Medialab stories: how to align ANT and digital methods?
Dominique Boullier, Digital Humanities Institute, EPFL
(Translation -slightly revised- from the previous post in French)
Note: This is a blog post and not a paper. Please accept the flaws in this draft of a potential paper (translation included!)
The time for reviews has apparently come for medialab (cf. Venturini et al. 2017), I will do my part, because there are indeed many lessons to be learned from this experience. My point of view will differ quite clearly from the article mentioned above, which notes first and foremost the work of constructing digital methods carried out with medialab, because I always situate digital methods within the conceptual framework that they put in place, sometimes without saying or knowing it. To better understand what is happening to us collectively, as social sciences, with the digital world, is indeed a task to which I have been devoting myself for the last 5 years. And the support of STS traditions is essential in this regard. The historical dimension is essential to understand the very special “assemblage” of medialab from 2009 to 2017. I would like to show here how a very particular alliance was formed between ANT (Actor Network Theory) in its Latourian version on the one hand and the numerical methods in their UTC version (University of Technology of Compiegne), i. e. web topology, on the other hand. This will recognize the limitations of the experiment, but it will also teach us a great deal about what is at stake, including the place of ANT in all social science approaches, how it has been used and how it could still be used. This will involve revisiting some of the beliefs spread around what the medialab and ANT were doing at the same time. Thus, it seems to me quite wrong to characterize the relationship ANT/digital by the “carbon copy” effect (the web duplicates the social), by qualiquantitative methods, or by the gains of knowledge allowed by a supposed zoom (exploration of the social by change of scale). To put all these points in their proper place, it is necessary to put the entire ANT in historical perspective and to give an account of the technical methods and devices that have enabled it to exist, as has always been the case with the STS, and ANT has became a reference framework for them. Keeping in mind the digital approach, I will therefore first give a brief historical overview of ANT and its specificities before turning to digital methods in detail. I would like to make it clear that all the points of view I am going to develop were born within this setting provided by the medialab, of which I was scientific coordinator with Bruno Latour until 2015. They resume debates that have been sketched out within the team itself, notably for the writing of the article “The whole is smaller than its parts”. In this article, I use references to the work I’ve been doing on this third generation social science blog, and I test the hypotheses I’ve been working on for almost five years on the era of quantification that we live in today.
How does ANT make a difference in the social sciences?
I have already positioned (Boullier, 2015,2016) clearly the ANT and its reference authors (Callon, Latour and Law) as the founding authors of a new era of quantification, which I have named the third generation of social sciences. Some were very surprised by this choice, as digital and quantitative seem to have remained peripheral in the work of all the authors of the ANT. However, I persist because it is not digital technology that is at stake here, but the cognitive framework that it proposes, particularly with the Machine Learning, which was anticipated in a certain way by the posture of ANT. What should we retain from the ANT, indeed, if we want to reduce it (abusively) to a few “slogans”?
1/ Non-humans have their own agency, agency and power of action. The oldest perhaps remember the scandal represented by Michel Callon’s article on scallops. It was a systematic outcry, scandal and vendetta against those of whom I was one, who defended this thesis of the power to act of objects, things and in this particular case of the animals that are the scallops. The symmetrical description of the processes of delegation between fishermen and scallops was the prime target of detractors, and more than thirty years later, some still have scallops across their throats. A whole series of articles on objects were published in the same fashion and the authors’ writing skills even made them, mortal sin, stories accessible to the general public, in other words “anecdotes” instead of demonstrations. This shift towards taking serious account of the “roles” of artifacts, living things, etc., has become increasingly shared and was already taking up concerns from Foucault on devices, from Simondon on the way in which technical objects exist, from Mc Luhan on the media, etc. It is this shift that constitutes a serious path for the digital world, or at least a posture that allows us to be affected (Haraway) by the digital age at the risk of being treated as “technological determinists”, an accusation that completely ignores the thread of the narratives offered in the cases studied. But even more important and largely ignored until now, this provocative stance allowed the agency’s distribution to be reopened, the powers to act, hitherto devolved solely to humans, according to the modernist dogma (or naturalist to use Descola’s ontologies) either in aggregated form of social structure or in individualized form of preferences and strategies. What B. Latour said again in Reassembling the social/ Re-doing sociology: “Can we anticipate a social science that takes seriously the beings who make people act? “.
