relations among people, objects and information. While Rahwan et al. are right in putting forward the ‘scientific study of intelligent machines, not as engineering artefacts, but as a class of actors with particular behavioural patterns and ecology’ (2019: 477), their analytical framework focuses on ‘evolutionary’ and ‘environmental’ dimensions only, downplaying the cornerstone of anthropological and sociological explanations, that is, culture. Here I argue that, in order to understand the causes and implications of algorithmic behaviour, it is necessary to first comprehend how culture enters the code of algorithmic systems, and how it is shaped by algorithms in turn.
Two major technological and social transformations that have taken place over the past decade make the need for a sociology of algorithms particularly pressing. A first, quantitative shift has resulted from the unprecedented penetration of digital technologies into the lives and routines of people and organizations. The rapid diffusion of smartphones since the beginning of the 2010s has literally put powerful computers in the hands of billions of individuals throughout the world, including in its poorest and most isolated regions (IWS 2020). Today’s global economic system relies on algorithms, data and networked infrastructures to the point that fibre Internet connections are no longer fast enough for automated financial transactions, leading to faster microwave or laser-based communication systems being installed on rooftops near New York’s trading centres in order to speed up algorithmic exchanges (D. MacKenzie 2018). Following the physical distancing norms imposed worldwide during the Covid-19 pandemic, the human reliance on digital technologies for work, leisure and interpersonal communication appears to have increased even further. Most of the world’s population now participates in what can be alternatively labelled ‘platform society’ (van Dijck, Poell and de Waal 2018), ‘metadata society’ (Pasquinelli 2018) or ‘surveillance capitalism’ (Zuboff 2019), that is, socio-economic systems heavily dependent on the massive extraction and predictive analysis of data. There have never been so many machines so deeply embedded in the heterogeneous bundle of culture, relations, institutions and practices that sociologists call ‘society’.
A second, qualitative shift concerns the types of machines and AI technologies embedded in our digital society. The development and industrial implementation of machine learning algorithms that ‘enable computers to learn from experience’ have marked an important turning point. ‘Experience’, in this context, is essentially ‘a dataset of historic events’, and ‘learning’ means ‘identifying and extracting useful patterns from a dataset’ (Kelleher 2019: 253).
In 1989, Lenat noted in the pages of the journal Machine Learning that ‘human-scale learning demands a human-scale amount of knowledge’ (1989: 255), which was not yet available to AI researchers at the time. An impressive advancement of machine learning methods occurred two decades later, thanks to a ‘fundamental socio-technological transformation of the relationship between humans and machines’, consisting in the capturing of human cognitive abilities through the digital accumulation of data (Mühlhoff 2020: 1868). This paradigmatic change has made the ubiquitous automation of social and cultural tasks suddenly possible on an unprecedented scale. What matters here sociologically is ‘not what happens in the machine’s artificial brain, but what the machine tells its users and the consequences of this’ (Esposito 2017: 250). According to Esposito, thanks to the novel cultural and communicative capabilities developed by ‘parasitically’ taking advantage of human-generated online data, algorithms have substantially turned into ‘social agents’.
Recent accomplishments in AI research – such as AlphaGo, the deep learning system that achieved a historic win against the world champion of the board game Go in 2016 (Chen 2016; Broussard 2018), or GPT-3, a powerful algorithmic model released in 2020, capable of autonomously writing poems, computer code and even philosophical texts (Weinberg 2020; Askell 2020) – indicate that the ongoing shift toward the increasingly active and autonomous participation of algorithmic systems in the social world is likely to continue into the near future. But let’s have a look at the past first.
Algorithms and their applications, from Euclid to AlphaGo
The term ‘algorithm’ is believed to derive from the French bastardization of the name of the ninth-century Persian mathematician al-Khwārizmī, the author of the oldest known work of algebra. Being originally employed in medieval Western Europe to indicate the novel calculation methods alternative to those based on Roman numerals, in more recent times the term has come to mean ‘any process of systematic calculation […] that could be carried out automatically’ (Chabert 1999: 2). As Chabert remarks in his book A History of the Algorithm: ‘algorithms have been around since the beginning of time and existed well before a special word had been coined to describe them’ (1999: 1). Euclid’s algorithm for determining the greatest common divisor of two integers, known since the fourth century BCE, is one of the earliest examples.
