Reclaiming artificial intelligence accounts: A plea for a participatory turn in artificial intelligence inquiries

Abstract

How to participate in artificial intelligence otherwise? Put simply, when it comes to technological developments, participation is either understood as public debates with non-expert voices to anticipate risks and potential harms, or as a way to better design technical systems by involving diverse stakeholders in the design process. We advocate for a third path that considers participation as crucial to problematise what is at stake and to get a grip on the situated developments of artificial intelligence technologies.

This study addresses how the production of accounts shape problems that arise with artificial intelligence technologies. Taking France as a field of study, we first inspected how media narratives account for the entities and issues of artificial intelligence, as reported by the national press over the last decade. From this inspection, we identified four genres and described their performative effects. We then conducted a participatory inquiry with 25 French artificial intelligence practitioners’ to ground artificial intelligence in situated experiences and trajectories. These experiential accounts enabled a plural problematisation of artificial intelligence, playing with the geometries of artificial intelligence and its constituencies, while diversifying and thickening its problems.

To conclude, we discuss how participatory inquiries, through experiential and plural accounts offer a refreshing weaving of artificial intelligence problems into the fabric of its deployments. Our participatory approach seeks to re-politicise artificial intelligence from practitioners’ situated experiences, by making the ongoing relationships between past trajectories, current frictions and future developments tangible and contestable, opening avenues to contribute otherwise.

This article is a part of special theme on Analysing Artificial Intelligence Controversies. To see a full list of all articles in this special theme, please click here: https://journals.sagepub.com/page/bds/collections/analysingartificialintelligencecontroversies

[P1] If there were a button to shut down now and forever all AI projects, good and bad, would you press it?

[P2] This is an interesting point, as we have questioned the very existence of AI since the beginning of this workshop…

[P3] But once again, we do not know how it is defined, nor can we delineate its infrastructure.

After a short pause:

[P4] Is the political issue at hand a matter of ‘unplugging the technology’? Or does it require to revisit and unpack how technologies are associated with use and users? Although, I find your question fascinating: a world without AI? Let’s think about it in all its consequences…

[P5] We should probably find ways to discuss collectively what a good use of AI technologies would entail.

This discussion between artificial intelligence (AI) practitioners happened during a workshop we organised in Paris in 2022. Despite its somewhat amusing tone, it gives a good idea of how diverse and divergent conceptions of AI can be. From a global entity that can be ‘shut down’, or ‘unplugged’ – if it even exists? – to local associations of human and non-human entities, that should be democratically assessed – with what moral compass?

It is a good snapshot of the tensions we witness on a global scale. From the pause suggested by tech giants (Pause Giant AI Experiments, 2023), to the permanent international scrutiny of both breakdowns and successes of computational technology innovations, we are witnessing a great cacophony in AI accounts and we endure a continuous flow of conflicting news surrounding AI. Even academic studies of AI put us in a somewhat schizophrenic position, both affirming and denying its reality (Jaton and Sormani, 2023). In this noisy context, how to make sense of AI controversiality, and from whose perspective? This paper investigates how to reclaim ownership over the framing of AI-related problems to bring forth meaningful and pressing issues.

As part of the ‘Shaping AI’ project, ¹ our research is interested in ways of participating in the development of AI technologies. Although AI studies have taken a ‘participatory turn’ (Delgado et al., 2023), we argue that participation as a means to AI inquiries, has been overlooked. In this study, we rely on the recent developments in participatory design (PD) methodologies: instead of operationalising PD principles in the design and deployment of AI systems, we apply them in the study and problematisation of AI. In line with STS and ethnographic approaches (Seaver, 2017), we focus on the situations of AI. Equipping a situational focus means relocating empirical inquiries to consider the dynamics of development from concrete situated activities. AI problems are configured and articulated in relation to broader social, political, economic and technological processes, so we need to weave transnational, geopolitical and historical perspectives with situational particularities to better identify and problematise the objects at stake. We engaged in a collective political endeavour to reconstruct the progressive co-constitution of AI with the social and political fabrics of the situations in which it is realised. Through this participatory approach, we aim to thicken descriptions of how AI is actually realised, thereby reconfiguring entities and actors. But more importantly, our research seeks to enable the formation of meaningful problems; ones that reinforce the agency of the persons who experience these situations in order to influence future developments. Therefore, we advocate for a careful design of accounting processes to make the multi-scalar situations of AI realisations visible and disputable.

To grasp the interplay between the different scales and sites of AI development, we chose to ground our study in the French context. In addition to an easy access and intimate knowledge, we believe that France offers a valuable standpoint to pluralise the problematisations of ongoing AI developments, since many AI studies focus on North American settings. While France is not a leading country in the production of AI models, it is known for its high standards in mathematics education, with a deplored brain drain towards North American companies, as recent success stories attest (e.g. Hugging Face or Mistral AI). To maintain a ‘national sovereignty’, President Macron, who praises the start-up model of Silicon Valley, launched a national strategy in 2018 (now in its second phase). This strategy fostered and financed AI research clusters made out of public and private organisations, and has equipped the country with a supercomputer (called ‘Jean Zay’, housed at the French National Centre for Scientific Research (CNRS)). There is a strong political push to experiment AI in all sectors, which reconfigures workplaces and economies in a fast pace, starting with state services. Lastly, the whole French tech ecosystem actively took part in the European effort to draft the EU AI Act.

With France as a point of departure, we used a two-fold methodology to account for AI problematisations. First, relying on issue mapping methods applied to a corpus of news articles spanning over ten years, we inspected how French media narratives frame AI problems. Our results led to the identification of four typical forms of media accounts and we discuss their performativity at different scales.