2/ The agency’s move is based on a well-known but little insisted loan, a borrowing from Greimas’ semiotics. The concept of actant allowed the potential arrangement of any entity that populated the narratives to be aligned beyond Propp’s tradition. While Greimas’ formalism was certainly not preserved, the principle allowed for more open stories to be told and the concept of allies to be formalised, in particular, which extended the idea of adjuvants and opponents (without this being done in a strategic perspective, contrary to what some interpretations thought). One of the most striking misunderstandings for the ANT will be precisely the use of the term “network” to characterize what is above all a narrative, which makes connections emerge, neither strategically nor structurally, while network analysts still treat them according to these two points of view. At the same time, this makes it possible to point out the key importance of narratives, which will be the preferred methods of restitution by all the authors of case studies, often very inspiring, and to underline the weight of words, i. e. the importance of publications in the follow-up of the actors. The first article published by B. Latour with P Fabbri was about the rhetoric of scientific articles and the importance of these textual forms anchored in different truth regimes will be a constant concern for him. More generally, the Scientometry taken up by Callon and Courtial in particular, will experience a new youth thanks to the serious consideration of all the components of scientific activity published, including terms and concepts, which live their lives within scientific networks. The promising future of Scientometry in this context will not be exploited, however, because, firstly, in an inexplicable way for B Latour, Mr. Callon will give up Scientometry altogether at the very moment when the power of digital technology would have enabled him to forge a standard of non-positive quantitative studies (“Réseau-LU” was the last attempt in this direction) and secondly, the work on quotations was quickly reduced to the analysis of keywords, without any power to act but treated only as indicators of the activities of institutions or individuals.
3/ The importance of these narratives, however, is an important feature of the method, consistent with the discussion of innovations characterized by their uncertainty (hence the principle of symmetry between those who have succeeded and those who have failed, and the methodological advice of “following the actors”), consistent also with a deliberate emergentist approach to the social order. By recounting all the mediations that make a phenomenon, a scientific fact, an innovation or an institution, this is in direct opposition to all the social sciences that take for granted determinant variables or possible computable spaces of choice. The price to be paid remains an impossible comparison, a weak formalisation, the use of concepts that are conducive to misunderstandings (such as translation or attachments) and even become a quality of the story. Let us not forget, however, that it is thanks to this method that the exploration of unstable or ongoing situations and processes was possible, against all those who would like to reduce to a few indicators the analysis of all observable processes on digital networks. This non-positivism, this ability to remain in uncertainty about the concepts or the status of the entities involved (hence the use of quasi-subjects or quasi-objects borrowed from M. Serres) remains a salutary precautionary principle at a time of galloping algorithmic positivism. In a line similar to Deleuze’s, the analysis of folds or lines of flight must remain the major concern if we want to avoid being captive of our own models. However, it is clear from the outset that this poses a cultural problem not only with the classical social sciences but with all the research in computer science.
This cultural shock provoked by ANT is recognized everywhere, experienced by researchers, always vigorous in all social sciences and in the natural sciences as well, who have launched the trial in relativism on several occasions. But it must be acknowledged that it has been amplified, or even provoked by the ANT’s claim to settle their fate definitively in all the pre-existing social sciences, in a tabula rasa move, which is certainly enjoyable but counter-productive for one part, and above all contrary to the very principles of the STS founded by ANT. ANT itself should have considered itself only as a point of view, and above all as inseparable from its methods of exploration and the objects it chooses. This will have a considerable importance in the difficulty of the medialab in dialoguing with other social science approaches and will thus help to limit its influence, especially at Sciences Po, even though this splendid isolation could be presented as a title of glory sometimes. Why is it that ANT formalised in particular in the book “Re-assemling the social/ redoing sociology” did not apply to itself its own premises of sociology of sciences?
From one point of view to tabula rasa: an extension that contradicts the sociology of science itself
The systematic proposal of this book is to ask researchers to trace all the mediations of any phenomenon one by one and to refuse the “double-click” operations that allow entities whose existence cannot be justified, such as social structures or market laws, to act. This level 2, convened at will, is opposed to a method specific to the ANT for monitoring actors and mediations (and we can clearly see what in this monitoring can largely benefit from advances in digital technology). For sure, the critical analysis of the supposed power of structures or the market can be sometimes devastating: for example, when B. Latour, in another paper, literally dismantled the Durkheimian approach to religions. However, it cannot ignore what Bruno Latour’s sociology of sciences, in particular, has clearly shown: the “social structures” and their “causalities” have been constructed, admittedly, often arbitrarily, but they took the support of census quantification mechanisms to statistically construct the idea of a whole, comprehensiveness and laws applying to these aggregates, regardless of the action of individuals. This process has a history, the contributions of Desrosières have shown it widely but they have been exploited by ANT only to show the conventional aspect of these semi-overhangings –surplombs- (level 1.5 and not level 2) that the actors give themselves. However, among these actors, social scientists who use these statistical conventions to assign an agency to structures through scientific calculations and demonstrations must be considered separately, as they work to produce this level 2, whereas for ANT it is only a moment of conventional stabilisation of mediations that the actors give themselves. I put this period in perspective by talking about 1st generation social sciences because it seems to me above all important to recognize this validity of a claim to the agency of structures. What Latour grants to fetishes, (well-built and therefore powerful) why should we deny it to statistics and causalities built by generations of researchers from now on? Putting all these “believers” on levels 2 in the same bag is both ridiculing them in the name of a status overlooking the science made that has always been avoided by the ANT and it is both mixing apples and oranges (structures and the market) and despising distinctions that these approaches have spent considerable time building. No, the market approach, preferences and rational agent decision models, are not equivalent to that of a Durkheim or Bourdieu. To refuse to take sides in this quarrel is one thing, to reject them together because they ignore the mediations that make them exist is too easy and not historical enough.