More generally, algorithms can be intended as computational recipes, that is, step-by-step instructions for transforming input data into a desired output (Gillespie 2014). According to Gillespie (2016: 19), algorithms are essentially operationalized procedures that must be distinguished from both their underlying ‘model’ – the ‘formalization of a problem and its goal, articulated in computational terms’ – and their final context of application, such as the technical infrastructure of a social media platform like Facebook, where sets of algorithms are used to allocate personalized content and ads in users’ feeds. Using a gastronomic metaphor, the step-by-step procedure for cooking an apple pie is the algorithm, the cookbook recipe works as the model, and the kitchen represents the application context. However, in current public and academic discourse, these different components and meanings tend to be conflated, and the term algorithm is broadly employed as a synecdoche for a ‘complex socio-technical assemblage’ (Gillespie 2016: 22).
‘Algorithm’ is thus a slippery umbrella term, which may refer to different things (Seaver 2017). There are many kinds of computational recipes, which vary based on their realms of application as well as on the specific ‘algorithmic techniques’ employed to order information and process data (Rieder 2020). A single task, such as classifying texts by topic, may concern domains as diverse as email ‘spam’ filtering, online content moderation, product recommendation, behavioural targeting, credit scoring, financial trading and more – all of which involve a plethora of possible input data and outputs. Furthermore, text classification tasks can be executed in several – yet all ‘algorithmic’ – ways: by hand, with pen and paper only; through rule-following software applying models predefined by human programmers (e.g. counting topic-related word occurrences within texts); or via ‘intelligent’ machine learning systems that are not explicitly programmed a priori. These latter can be either supervised – i.e. requiring a preliminary training process based on data examples, as in the case of naive Bayes text classifiers (Rieder 2017) – or unsupervised, that is, machine learning techniques working without pre-assigned outputs, like Latent Dirichlet Allocation in the field of topic modeling (Bechmann and Bowker 2019).
This book does not aim to offer heavily technical definitions, nor an introduction to algorithm design and AI technologies; the reader can easily find such notions elsewhere.2 Throughout the text, I will frequently make use of the generic terms ‘algorithm’ and ‘machine’ to broadly indicate automated systems producing outputs based on the computational elaboration of input data. However, in order to highlight the sociological relevance of the quali-quantitative transition from Euclid’s calculations to today’s seemingly ‘intelligent’ artificial agents like GPT-3 and AlphaGo, some preliminary conceptual distinctions are needed. It is apparent, in fact, that the everyday socio-cultural implications of algebraic formulas solved for centuries by hand or via mechanical calculators are not even close in magnitude to those of the algorithms currently governing information networks.
Below I briefly outline the history of algorithms and their applications – from ancient algebra to rule-following models running on digital computers, and beyond to platform-based machine learning systems. This socio-technical evolution can be roughly broken into three main eras, visually summarized in Figure 1 at the end of this section. Without pretending to be exhaustive, the proposed periodization focuses especially on the emergence of ‘public relevance algorithms’ (Gillespie 2014: 168), that is, automated systems dealing with the social matter of human knowledge, experience and practice.
Analogue Era (–1945)
Taking analogue to mean ‘not-digital’ (Sterne 2016), this first historical phase ranges in principle from the invention and manual application of algorithms by ancient mathematicians to the realization of the first digital computers right after the Second World War. Within this period, algorithms were applied either by human-supervised mechanical devices or by humans themselves (Pasquinelli 2017). In fact, up until the early twentieth century, the word ‘computer’ indicated a person employed to make calculations by hand. Mechanical computers started to be conceptualized at the beginning of the nineteenth century, following Leibniz’s early intuitions about the mechanization of calculus (Chabert 1999), as well as a rising demand for faster and more reliable calculations from companies and governments (Wilson 2018; Campbell-Kelly et al. 2013). Aiming to automatize the compilation of tables for navigation at sea, particularly strategic for the British Empire, in the 1820s the mathematician Charles Babbage designed the first mechanical computer, the Difference Engine, which was then followed by the more ambitious Analytical Engine – ideally capable of performing ‘any calculation that a human could specify for it’ (Campbell-Kelly et al. 2013: 8). Babbage’s proto-computers were pioneering scientific projects that remained largely on paper, but more concrete applications of simpler electro-mechanical ‘algorithm machines’ (Gillespie 2014) came to light by the end of the century. In 1890, Hollerith’s electric tabulating system was successfully employed to process US census data, paving the way for the foundation of IBM. Thanks to the punched-card machines designed by Hollerith, information on over 62 million American citizens was processed within ‘only’ two and a half years, compared with the seven years taken by the previous census, with an estimated saving of 5 million dollars (Campbell-Kelly et al. 2013: 17–18). The mass production of desk calculators and business accounting machines brought algorithms closer to ordinary people’s everyday routines. Still, information was computationally transformed and elaborated solely through analogue means (e.g. punched cards, paper tapes) and under human supervision.