Then, to gain plural perspectives on AI situations, we designed an ‘accounting dispositif’ to conduct a participatory inquiry with 25 French ‘AI practitioners’. After eliciting how AI is realised in their personal trajectories, the co-inquirers accounted for problematic situations encountered in the midst of their activities, along with their attempts to make these situations evolve. We use the French term souci (mix of care and concern) to coin both the affective dimension of felt problems within one’s experience and the concrete operations one engages in to transform a situation, involving a valuation process (Dewey, 1938; Quéré, 2012). To respect our co-inquirers’ voices in their own words, we will punctuate our analysis as much as possible with snippets of data, either verbatim or detailed scenes (like the dialogue at the very beginning). Thanks to plural and rich accounts, we were able to analyse how situated problematisations of AI weave together the construction of AI problems (their delineations and justifications), with the types of grips one has from their situated practices (constraints and resources, habits,…). We synthetised our results in a problem space and examine its usefulness to shape meaningful alliances and redistribute power dynamics. Note that all materials of this study are available on a dedicated website, ² to complement the synthesis proposed here.

We conclude by discussing how this two-step study contributes to (a) re-equipping a multi-scalar understanding of AI developments and (b) discussing how a participatory turn in the study of AI could enable a genuine reopening of its trajectory.

How media narratives account for and perform AI

‘AI is always inherently accompanied by narratives, fantasies, and promises. (…) For instance, the narrative I carried out as a journalist was that AI is powerful – whether positively or negatively. An example: an article I wrote in 2017 about the final match of AlphaGo, where I mentioned that DeepMind would focus on new challenges like curing diseases, reducing energy consumption, and inventing revolutionary new materials. (…) However, over the course of my career, I have seen the media’s treatment of AI change direction. We started to see topics like algorithmic biases and AI replacing or eliminating jobs. So, the narrative I participated in is one of AI’s power and people’s fatalism in the face of that power.’ (L., journalist)

Studies on the media coverage of AI argue that media play a major role in setting the agenda and shaping public opinions as they define expectations and issues associated with emerging technologies (Fast and Horvitz, 2017; Chuan et al., 2019). They contribute to perceptions of AI, by smoothing or roughening technological or scientific details, and amplifying voices and tropes while silencing others (Bareis and Katzenbach, 2022; Hansen, 2022). In line with these works, we started to investigate how media narratives account for AI and its problems. How is AI depicted and what entities are associated with it? How are they configured in practice and what agency do they have over concrete AI developments?

The term ‘AI’ encompasses multiple technologies and it gained traction because of its ability to bring together such a wide range of entities in a single discourse (Katz, 2017). Inspecting AI in this sense means being able to identify and disentangle from its floating and monolithic interpretation, the plethora of novel objects endowed with agency and capable of exerting influence in a specific social context (Latour, 2007).

We applied key principles of issue mapping (Rogers, 2013; Marres, 2017; Venturini and Munk, 2022) to analyse how AI has been both accounted for and constructed by the French press over the last decade. The national press is a well-defined initial source that constitutes a relevant proxy to define how and where different interest groups stabilise the nature, functions and future of techno-scientific objects.

Inspecting themes and their tonality

We created a corpus of press articles ³ using the 300 sources available on the Europresse service, comprising a broad variety of newspapers (general and specialised, national and local), and spanning the period 2011-2021 ( $n=46,000$ articles).

Previous research on technological innovations shows that media tend to report and amplify both the positive and negative impacts of AI (Nguyen and Hekman, 2022; Cave and Dihal, 2019). We took advantage of this amplification effect to generate a clearer picture of predominant AI issues in media accounts, according to both frames. So, we made a preliminary categorisation of our articles in terms of their tone: Whether they were intended to cast a light on AI benefits or its potential harms. Training an algorithm, ⁴ we obtained a first category (20% of the corpus) putting to the fore the ‘promises and benefits’, and a second one (9% of the corpus) expressing the ‘critiques and threats’ of computational technologies. ⁵ The final corpus comprises 13,000 articles, and while the articles presenting the ‘promises and benefits’ of AI outnumber those presenting the ‘critiques and threats’, their ratio remains stable over time, which contradicts previous studies conducted on fewer mainstream outlets (Nguyen and Hekman, 2022).

The corpus reflects 11 main themes, analysed with a co-occurence network. ⁶ (Figure 1-top part). A topological analysis shows that most themes (8) are closely connected, with at the centre, the theme that links AI to society. In contrast, themes related to health, education and justice are more peripheral.

Figure 1.

(a) Semantic network (6802 nodes (n-grams); 75,432 links (co-occurrences)). (b) Cluster projection by tone.

In terms of tonality, each term of the network got a score projected onto the graph relative to its frequency in the ‘promises and benefits’ or ‘critiques and threats’ sub-corpus ⁷ (Figure 1). We observe a significant polarisation of our corpus: themes associated with critique mainly occupy the right side of the network, plus a few localised pockets in other parts of the graph. Critical themes are primarily related to education, justice, defence/security, but also to ethical issues regarding the development of AI in society. In contrast, themes on the left side of the network are more closely associated with promissory narratives, mainly related to AI R&D in general as well as in specific sectors, i.e., healthcare, art, research, commerce, and finance.

These results, both in terms of themes and tonality, are in line with similar studies (Cools et al., 2022; Chuan et al., 2019), and generally, we note a widespread emphasis on the positive impacts of AI. More precisely, when compared to the Anglophone press (Crèpel and Cardon, 2022), themes such as digital labour or automated weapons are similarly prominently criticised in France, while others, such as autonomous vehicles or health applications, take on a positive tone. In addition, the peripheral position of some themes (health, law and education) may be a good indicator of controversies and issues specific to French cases. But, more interestingly, we complemented this first inspection with an analysis of the entities present in each cluster, which led to identifying genres of media narratives.