Criticism of all these approaches as aggregates is not enough, because a network that is built up and that lasts also ends up naturalizing itself as an aggregate (it will be important in digital approaches), and even the narrative that ANT produced may become a meme as the one of the scallops, which goes beyond their author. In one case, indicators treated as variables of pre-selected social positions are aggregated; in the other, individual expressions of preferences are aggregated. Sampling is not only a refinement of statistical work, as E. Didier has shown, but also makes it possible to prove the existence of other entities, previously inaccessible, namely those individual preferences, in terms of choice of products, brands, etc. but also in terms of choice of opinion and electoral choices. What Weber had put forward in his work, insisting on the issue of values, found a vector of expression and calculation, which was not at all in the power of the census, for sure. That is why I proposed to consider the 1930s, along with sampling and opinion polls, as the landmark for a second generation of social sciences. However, beyond genealogy, it seems more important to me now to insist that these methods make it possible to produce, to make exist and to render calculable the specific agency of individual preferences, including the influence, that Katz and Lazarsfeld will theorize. The social sciences accept to lose the exhaustiveness of these enumerated agents (but only a few variables are listed) and to go deeper, domain by domain, into their individual preferences, whether they are declared, observed or expressed practices, which can be examined in detail under controlled conditions of representativeness.
Many believed that ANT was thus proposing a new version of this vision of preferences by highlighting the role of certain innovative entrepreneurs (in Schumpeter’s tradition) or some bold researchers as actors-strategists to explain the scientific facts. All these actors would produce network, which became a recognized resource of any strategy, and the term “actor-network ” was sufficient to encourage a shift in meaning. It must be recognized that when the starting point of the narratives focuses on a hero like Eastman, even though the object of the study is to show how a network is set up to invent the amateur photographer, the reader keeps in mind this starting point, the hero, who tends to become the strategist constructor of a network, even if it is shown that he has failed several times before succeeding. By telling the story, it is certainly quite easy to get rid of structures as actants, but it is more difficult not to accept individuals as actants, and even more difficult to succeed in the narrative bet of making the Berlin key or topofil the central actor. The tendency of some students to copy B. Latour’s narrative style without mastering the basic notions led me to recommend the refusal of this type of writing for those who would not have the talent of a Callon or Latour, which means few people! More importantly, ethnographic distribute agency as much as quantitative methods do, sometimes without the same concern for control, but they cannot help but do so, and any hint of an entry point into the narrative will give rise to interpretation, despite the author sometimes. This puts us on the path of the need to explicitly choose a point of view: that of structures, that of preferences and agents, or that of the entities that circulate, whose objects or quotations follow-up could demonstrate.
A risk: Emergentist ANT becoming a total explanation
But choosing one point of view does not mean disqualifying the other points of view. However, we will see that, no matter how digitally equipped they are, the pioneers who combine digital and ANT still have to prove that they are showing and demonstrating new processes, new concepts that have not been approached by traditional methods and specialists in the fields concerned. However, the review article that we have quoted fails to make this comparison, in terms of results, concepts, interpretations, etc. and it would be very difficult to indicate a significant result outside the development of platforms and datasets.