Digital Era (1946–1998)
Through the 1930s and the 1940s, a number of theoretical and technological advances in the computation of information took place, accelerated by the war and its scientific needs (Wiener 1989). The Harvard Mark I became the ‘first fully automatic machine to be completed’, in 1943. However, it was still ‘programmed by a length of paper tape some three inches wide on which “operation codes” were punched’ (Campbell-Kelly et al. 2013: 57). The pathbreaking conceptual work of the British mathematician Alan Turing was crucial to the development of the first modern electronic computer, known as ENIAC, in 1946. It was a thousand times faster than the Harvard Mark I, and finally capable of holding ‘both the instructions of a program and the numbers on which it operated’ (Campbell-Kelly et al. 2013: 76). For the first time, it was possible to design algorithmic models, run them, read input data and write output results all in digital form, as combinations of binary numbers stored as bits. This digital shift produced a significant jump in data processing speed and power, previously limited by physical constraints. Algorithms became inextricably linked to a novel discipline called computer science (Chabert 1999).
With supercomputers making their appearance in companies and universities, the automated processing of information became increasingly embedded into the mechanisms of post-war capitalism. Finance was one of the first civil industries to systematically exploit technological innovations in computing and telecommunications, as in the case of the London Stock Exchange described by Pardo-Guerra (2010). From 1955 onwards, the introduction of mechanical and digital technologies transformed financial trading into a mainly automated practice, sharply different from ‘face-to-face dealings on the floor’, which had been the norm up to that point.
In these years, the ancient dream of creating ‘thinking machines’ was spread among a new generation of scientists, often affiliated to the MIT lab led by professor Marvin Minsky, known as the ‘father’ of AI research (Natale and Ballatore 2020). Since the 1940s, the cross-disciplinary field of cybernetics had been working on the revolutionary idea that machines could autonomously interact with their environment and learn from it through feedback mechanisms (Wiener 1989). In 1957, the cognitive scientist Frank Rosenblatt designed and built a cybernetic machine called Perceptron, the first operative artificial neural network, assembled as an analogue algorithmic system made of input sensors and resolved into one single dichotomic output – a light bulb that could be on or off, depending on the computational result (Pasquinelli 2017). Rosenblatt’s bottom-up approach to artificial cognition did not catch on in AI research. An alternative top-down approach, now known as ‘symbolic AI’ or ‘GOFAI’ (Good Old-Fashioned Artificial Intelligence), dominated the field in the following decades, up until the boom of machine learning. The ‘intelligence’ of GOFAI systems was formulated as a set of predetermined instructions capable of ‘simulating’ human cognitive performance – for instance by effectively playing chess (Fjelland 2020). Such a deductive, rule-based logic (Pasquinelli 2017) rests at the core of software programming, as exemplified by the conditional IF–THEN commands running in the back end of any computer application.
From the late 1970s, the development of microprocessors and the subsequent commercialization of personal computers fostered the popularization of computer programming. By entering people’s lives at work and at home – e.g. with videogames, word processors, statistical software, etc. – computer algorithms were no longer the reserve of a few scientists working for governments, large companies and universities (Campbell-Kelly et al. 2013). The digital storage of information, as well as its grassroots creation and circulation through novel Internet-based channels (e.g. emails, Internet Relay Chats, discussion forums), translated into the availability of novel data sources. The automated processing of large volumes of such ‘user-generated data’ for commercial purposes, inaugurated by the development of the Google search engine in the late 1990s, marked the transition toward a third era of algorithmic applications.