Four genres of media narratives staging diverse entities

Pursuing further the investigation of our corpus, we extracted and manually annotated the entities present in each cluster. The following categories emerged through an iterative coding process: Technical entities, data entities, people, public figures, companies, institutions and topics.

Although the clusters are not homogeneous in the way they account for AI, analysing the arrangements of the entities in each one revealed four typical narratives of AI, i.e. four genres. These four genres are organised on a two-axes matrix, from projection to realisation and from negative to positive (Figure 2). Exemplary snippets of articles illustrate how each genre arranges specific types of entities (Table 1). We then discuss their performativity on AI developments.(Figure 3).

Figure 2.

Four genres of artificial intelligence (AI) narratives present in French media, organised in a matrix (from projection to realisation and from negative to positive). The genres are described according to their clusters, topics and narrative structures.

Figure 3.

The performativity of four genres of media narratives.

Table 1.

Excerpts from articles present in our corpus, that illustrate four genres of AI narrative present in French media.

GENRE	EXAMPLE 1 (excerpt from article-our translation)	EXAMPLE 2 (excerpt from article-our translation)
A $\bullet$ Abstract Critiques	When Machines Surpass Humans (Les Echos-6/05/17) It will take more than 40 years for artificial intelligence to reach a sufficient level to compete with humans in all tasks, and over 100 years before it can take over all jobs. This is the result of a survey of 300 experts in the field.	Artificial Intelligence in Search of Ethics to Remain Safe for Humanity (Le Figaro-7/28/17) Ethical reflection is gaining momentum worldwide. It is regularly reignited by alarmist statements from Elon Musk, the founder of Tesla, who considers artificial intelligence as ‘the greatest risk our civilisation will face,’ or from the scientist Stephen Hawking.
B $\bullet$ Global Promises	AlphaZero, the Chess Prodigy Algorithm (Le Monde-12/20/17) The artificial intelligence algorithm AlphaZero from the British company DeepMind - acquired by Google in 2014 - had defeated Stockfish, the world champion of chess programs in 2016, which was far superior to any human champion. (…) This is the generic version of AlphaGo, the algorithm that over the past two years has demonstrated the superiority of machines over Homo sapiens in the game of Go.	Researchers Equip a Computer with Artificial Neurons (Le Monde-8/08/17) The machine designed by an international team recognised voices using a component that imitates the human brain. After algorithms beat humans in games like Go and poker, and machines excel at tasks such as image recognition, text translation, and autonomous driving, an international team has achieved new progress with a success rate of over 99.6%.
C $\bullet$ Promising Experiments	Diabeloop Designs an Insulin Pump That Automatically Calculates the Dosage to Administer (L'Usine Digitale-9/18/19) According to the French National Institute of Health and Medical Research (INSERM), the number of people with type 1 diabetes has been steadily increasing for the past twenty years, at a rate of 3 to 4% per year. (…) This is exactly what Diabeloop aimed to transform by creating DBLG1, a device that includes an algorithm capable of accurately calculating the amount of insulin needed to be injected.	‘Artificial Pancreas’ Tested in Evry to Improve the Daily Life of Diabetics (Le Parisien-5/08/16) ‘A glucose sensor (placed on the abdomen, for example), smaller than a USB key, constantly sends information to a terminal which, using a complex algorithm, calculates the required insulin dose for the patient. This dose is immediately injected using a pump,’ explains Erik Huneker. The goal is to provide more daily comfort for diabetics.
D $\bullet$ Local Controversies	New APB Procedure, a Politically Explosive Project (Le Monde-10/10/17) The postbaccalaureate admission procedure (APB) will change its name at the start of the 2018 academic year. It will still use an algorithm, ‘but we are working to find other solutions so that it is not the sole decision-maker,’ stated Frédérique Vidal on September 28 during her back-to-school press conference. The Minister of Higher Education was responding to the National Commission on Informatics and Liberty (CNIL), which, on the same day, issued a notice to the ministry to ‘stop making decisions concerning individuals based solely on an algorithm.’	Parcoursup Explicitly Leaves a System to Administer, Rank, and Order the Dreams of a Generation (Le Monde-5/31/18) The soughtafter optimality criterion is to maximise the number of registered individuals, rather than the actual satisfaction of the students.We are subjecting young French people to an uncertain experiment, moreover on a large scale, while betraying our own social promise of an open university. Parcoursup is in a way a poor technical response to the real issue of the lack of university resources.

Abstract Critiques, the genre of the top-left quadrant of the matrix, is strongly present in the network’s central themes and refers to broad and vague entities (‘civilisation’, ‘society’, ‘AI’, ‘cyborg’, ‘machines’). Its main narrative, presenting a direct and deterministic relation between AI and society, insists on potential risks of computational technology – loosely designated (‘transhumanism’, ‘ethics’) – to democracy and ‘humanity’. Social actors remain quite abstract, with little reference to French social and political environments and specificities; rather, prominent thinkers (Jeremy Rifkin), political leaders (Macron) or global institutions (‘Church’, ‘UNESCO’) are mentioned, although often, the main actants are not people but the technology itself or ideologies such as the neoliberal model embedded into AI development that is pointed out.