In fact, ANT has always chosen a point of view that I call an emergentist one, which has a philosophical filiation from Nietzsche to Deleuze, Bergson and Tarde. But this approach is in fact very relative to the state of knowledge on a given process. To bring out the role of non-humans, texts and the uncertainty that presides over innovation or science is to adopt an emergentist posture, which cannot ignore, however, that what emerges eventually stabilizes or disappears, and that this stage is also part of a consequent constructivism (as Latour says for science ready made and science in the making). Everything is built up, everything is emerging, but we live in our world only by getting used to it and folding back what has been unfolded only on rare occasions. We must therefore constantly look for what is emerging within the emergence itself, what “remains” we could say, what “surpasses us” because “well constructed” as Latour says, this “vif”, which Deleuze speaks of, without despising what reconstitutes territory and stasis. Durkheim delivers a mythical vision of religion, but he does so from the point of view of the power of structures to act, and no one can ignore the fact that, taken at a certain point in their history, the mediations that have made religions emerge have become natural and seen by “the actors themselves” as acting. When he uses this particular point of view to retrospectively look back at the genealogy of the elementary forms of religious life, he adopts a method that is precisely not suited to this, whereas ANT would be much better since it does not presuppose what has to be demonstrated, namely religion and its social force of imposition, the emergence of which must be followed precisely. This alone should help to give these sciences a status while refusing to give them ALL the power of explanation, as often do those who practice these sciences of social structures or even the sciences of individual preferences (the neo-classical economy not to name it). What the other points of view gain, however, is a capacity for comparison, certainly in reduced universes, but which allows them to extend their world of reference. However, ANT, by its requirement to monitor mediations, cannot claim to compare, calibrate or even discuss the validity of its observations in precise terms. It seems to me that if ANT were to assume its emergentist position, it should refuse to substitute itself for the other total explanations, and should insist on what they lack, namely a distribution of agency to non-human entities, in particular signs and objects, which circulate and are in the process of being set up, “in the making”. According to B. Latour, the “sociology of associations” (another definition of ANT) should be a “renewal” of the “sociology of social” (p. 375). We can understand “renewal” in two ways:
-one as the constant attention to emergence, to new candidates for existence, as a repopulation of what the aggregations of other sociologies reduce in causes (structure or market), as a equivalent of “mending” in some way, to reweave in a more finer way or closer to what emerges,
– the other as a disqualification of these other sociologies, a form of start from scratch, to deny any existence to structures or individual preferences in the possible accounts of the social and to claim that only associations are the legitimate viewpoint on the social, point of view which can claim to be independent of the state of its investigative devices and produce the meta-language that allows to put in their place the other points of view.
The omnipotence of a Bourdieu stating that “youth is only a word” and re-using the formula “are only” to impose a structural, deterministic point of view on any phenomenon was unbearable, but we can in no way do the same to criticize it by claiming that the social “is only” a matter of emerging associations. A pluralist posture is not a posture of conciliation, nor even diplomacy, but only a salutary reminder that ANT, habitus or game theory are only various points of view, which give back a distributed agency to certain entities, but never to all together, and that it is good to remember that it is not so easy to get off the rails of social reproduction, just as it is necessary not to forget that words, things, living things, climates, etc. affect us, far beyond the effects of structure, in situations and unexpected moments. Nevertheless this remains ONE point of view that cannot erase our individual capacities to make decisions, our preferences, in some circumstances, and only these ones.
Among the mediations that make us act, the quotes, as I said, have not been sufficiently exploited by Scientometry, because it should have used more specific linguistic methods, beyond Leximappe for example, but other mediations such the objects and their properties always required more and more detailed investigations. It was also one of my implicit criticisms of the approach to innovation as translation versus Everett Rogers’ diffusion theory in my 1989 article. While the criticisms made to Rogers were certainly relevant, E. Rogers was nevertheless more precise about the features, the properties of the innovations that allowed them to circulate. He attributed a stronger agency to the specificities of the objects or methods disseminated than did the theory of translation, which told stories that were certainly finer but always unique and incomparable, where everything was assembled despite the attention paid to the specific mediations that certain features of technologies can be. Let’s point out that I have always rejected this term of “assemblage”, which is very popular in English-speaking countries, because it allows us to skillfully conceal the absence of a point of view, and to take “everything” into account although we are aware that any observation and ana-lysis presupposes choices and therefore a point of view to reconstruct the history of this “assemblage”.