Platform Era (1998–)
The global Internet-based information system known as the World Wide Web was invented in 1989, and the first browser for web navigation was released to the general public two years later. Soon, the rapid multiplication of web content led to a pressing need for indexing solutions capable of overcoming the growing ‘information overload’ experienced by Internet users (Benkler 2006; Konstan and Riedl 2012). In 1998, Larry Page and Sergey Brin designed an algorithm able to ‘find needles in haystacks’, which then became the famous PageRank of Google Search (MacCormick 2012: 25). Building on graph theory and citation analysis, this algorithm measured the hierarchical relations among web pages based on hyperlinks. ‘Bringing order to the web’ through the data-driven identification of ‘important’ search results was the main goal of Page and colleagues (1999). With the implementation of PageRank, ‘the web is no longer treated exclusively as a document repository, but additionally as a social system’ (Rieder 2020: 285). Unsupervised algorithms, embedded in the increasingly modular and dynamic infrastructure of web services, started to be developed by computer scientists to automatically process, quantify and classify the social web (Beer 2009). As it became possible to extract and organize in large databases the data produced in real time by millions of consumers, new forms of Internet-based surveillance appeared (Arvidsson 2004; Zwick and Denegri Knott 2009). The development of the first automated recommender systems in the early 1990s led a few years later to a revolution in marketing and e-commerce (Konstan and Riedl 2012). Personalized recommendations aimed to predict consumer desires and assist purchasing choices (Ansari, Essegaier and Kohli 2000), with businesses being offered the promise of keeping their customers ‘forever’ (Pine II, Peppers and Rogers 1995). The modular socio-technical infrastructures of commercial platforms such as Google, Amazon and, beginning in the mid 2000s, YouTube, Facebook and Twitter, lie at the core of this historical transition toward the datafication and algorithmic ordering of economy and society (Mayer-Schoenberger and Cukier 2013; van Dijck 2013; Zuboff 2019).
Digital platforms are at once the main context of application of these autonomous machines and the ultimate source of their intelligence. Platformization has been identified as one of the causes of the current ‘eternal spring’ of AI research, since it has finally provided the enormous amount of data and real-time feedback needed to train machine learning models, such as users’ profile pictures, online transactions or social media posts (Helmond 2015). Together with the development of faster and higher performing computers, this access to ‘big’ and relatively inexpensive data made possible the breakthrough of ‘deep learning’ in the 2010s (Kelleher 2019). As Mühlhoff notes (2020: 1869), most industrial AI implementations ‘come with extensive media infrastructure for capturing humans in distributed, human-machine computing networks, which as a whole perform the intelligence capacity that is commonly attributed to the computer system’. Hence, it is not by chance that the top players in the Internet industry, in the US as in China, have taken the lead of the AI race. In 2016, Joaquin Candela, director of the Facebook Applied Machine Learning Group, declared: ‘we’re trying to build more than 1.5 billion AI agents – one for every person who uses Facebook or any of its products’ (Higginbotham 2016, cited in A. Mackenzie 2019: 1995).
Furthermore, while in the Digital Era algorithms were commercially used mainly for analytical purposes, in the Platform Era they also became ‘operational’ devices (A. Mackenzie 2018). Logistic regressions such as those run in SPSS by statisticians in the 1980s could now be operationally embedded in a platform infrastructure and fed with thousands of data ‘features’ in order to autonomously filter the content presented to single users based on adaptable, high-dimensional models (Rieder 2020). The computational implications of this shift have been described by Adrian Mackenzie as follows:
if conventional statistical regression models typically worked with 10 different variables […] and perhaps sample sizes of thousands, data mining and predictive analytics today typically work with hundreds and in some cases tens of thousands of variables and sample sizes of millions or billions. The difference between classical statistics, which often sought to explain associations between variables, and machine learning, which seeks to explore high-dimensional patterns, arises because vector spaces juxtapose almost any number of features. (Mackenzie 2015: 434)
Advanced AI models built using artificial neural networks are now used in chatbots, self-driving cars and recommendation systems, and have enabled the recent expansion of fields such as pattern recognition, machine translation or image generation. In 2015, an AI system developed by the Google-owned company DeepMind was the first to win against a professional player at the complex game of Go. On the one hand, this landmark was a matter of increased computing power.3