Global Promises, at the top right, associates quite specific technical and data entities (‘AlphaGo Zero’, ‘stockfish’, ‘neural networks’, ‘signal’, ‘brain’) with top international and French researchers in fields ranging from physics to cognitive sciences. In a similar deterministic fashion, the main narrative here revolves around the uprising of an autonomous AI, developed in a lab, able to compete with human intelligence. It stages a representative (a proxy) for both parties (IBM Watson versus champions Brad Rutter and Ken Jennings or AlphaGo versus Lee Sedol). Although we note the presence of few French business structures (‘Hub France IA’), the main actors of the genre are research labs of large IT companies (‘Google’, ‘OpenAI’) – or start-ups they acquired, and the technologies they develop. They sustain this narrative through regular and spectacular demos deployed in game-like or simulated environments.

Now in the bottom part of the matrix, narratives focus on specific technologies deployed in society and detail how they reconfigure local entities in richer and more complex ways.

On the right side, we acknowledge a genre encompassing themes related to health, web tech, finance, in which detailed narratives are built around innovative solutions and applications brought to the market that optimise domain-specific activities through computing technologies. In the case of health for example, many individual practitioners are staged promoting current AI developments (mainly devices or products from French start-ups), highlighting their benefits for specific individuals (‘patients’, ‘clients’). It relies on the authoritative voices of well-defined groups of specialists (French hospitals and research centres under the control of national institutions), while specifying data entities (‘sugar levels’, ‘insulin’, ‘hormones’) that feed AI models, justifying their collection and exploitation with regards to the alleged benefits.

With similar attention to specific local configurations, but focused on actual denunciations of technological negative effects, the last genre provides extremely rich and structured arrangements. Prominent in isolated themes in the network, constituted around specific French controversial cases (Education, Justice, Labour), such narratives account for the rare occasions of concrete political transformations. Groups of affected persons and networks of activists emerge to vocalise harming effects of computational technology from their embodied perspectives and tend to change their situations, through legal actions or other means (see the example of Parcousup, Table 1, line D). Such narratives stage counter-inquiries, complex reconfigurations and interesting reproblematisations of issues, that in some cases, manage to impact ongoing developments.

Synthesising media narratives’ performativity

The top of the matrix (Figure 3) characterises monolithic narratives. Both genres acknowledge and disseminate the existence of a powerful cohesive AI that will greatly disrupt ‘society’ and affect ‘humanity’, for the better or the worse. In addition, both project AI issues, i.e., effects and consequences in the world are speculated, anticipated. They generate both anxiety and fascination and play a strong performative role (Borup et al., 2006), mutually reinforcing each other. In critical narratives, problems are too often abstracted and translated into a set of moral principles (Jobin et al., 2019)., which echoes the ‘principled’ regulations (Mittelstadt, 2019), recently criticised, that might be found in ethical charters and guidelines. A similar abstraction plays out in the global promises: A wealth of technological objects for attracting and structuring money, people, and institutions is brought to the fore, for which issues are reduced to problem-solving processes, approached through games and simulations. In this context, performance is linked to the economy of promise (Joly, 2010), which legitimises the development and introduction of AI in society.

Highly particularised human actors emerge once narratives account for situations where AI is realised, embedded in specific national economies, research and industrial environments, legal frameworks, and experienced by the public. At the bottom of the matrix, the two genres of narratives perform in opposite ways, although mutually constitutive. One aims to persuade public opinion of transformative socio-economic impacts of computational technologies. This rhetoric is powerful in involving and aligning local actors and resources in the active reorganisation of social life through the development of experiments, prototypes, alpha versions and testings, ultimately turning the world into a laboratory (Marres and Stark, 2020). But testings provide fertile grounds for the emergence of situated problems and local controversies: they create the spill-overs and cases that we see in the fourth genre. Actual experiments generate troubles for those who are not involved and ultimately do not benefit from such AI realisations, leading to counter-inquiries that investigate systems functioning and constitution or even counter-actions, such as appeals and mobilisations.

These genres, largely controlled by big-tech and media players, grant existence to particular entities and arrangements and perform AI both as a global issue (AI monolith) and as specific configurations (AI situations). Each genre develops its own agency and puts in motion different operations (Figure 3): A moral agency that acts primarily on normative processes, a legitimisation one demonstrating AI’s power through controlled experiments, a persuasion agency aligning actors through testings ‘in the wild’, and a political agency, counter-acting from specific cases.

Reading the matrix vertically (Figure 3), the analysis of the genres also suggests that the issues of AI are constituted and dealt with through two main modalities: either through tests (in controlled settings or in the world) or through regulations of computational technologies (to prevent risks or condemn harming effects). If the first modality is mostly at the initiative of AI systems’ producers or investors, the second one engages more diverse stakeholders and mainly develops in policy arenas. These two modalities concentrate most efforts and compete to influence future developments. Neglecting other modes of participation, it constrains the vision of technological development to a binary and caricatured perspective (either pro-innovation or pro-regulation).

These wide-spread narratives create the main frames that account for AI issues in France, structuring the field of AI and stabilising its ‘thingness’ (Suchman, 2023). But, such predefined and limited views offer little room for other players to take part, and for genuine bifurcations in the developments of computational technologies. So, what if we were to develop renewed grip over AI trajectory?