In this disqualification operation of all other approaches focused on other agencies (apparently only, as we will see it), ANT lost the opportunity to dialogue with currents such as social analysis of networks in particular. The medialab suffered from this isolation, despite the arrival of V. Lepinay and despite my efforts to forge links with researchers like E. Lazega. The misunderstanding about the concept of network was complete, admittedly, but it was compounded by the rejection of any structural approach to networks and even Lazega’s neostructural approach, which skillfully combined structures and preferences, i. e. network structure and the specific role of certain nodes. However, as we will see, the digital approach to networks was going to exacerbate this misunderstanding even though the methods adopted tipped the balance towards a structural approach to web topologies, far from any emergentist approach. We can understand the need to distinguish ourselves from a constant debate between Bourdieu and Boudon (to simplify), which was the daily life of sociologists in the 70s and 80s. This search for hegemony by each point of view made the discussion particularly sterile, but the vision of “Reassembling the social” consisted in putting them back to back in order to better claim to account for any social process in the making, while the other points of view only dealt with the social ready made. This radical and hegemonic constructivism seems in fact contradictory to the premises of the STS that contributed so much to the development of the ANT. Criticising the claim to the “all” of other points of view cannot lead to the claim to “everything is constructed”. To put it more simply, neither structure nor invisible hand nor flat world, but only analytical points of view on the world and equipped with specific tools and methods. For my part, I have tried to parallelize these points of view, in heritages, arbitrations and neighborhoods ( heritage, arbitrage, voisinage, Boullier, 2010) without ever claiming to provide a full account of any phenomena whatsoever. This methodological pluralism seems to be the only one compatible with a historical and reflective vision of social science work.
Digital method at the medialab and ANT: an incoherent alignment
The status of ANT in the history of the social sciences was therefore clearly not situated within the framework of a new era of quantification, as the methods adopted were primarily ethnographic, qualitative and casuistic. The coupling with digital has been tried several times, via scientometry with the Candide software for instance, a very interesting attempt to equip the analysis of the linguistic material of scientific texts. Its extension in the Réseau-LU software was also promising but very complex to manipulate, which led me, after several attempts, to advise my doctoral students not to use it for their work. However, the combination of analyses of social positions (authors) and semantic positions (concepts or terms) still remains a research issue at the moment, as the Cortext platform shows. In all cases, these were positions, although the methods looked for emergence, which was already an undeniable twist of the approach in terms of actor-network, because we did not follow the actors, we overhang them. This trend was accentuated when the medialab was launched by the choice of some numerical methods through the recruitment of engineers from the University of Compiegne (and myself from UTC as well) all trained by Franck Ghitalla. Places and names play a role in these stories of scientific choices: this filiation is not anecdotal because it indicates a choice among the digital methods, dictated by availability and public reputation, and not by the requirements of ANT. F. Ghitalla, a web studies pionneer in France, operationalized what Kleinberg had said in 1998, starting from scientometry and translating the topology of scientific networks to the topology of the web. Kleinberg’s work inspired Brin and Page for their Page Rank with the success we know. The calculated scores are fundamentally based on this idea of preferential attachment, which is in some ways a second-generation social science, that is, the value of preferences and choices. Web site administrators declare preferences by attaching themselves to other sites through hyperlinks and D. Cardon showed in his genealogy of Page Rank that it produces a form of vote, claimed as such by Google, and supposed to say something of the merit of a given page. Ghitalla’s approach was somewhat different since, as a structuralist by training, in the same languistics school as me (the theory of mediation of J. Gagnepain), his research orientations led him to use topology to detect relational structures between sites, a latent structure that allowed him to show affinity groupings by visualizing calculated clusters. Visualization techniques were crucial for him, an image specialist and trained in Bertin’s approach, as they were the only way to access the latent structure, aggregating the preferences of each site thanks to centrality calculations, enabled by graph mathematics and made operational by eigen-vectors’ techniques. From these hypertext links, nothing was said about the dynamics of specific entities (except to reconstitute evolutions using photographs of graph images) but nothing was said about preferences as such, which Google exploited to produce a ranking (not clusters). The question that G. Fouetillou posed in his thesis since the early 2000s on the status of hypertext links was in fact relevant in terms of social science viewpoints: to insist on nodes (and classify them as Google) or to insist on the structure (and produce graphical images like Ghitalla) from the same material and identical calculations. Ghitalla’s approach was adopted in the medialab, taking into account the reputation of the Gephi graph visualization software developed by Mr. Jacomy, a former student of Ghitalla and recruited at the medialab fairly quickly. This choice was therefore opportunistic and yet perfectly justified given the market for available technology. However, in 2009, the Facebook and Twitter platforms were already created with a whole battery of new metrics, but the question of the dynamics of social networks was thus absent from the Medialab’s work programme, since it actually focused on hyperlinks and websites.