A third path to participation towards problematisation

‘I believe there is a kind of third wave in AI studies. Some, in the social sciences today, following a techno-critical approach criticise part of Latour’s successors and consensus conferences…What someone like Fressoz is saying is that this generation of STS researchers acts as if, since Ulrich Beck, the issue of socio-technical risks was addressed. Fressoz is reminding us that modern societies, at least since the 19th century, and in the face of industrialisation, have always been reflective. There were struggles; they were just completely invisibilised by history. In fact, highlighting them and narrating this history is a way to remind us of the contingency of the order we are in.’ (Q., researcher and activist)

Escaping the double-bind of participation in socio-technical developments

Participation has gained traction (Magassa et al., 2017; Young et al., 2023), whether it be in AI research (Rahwan, 2018; Birhane et al., 2022), AI systems developments (Martin et al., 2020) or AI governance (Gilman, 2023; Tabassi, 2023; Lee et al., 2019). Numerous initiatives creatively implemented meaningful participation to address public relations with AI, and several literature reviews synthesised these efforts (DataJusticeLab, 2021; Delgado et al., 2023; The use of public engagement for technological innovation, 2021) – with careful accounts of ‘participation washing’ mechanisms (Ahmed, 2022; Sloane et al., 2022).

But despite a call for ‘more participation’, ways of seeing participation remain largely informed by normative and instrumental traditions (Chilvers and Kearnes, 2020). Inherited from two distinct disciplinary streams (political science and STS on the one side, PD and interaction design on the other), participation is described and analysed primarily through external categories (‘democracy’, ‘society’, ‘technical systems’) that tend to replay the classical divide between the social and the technical. This binary vision ends up engaging different voices either in policy-making processes (referred to as public participation) (Callon et al., 2001), or in the design of technological systems themselves (Bødker and Grønbæk, 1991).

Recent convergences in STS and PD research tend to bridge these two approaches by renewing with a pragmatist heritage. Focusing on socio-material practices (Marres, 2012), and insisting on the relational and co-productive dimensions of participation, which contrasts with conventional argumentative/deliberative perspectives, they aim at infrastucturing publics through collective inquiries (DiSalvo, 2022; Ricci, 2019). Such approach does not seek to probe the preferences of individual citizens, but rather to track the ever-evolving constitution of values and knowledges as experienced in a diversity of socio-material collectives, thereby addressing questions of justice, equity, and public accountability at systemic levels. A similar pragmatist perspective has been articulated in the French sociology of public problems (Cefaï, 2016; Chateauraynaud, 2022). This line of research unpacks why and how collective inquiries can help diversify ways of participating, insisting on relational processes, such as the inter-objectivation (Zask, 2004), i.e. the formation of common objects and their plural problematisation.

AI is a textbook example of the continuous co-production of the social and the technical: not only does it reconfigure the relationships between science, technology, and society, but it ‘co-opts the world.’ (Barocas, 2019). To understand this co-production, we advocate for an expansion of social science tools, in the vein of previous pleas (cf. Sociology of testing (Marres and Stark, 2020), Remaking participation (Chilvers and Kearnes, 2020)), towards a pluralisation of accounts. Collectively accounting for AI situations, examining and valuing diverse experiences with AI, will vary the forms and definitions of AI as an object of inquiry, casting issues in a new light. Questioning the consistency of an object along with its issues is what we refer to as problematisation. With this goal in mind, we engaged with experiential experts (Magassa et al., 2017) – AI practitioners – as co-researchers to account for and problematise AI from their situated perspectives. Embracing an interventionist and design-oriented stance, we propose to design a reflexive participatory dispositif that seeks to regain some grip on AI and its problems from a broad diversity of practices.

The loop of AI practices: A heuristic tool for collective inquiry

Given a broad definition of AI – a computational problem-solving method (model) that transforms diverse inputs (data) into optimal outputs (instances) ‘to achieve goals in the world’ (McCarthy, 2007), we chose to represent the diversity of data and computational practices that contribute to the continuous co-production process of the social and the technical as a loop (Figure 4).

Figure 4.

The heuristic loop of data/compute-intensive practices.

The left-hand side of the loop (from the world, where data is extracted and fed into computational models) situate the practices related to the problem-formulation phase of AI and to the data-intensive work. The right-hand side (from the models to their implementation into instances, such as products and interfaces, that feed back into the world) situate the practices associated with the deployment and integration of computational systems into the fabric of people’s activities and refers also to actual experiences with these computational instances.

This representation echoes a framework proposed and used by institutions like the OECD (OECD Framework for the classification of AI systems, 2022; Tabassi, 2023). Here, it does not serve as a descriptive model, but as a heuristic tool to represent in a shared space a plurality of situations, practices and operations through which AI is realised and circulates. As part of our accounting dispositif, it is a means to diversify problems while connecting them to their situated standpoints.

The soucis of AI: A situated problem space

Our second methodological operation consisted in a participatory inquiry to ground AI within situated experiences that contributed to the specific arrangements of computational developments in the French context. Our final objective was to collectively problematise these arrangements. Understanding at once problems and the situated life forms that shape them requires thorough and sensitive accounts, hence the need for an appropriate dispositif to intensify and discuss relevant parts of their experience.

Enrolling ‘soucieux practitioners’

One of our first challenges was to enrol co-inquirers. How to proceed? Who to target, using which criteria? Where participatory processes usually divide experts from non-experts, we were interested in engaging with ‘practitioners’ to focus on their experiential knowledge, without a priori defining specific activities or environments. As Seaver (2017) argued, it seems that no one feels legitimate enough to be fully associated with AI or to be called ‘AI practitioner’; it is always someone else’s core activity or skill. Adopting a reverse logic, we started to pay attention to a multitude of individuals who made their experiences with AI public. Looking for plurality instead of representativity, we carefully balanced several criteria with a main objective, reflected in the adjective ‘soucieux’: to meet with people already engaged in a sort of inquiry of their own, i.e., actively committed to sorting out, valuating, and problematising a situation that matters to them, where AI is realised, as to influence its development.

In addition to listing individuals mentioned in our news corpus, we used Twitter as a probe to iteratively refine our criteria: We built an initial dataset of 236 seed accounts, that we automatically expanded by crawling their list of followers. Tweets revealed a diversity of engagements (practical, reflective, investigative, critical,…) and variations in the intensity of people’s concerns (Table 2). Then, we also listened to suggestions made by our co-inquirers to complement our selection.

Table 2.

Examples of two tweets showing engagement.

Tweet A	Tweet B
‘WHY I LEFT THE SEXIEST JOB OF THE 21ST CENTURY. After 3 years of working at [xxx], a data scientist, resigns to better denounce the role of digital technology, big data, and AI in ecological and social devastation’	‘Hey, I’m conducting a survey to understand how data science teams work with sensitive data! Interested?’

Several criteria guided our choices: Activities, profession and status, gender, background, type and level of concerns. Many of these individuals have multiple backgrounds and activities in different settings, which enrich AI accounts. As Chateauraynaud and Debaz (2017) argue, the more actors multiply their positions, the more they have a ‘grip’ on a problem: they control networks and master, even shape, tests and regimes of justification. But, we deliberately excluded individuals who already benefited from a significant degree of media attention.

We reached out to a total of 56 individuals, out of which 25 (19 men and 6 women) eventually engaged in the research process, as anchor points to ground, historicise and re-problematise AI. The co-inquirers’ practices, all related to AI, are characterised by a striking degree of heterogeneity, e.g. reporting on the digitalisation of public services or pushing it, creating deep learning models applied to sound, building a community of AI developers, optimising AI models to reduce their environmental impacts, or studying regulations of AI (Table 3 details three of their profiles).

Table 3.

Examples of co-inquirers’ profiles.

Practitioner A $\bullet$ medical imagery	Practitioner B • automated video analysis	Practitioner C • critical art
A tenured professor and research fellow at one of the four French Artificial Intelligence Institutes, created in 2018 as part of the French national initiative on AI. She specialises in biomedical imaging and is interested in building models for brain tumour detection. Her approach emphasises the reliability of the data – clinical, research, synthetic – and their	An entrepreneur, CEO of a company developing artificial intelligence (AI)-based video analytics for urban management and security. He develops business opportunities with various stakeholders, including both public and private entities, and adapts digital products to meet their specific needs. Part of his work is concerned with AI ethical and legal issues.	An art curator and teacher, formerly a journalist. She considers critical digital art as an observatory for detecting weak signals of problems with technology. Her interest has led her to collaborate with digital artists who work at the intersection of art and science and develop a sort of counter-expertise infrastructures.

Grounding AI in practitioners’ experiences (first encounter)

According to a pragmatist perspective, our inquiry process sought the creation of a common object of inquiry (Zask, 2004). The first encounters consisted of two hours of individual, in-person, lively discussion about the co-inquirers’ personal histories. They started with a formal prompt: ‘Can you describe your activities related to AI and how you came to engage in them? How is AI realised within these activities?’ Some follow-up questions focused on co-inquirers’ attachments and concerns, to delineate current problematic situations. These conversations accounted for incredibly detailed life trajectories and, as the excerpts from Table 4 illustrate, biographical storylines revealed key episodes of the French AI history. They account for the entanglements of global and local elements interacting within milieux and networks, education, modus operandi and tools, funding mechanisms, norms, values, and so on, all rooted in situated activities.

Table 4.

Excerpts from conversation with co-inquirers.

Voice A

Voice B

‘And later, at Etalab, the idea of creating a data science team around the General Data Administrator emerged. [..] The genealogy of the General Data Administrator notion is interesting. It is the explicit French translation of ‘Chief Data Officer’ that was created under Obama. There was a strong inspiration from the work done in New York by Mike Flowers. They did very operational work to identify unhealthy buildings and where firefighters should go. […] So, I improvised as a data scientist within a small team of four people. We weren’t talking about AI at the time, it was something that was a bit over my head.’

‘I think it’s the geopolitical context: we’ve seen these technologies coming and changing policing practices, Palantir experimenting with predictive policing in the US in 2017, facial recognition being used in China. And the realisation that it’s happening in France and that we’ve completely missed it all this time, that there’s a real lack of debate and expertise and that we need to build it up. It all started in November 2017, we became aware of this project in Marseille; and later with a public meeting I organised in March 2018 to make the documents known.’

Our team extracted many concrete items from the transcripts: from events to papers; from projects to laws, and from data repositories to tweets and memes. We carried out extensive online research, looking for traces and evidence of these specific items, ending up with an archive of over 1000 documents that rematerialise AI history into concrete episodes. Figure 5 presents four of these documents to give a sense of their materiality and diversity.

Figure 5.

(a) ‘Blog post (Benesty (2016)) about judge impartiality. (b) Download section of CamemBERT website, a French language model, developed by INRIA and Facebook AI Research (Muller B , n.d.). (c) Portfolio of projects supported by the LabIA at Etalab (Portefeuille des projets - Etalab, 2001). (d) Personal certificate of an online course (MOOC) on Machine Learning proposed by Stanford University, taught by Andrew Ng. Source: personal documentation, date: 2016.

At this stage, we suspended any interpretation, focusing instead on the activation of these documents as ‘material to be used’ (Zask, 2004) for problematisation.

Problematising AI from practitioners’ soucis (second encounter)

For each co-inquirer, we printed out the documents gathered from their biographical accounts ( $\pm 45$ per person). During a second encounter, we invited them to explore a wall where these documents were displayed. After a moment, we prompted them: ‘You can choose one or more documents that, in your opinion, exemplify a problematic dimension you see with AI as it is realised in your activities’. They then brought these documents to a table mounted with a camera (Figure 6) for inspection (Figure 7). When they felt they had saturated the AI-related issues brought to light by one batch of documents, they went back to the wall and started a new thread of account by selecting new documents.