The medialab was paradoxically in charge of developing massively structure-centred methods for exploring the web, through images of aggregates derived from links, whereas ANT had always advocated dynamically following the emergence of actors-networks. As the entire linguistic part of the LU-network used experimentally for controversial projects did not delivered the expected outcomes, the medialab thus focused on techniques that were in fact contradictory with ANT. One might have thought that this would have allowed a connection to the social analysis of networks, since the methods and concepts (such as Burt’s structural holes, for example) were essentially the same. But on the one hand, the web could never be considered as a complete network, which is difficult to take into account for the structural analysis of networks, and on the other hand, the claim of ANT never to evoke the very idea of structure, or even patterns, made any dialogue impossible.
The talking points trap: carbon copy, qualiquanti and zoom
In fact, the work carried out in the vein of these topologies makes it possible to report on the positions of entities immediately qualified as actors, by a daring shortcut. Indeed, the theme of the web as carbon copy was one of the first “talking points” forged by T. Venturini, which provoked a lot of confusion, only removed by the occasional recourse to the vigilant digital methods of Rogers and Marres. However, the interpretations were rapid or even immediate in terms of social proximity or hierarchies between clusters or between nodes, immediately translated into social actors, groups or communities. “Carbon copying ” thus produced a very surprising digital positivism for researchers trained in STS and all the methodological precautionary advice, the framing of which by the” issues” of N. Marres, did nothing about it. It should be noted that F. Ghitalla drew the most coherent conclusions from it, since he abandoned the study of the web to work on complete or almost complete documentary corpus (such as patents for example) and make sure that what he states can be directly linked to social actors observable with traditional methods, without the mediation of platforms, made up of robots, fake accounts, multiple accounts (boyd and Crawford). He is thus approaching the structural analysis of networks while using the same network topology techniques. The medialab never did this aggiornamento that would have been painful because contradictory with the reference theory, ANT.
Bruno Latour and others saw the problem well and especially when combinations were attempted with a temporal visualization of the networks, as if the problem was a technical issue or should be limited to the addition of a new feature, time, in the topologies. Current multilayer networking techniques and Machine Learning should make it possible to move this issue forward, but only if specific skills are available and the choice is made to make this issue a priority, in order to better digitally extend the powerful intuitions of ANT, as one viewpoint among others. The work of students trained in “controversy mapping” was in fact the best expression of this dynamic requirement to trace entities and not networks structures or influentials. The term “mapping” raised debates that were akin to the dispute about which entities we should follow. Within the framework of the Forccast programme, several methodological and technical innovations in the restitution of controversies were developed. But significantly, however, it was above all the timelines that were the most telling support to restore controversy. I have always thought that this form of visualization was the most powerful way to restore dynamics, at the expense of other connectivity elements, inevitably. Tarde predicted the era of newspapers made up primarily of curves: he would be delighted to observe continuous financial TV channels (Bloomberg) or managers’ dashboards, or the proliferation of polls during elections or data graphs in economic supplements. These graphic forms, these cognitive technologies, are actually used in social listening agencies much more often than graph images on the dashboards they offer brands. It is therefore not surprising that Kleinberg, the one who founded, we could say, the study of the topology of the web, was directly interested in the “bursts and cascades” as early as 2002. Then he developed in 2009 with Leskovec a meme tracker based on curves and propagation, that became a large part of the studies of Twitter (Boullier, 2017 to be published). I’m trying to develop a meme tracker at EPFL and the questions asked are more complex than a simple timeline, as one can suspect! But this makes it possible to account for these emergence processes, sometimes called buzz, which constitute one side of the social and only one of its faces, even if from my opinion, it would have been the one that Tarde would have favoured and that ANT had thought before digital networks emerged. The question representation of time was thus technically removed from the work of the medialab, which reinforced the gap with ANT, despite the work that I was able to conduct on events in particular. However, works such as the remarkable “factory of the law”, mentioned in the article mentioned above, are in fact time tracking devices, making it possible to follow changes in amendments and legal texts. This makes total sense in relation to ANT, or even for Scientometry, but this dimension is never presented as a radically non-structural point of view: indeed, the emphasis is put on correlation studies with the properties of the authors of the amendments, for example. However, it is possible to put forward this mechanism from an emerging point of view, which gives a real agency to legal texts as such. But to do so, however, it would still require that researchers in the field take the tool seriously to produce analyses and not just data. The possible reduction of the medialab to a producer of resource platforms would be very disappointing if we consider the theoretical power of its founder.