Figure 6.

Setting of the second meeting at the Cité des Sciences et de l’Industrie, with the wall of documents and the working table.

Figure 7.

Screenshots from 4 different video recordings, where we see the co-inquirers using (grouping, pointing, sorting out, discarding) the documents they picked from the wall of documents.

This setting (the wall, the documents, the table) offered the co-inquirers the opportunity to objectify elements that matter from their personal vantage point and address them as problems. We use the French word souci to designate the affective and practical processes aimed at getting a grip on a problematic situation. Se soucier – both caring and being concerned with something – attests that a situation, perceived as problematic or unsatisfactory, has triggered a movement towards an active transformation. Adding to the STS vocabulary (e.g. attachments (Hennion, 2004), issue and concerns (Marres, 2005) or troubles (Emerson and Messinger, 2012)), soucis accounts for both critical and transformative processes. Referring to mundane and situated experiences, it insists on the interactions between practical operations and forms of problematisation.

Our co-inquirers discussed how their soucis operate and configure situations, assessed their importance, and evaluated their consequences. They also detailed how they engage in transformative operations towards their resolution. Therefore, the whole dispositif acted as a trigger to put into motion a process of valuation (Quéré, 2012).

Grounding AI in intimate stories, schemes of activity, workflows, organisational routines and norms, such accounting dispositif offered new grips to problematise AI from its concrete performances: latent, problems became tangible; explicit, they thickened through thorough descriptions.

We ended up with 62 hours of video material. After an iterative coding process using a grounded theory approach in a video editing software, ⁸ 19 soucis emerged in the shape of roughly 12 minutes video-montages each. Five to ten voices echo each other in these videos, with clear cuts and rough transitions. This media strategy sought to highlight disagreements and gaps created by the heterogeneity of practices operating at different levels and domains. Composed in a dialogic spirit, our montages aim at forming collective soucis while highlighting the multiple and sometimes contradictory ways of problematising them. These compositions aim to grasp what makes these perspectives simultaneously possible, and foster discussions on the grounds that sustain them, despite their diversity and their contradictions.

Generating a problem space

How do co-inquirers problematise the indeterminate and ambiguous situations in which they participate where AI is realised? Qualitatively, we listed all acting entities of our 19 video montages and identified recurring patterns that reveal conflicting ways of problematising AI, all rooted in specific practices. Making use of the loop (see Figure 4) to map these patterns, dividing lines separate problematisation spaces that define simultaneously the problematic object itself, the range of acceptable actions to make it evolve and the values associated with them (the analytical process for one souci is presented in Figure 8).

Figure 8.

Coding scheme used on the videos – detail for the souci ‘greening AI’.

Repeating the operation for all soucis, we generated a problem space composed of four main types of divides, that actively influence to the unfolding of computational technologies. (Figure 9).

Figure 9.

French Problem Space of artificial intelligence (AI), derived from 19 soucis, and organised according to their modes of problematisation. (note that out of the 19 soucis, 3 are not represented here: 2 were mono-modal (‘Using AI for academic research’; ‘Investigating AI systems’) and one did not present any clear divides (‘Participating in AI’).

The first type divides the loop into three horizontal spaces. At the top, close to the model pole, an institutionalisation mode refers to practices aiming at establishing AI as a proper academic field. They are concerned with the delineation of AI as a discipline, and the elaboration of relevant metrics and benchmarks (including environmental ones) driven by the notion of generalisation. The middle space comprises practices concerned with AI production, i.e. the implementation and maintenance of AI-based applications on a large scale. Their soucis reflect infrastructural concerns, and are populated by legal, economic, material or social entities. For example, the souci of ‘datafying the world’ is approached from the perspective of consent, co-creation of specific datasets or open-data regulations, as opposed to debiasing strategies. Again, in ‘defining AI’, the production mode of problematisation focuses on the consequences that a definition would have – for instance, legally: the definition of AI in the new European AI Act will determine AI products regime and potential legal actions. Here, the purity or relevance of a scientific object is not so much at stake compared to the concrete production and management of functional systems. Finally, the third mode of problematisation seeks to deconstruct AI to shift prominent techno-scientific narratives and understandings of AI in order to draw attention to larger ecologies of computational logic. Deconstructing AI narratives entail investigating and unpacking discourses and systems, which can be made through archival practices, documentary research, fieldwork, reverse engineering, manufacturing one’s own datasets or systems. It is also concerned by means of publicising other narratives to reframe computation history (e.g. focusing on labour, or unpacking the joint development of State and information systems…).

The objective to develop further computational technologies is clearly shared and performed through practices that benefit from it – represented in the two first spaces, institutionalising and producing – which oppose them to the third space of practice. However, some trembling life transitions or deceptive transfers of technology from academia to real-life applications reveal the potential for unexpected alliances between the first and the third areas of practice.

The second type follows the same division, although the first space does not contribute to the discussion. The first two soucis of this category articulate theoretical considerations – respectively addressing the co-evolution of AI and society, and the intended role or impacts of AI studirefrefes – , while the three other ones are deeply situated and concrete. ‘Experiment’, ‘Develop for public services’, ‘Regulate’: They have in common to discuss the genesis of computational experiments, the logic and policies of funding bodies (public and private), the enrolment of actors, as well as the interplay between French regulations and the numerous exception mechanisms to bypass the common law. In that regard, the role of the State, both as regulator and experimenter, is a recurring issue. Two modes of problematisation stand out in these soucis: one focuses on assessing the risks of computer technologies for society (middle space) while the other seeks to re-politicise the issues (bottom space). Their positions are particularly contrasted when it comes to the means and goals of AI studies, towards an inflexion of socio-technical trajectories or towards a more radical questioning (Table 5).