Students of controversies also have another instruction, apart from the timeline, that of conducting ethnographic field surveys. Their pedagogical virtues are undeniable, but they also posed problems of restitution. Efforts have been made to develop relevant visual and narrative formats and this is certainly a very consistent approach with the ANT tradition. Some of the medialab’s work on climate has also attempted to mobilize this method in a much less systematic way. Yet, the inaugural discourse of the medialab was entirely based on this principle of “qualiquantitative methods”. Again, the limitations of this talking point should be noted. From a quantitative point of view, we have seen that the preferred approach was that of centrality scores from crawls on the web. However, the work was quickly reduced to classical calculations because the medialab did not have statisticians or people trained in high-level mathematics that graph theories might require. The rather basic attempts that I was able to make in this direction, for an international survey on habitèle (our personal data ecosystem) and for an investigation on digital books, with Maxime Crépel, forced us to call upon external skills with a complete decoupling effect with the usual digital methods. A programme of work on a “medialab influence ranking” that I had announced with Audrey Lohard in our book “Opinion mining and sentiment analysis” could not be carried out for the same reasons. It involved the complex use of indicators obtained in topologies to produce a proxy for the types of influence. It was therefore a question of admitting another point of view, that of influencers, i. e. the role of individual nodes in thematic networks. This was not a very late position but rather a classic second-generation position, the preferences of opinion leaders in the Katz and Lazarsfeld lineage for example. But the conceptual and methodological equipment of the medialab did not allow its implementation. This is not surprising, because a laboratory cannot have all the skills and orientations at once. However, the discrepancy between the announcements on the quantitative plea and the reality of the work focused on the centrality calculations should be noted to show the necessary effort to bring coherence that remains to be made.
Qualitative approaches on the other hand have not seen any particular innovation, whereas it is still necessary to combine linguistic material with indicators or traces of lower levels. We tried this with Maxime Crépel during our international investigation on “habitèle”, but the case is very complex, especially when we collect expressions in Mahrati, Korean, Arabic or Portuguese in the same investigation. Our database (available!) includes all of these cleaned and translated verbatims, but we have actually done the analysis work by hand and therefore in a much less systematic way than desired. This approach is a qualitatively demanding method of processing that is still very difficult to implement (besides the classical lexical analysis or some very reductive NLP methods). Another attempt was made with a program called “equipment for excellence” that had to archive interview-based survey reports and make them available to other researchers. An article by the authors of the device reports on this attempt, which proved to be unsuccessful, both for reasons of team management and for deep misunderstandings about what digital technology should enable in the archiving and study of these corpus. As we can see from the qualitative assessment, the difficulties of such an orientation are apparent, beyond the necessary linguistic skills that should be added, despite what the NLP entrepreneurs say (“our language analysis systems progress along with every linguist fired”). Language is still the main blockage in any computer processing, but Google’s recent advances in translation (and competitors such as DeepL) with the help of their computational power for deep learning can change the game more quickly than we think.
Until now, the solution to the deadlock in language processing was based on im-mediate visualization techniques. A dream of a shortcut of language issues(to the point of hoping to avoid legends on a map!)that also seduced the medialab and even Bruno Latour, so brilliant rhetorician, writer, storyteller and language lover. The place of design thus constituted a major feature of the medialab of this first period and certain achievements, notably in the context of the survey on lifestyles or climate, brilliantly illustrated the interest of visual quality, graphic, typographical, iconographic. However, as we have indicated for timelines, it is the scientific aims that should inspire designers’ creativity and not the virtuosity of these disciplines, where fashion effects are quite pronounced in inverse proportion to the requirements of justifying choices. It is here that the absence of tests and formalized approaches in cognitive sciences is felt because no debate between experts or between novices can help to decide on the relevance of certain choices. The concern for testing, at the level of algorithms, interfaces and visualizations of information should be part of any laboratory requirements. Because this testing is the only way to learn and not to adopt or reject solutions because of implicit a priori. The need to produce a language specific to dynamic visualizations, probably borrowed from cinema as Lecollinet or Manovich suggested, to extend Bertin’s graphic semiotics, is becoming crucial if we want to give users back a grip on their digital environment. However, this program is a vain wish for theory, while platforms like Google adopt a systematic testing policy called A/B testing, which doesn’t bother with theory and compensates with its learning power based on millions of users for a single variation of feature. Semiologists and designers alike have to be concerned when creativity depends on this kind of evaluation.
But the theoretical stakes for the social sciences are no less important. For the work on emergence, on propagations, on circulating entities, on the replications that I advocate as a coherent extension of ANT, requires inventing methods of exploration of this emerging world. The supposed im-mediate visualizations must be vigorously challenged and equipped in the publications by presenting other possible choices. Thus, in our article on zoom (Boullier, Crépel and Jacomy, 2015), we deliberately criticized a previous article that we had written, in which the clusters we presented were in fact “a ready-made network” whereas it was entirely built by a multitude of choices, sometimes highly debatable. We have thus shown how zoom is a pure illusion, since any recalculation of the graph requires modifying the parameters of spatializing and clustering algorithms to produce totally different results. And studying a segment of the graph sometimes requires changing the algorithm or feature previously considered.
This is all the more so since a considerable mutation occurs with Machine Learning (but it was already present when a serious exploration was carried out). Learning is not just about testing algorithms on learning sets and then transposing (executing)them to “real” data sets. It also involves testing the algorithms themselves to verify their performance against various criteria (including bias and variance). This means that the dynamics of the processes we are studying are doubling the dynamics of the calculation tools we use to account for them. This is what a company like Criteo does all the time, which sets prices for advertising placements on the web with a high degree of granularity on the fly. The equivalent in social science research does not exist yet, because the phenomena and problems dealt with are a priori selected as long-term (structure) or cycle (opinion) but not high-frequency ones. This is yet another direction that needs to be explored if we want to address replication processes, which are only ONE dimension of the social, but which will require skills of another kind and a radically new cultural posture of scientificity (approximation, induction, revision by learning and not by falsification). Because the key issue is not so much that of Big Data as that of Learning. We thus find a coherence between the object of study and the method, both focused on dynamic traceability. However, this dimension has not yet been taken seriously within the medialab. The focus on zoom, the change of scale, the granularity of the explorations, the revision of the whole and parts has been largely privileged (aggregation vs. situation, micro vs. macro that we go beyond but, by doing so, that we recognize). We will soon be showing in another paper, taking our article on “the whole is smaller than its parts” step by step, that we had made a mistake in favouring this spatial metaphor of the “whole”. Just as we have shown that the technical argument of zoom does not hold if we follow all the activities of changing points of view that this implies in fact, so it is necessary to consider that the argument of scales of exploration as a key for the digital does not hold either. Because ANT, like Tarde, is not interested in the “whole” or the “parts”, nor in the navigation between the two: it would only be to move from a point of view of aggregated positions to that of individualized positions, which can make a bridge between the approaches by the structures and by the individual preferences, but which would lead ANT to lose the emergence if it only follows these already defined positions. Everything that makes us act and goes beyond us is the radically new object of ANT that digital technology makes it possible to trace from now on. Neither the network structures, nor the nodes but what is circulating. And what circulates does not boil down to a decomposition of individuals or institutions into more elementary particles, but rather to the restitution of agency to these particles, which are specifically mediations, and not intermediaries, and which are therefore only captured as trajectories, as waves and not as positions. Therefore, what is a common feature of two profiles, for example a diploma, is no longer a common feature of individuals (because the profiles are not individuals and the individuals are precisely dividus, divided into features). This diploma is what circulates, makes people act, aggregates, creates routes, neighbourhoods and emergence. It is its circulation as a manifest feature according to “issues” (or situations) and not as a permanent quality that becomes the subject of the analysis. The same item (diploma) is thus treated differently and can be seen from a point of view of positions (and structures) or from a point of view of replications (and emergence) circulating and attaching beings between them. I consider then that this last point of view is much closer to the premises of ANT than the other and those of Tarde. However, digital methods must make it possible to identify the mention of the diploma as a meme, which dissolves the positivism of the diploma, as the sociology of positions and structures usually frame it. However, this does not prevent from using the diploma as an explanatory variable, from a structural effect point of view, of certain social behaviours or processes. This is only another point of view on the social, among the three possible. Getting rid of the idea of zoom in depth to think “point of view” is therefore another revision necessary to bring the ANT program supposed to inspire the first epoch medialab in coherence with the digital methods adopted.
The change of the “talking points” above mentioned necessarily implies that the continuity found by digital methods must no longer be extolled, which is another way of claiming to account for the “whole” better than other points of view. On the contrary, it is necessary to take on the analytical formatting of our points of view through our tools and concepts, and above all to ensure that some of them are consistent with others, and to ensure the alignment of the mediations that A. Hennion has proposed. In this context, it is quite conceivable to have digital methods to highlight the structural effects or the role of the influencers but also the power of action of the replicas, of these entities that circulate and that exceed us. This makes three distinct approaches, all equally legitimate but technically different, and impossible to combine except to keep the nostalgia of a map that would be the world as Borges described it. To mourn the abusive promises of digital technology, which basically only take up old idealistic promises, is not to abandon the power of the vision of the ANT pioneers who, before any real availability of digital methods, had been able to redistribute the agency in a new and profoundly non-modern way. It is in this respect that the choices of digital methods we make are also political, or even critical, while remaining pluralistic and historically reflective, because this is the condition for the possibility of scientific debate.