Table 5.

Verbatim illustrating the second type’s fault line.

A $\bullet$ Risk evaluation

B $\bullet$ Politics

‘There is obviously a question about the place of critique in the history of digital technologies, since we are moving from a critique that was first internal to the very community that was creating the technology - the hacker culture - and now, given the success of technological projects, it is coming out of the world. So it is no longer a corrective critique, but rather a fundamental critique of the role of technological innovation in our societies. Bruno Latour’s entire work showed that we should not double-click and go from intentions to effects. We must rather show how intentions are translated into devices that are approximations of those intentions and that involve a whole series of transformations that change the actual state of the initial project and strategy. In a sense, this somewhat strategic vision of the social world in which we have no socio-technical mediation is a real danger.’

‘I became somewhat disillusioned. These are questions that have increasingly piqued my interest since I began to question what I was trying to achieve as an engineer, applying a lot of technical expertise in the service of the common good. What do we do with technology? What extent of technical expertise should I employ, in what context, and for what purpose? I don’t want the discussion to revolve solely around whether an algorithm is ethical or not, but to delve deeper into what is involved when dealing with advanced technologies: what does it imply, and what social organization does it presuppose? Engineering students are often targeted with a one-way discourse that promotes the idea that big data and Industry 4.0 are what society needs today and that this is how they can be useful to us.’

This opposition plays out in the three situated soucis. One argues for a careful assessment of specific systems through testing. The other denounces strategies of trivialisation, perpetuating a regime of exceptionality, despite threats to civil rights or social justice (Tréguer, 2019). This last line of argument, usually associated with activism denouncing state surveillance, suddenly sounds different when it comes from civil servants themselves, who describe the development of AI projects with an insider’s experience. Bringing these voices together opens up avenues for unexpected alliances and creates new ways of responding to the development of computing technologies.

The third type organises three spaces of problematisation. Practices that characterise the upper left-hand space are concerned with the creation of efficient and optimised models for a given set of tasks or problems, thus characterised design mode. On the other hand, the integration mode focuses on integrating tailor-made AI-based products into organisations, workflows and routines. A striking contrast between these two modes can be seen in how practitioners justify their engagement in AI projects: some refer to morality, while the others refer to instrumental considerations (Figure 10).

Figure 10.

Souci ‘Justifying AI projects’ – Extensive list of verbs enunciated by practitioners positioned in the ‘design’ space vs. those of those positioned in the ‘integration’ space.

A third space remains stable at the bottom of the loop. In strong opposition to the first two ways of problematising, it resists a hegemonic computational logic and advocates for other kinds of solutions carrying other values. For example, one co-inquirer mentioned Hito Shteyerl’s artwork ‘The City of Broken Windows’ to contrast data-intensive policing technology with delicate care, two alternative worlds we can choose to live in.

The souci related to the transformations affecting both professions and labour is a great preoccupation, although perceived differently. With a focus on new tasks (design) or on work management (integration), two spaces of practice are concerned with setting up new scenarios. On the contrary, accounting for reconfigurations from situated realities, as opposed to projections of designers and intergrators, is a mode of resisting these transformations. Table 6 provides verbatim to illustrate these three modes.

Table 6.

Verbatim showing the fault lines present in the third type (in the case of the souci ‘Working with AI’).

A $\bullet$ Design	B $\bullet$ Integration	C $\bullet$ Resistance
‘I don’t know how many different AIs we would need to make a doctor, but it would be a lot…Because one AI might be good at segmenting tumors, another AI might be able to distinguish between dementia and non-dementia, another AI might be good at performing a specific task…These are all highly specific tasks. For example, in this image, there is a lesion, and it may be useful to know the volume of the lesion. I think a doctor has better things to do than spend three hours outlining all the slices to determine the volume of the lesion.’	As much as possible, we try to have a human in the loop. There are many reasons for this, including responsibility. When I’m faced with an interface where I have to select from a list, there may be a bit less creativity involved. In the specific case of Project D or in working situations such as aviation, there are well-defined procedures in any case. However, the human element does bring something valuable, and I don’t think the question is whether workers are devalued in this case. In fact, we rely on their expertise as it is up to them to make the final decision.’	Colleagues do not agree with the new tasks that have been imposed on them. They are constantly stuck doing desk work and end up with a final product of lower quality. What is important for municipalities is regular updates of the cadastral plan with field visits by surveyors to discuss the rates and particular local tax…If we don’t have precise knowledge of the field, then it’s not the same in terms of public service. We are constantly reassured that there will always be human intervention. Yes, but which one? If it’s clicking a button, is it really human intervention?’

Lastly, the fourth type offers a distinctive segmentation: it divides modes of problematising vertically, between a ‘normative’ mode on the right and an ‘appropriative’ mode on the left. We could say that some are busy creating the conditions in which computing technologies develop, which involves framing narratives or establishing standards, norms and values; while others are living with and adapting to these technologies. Some develop strategies and hacks to benefit from them, as well as counter-narratives and unexpected ways of relating to them, although many account for a degradation of their living conditions. These two modes mutually reconfigure each other: For example, the souci ‘data sharing’ reveals a moving terrain where normative and appropriative perspectives shape each other. Indeed, new data cycles actively reconfigure relations, especially (but not exclusively) between public and private actors. In France, the ‘Loi pour une République Numérique’, adopted in 2016, established an open data framework that changes the relationship between public administrations and their data production and use. The State now has a new mandate: that of a ‘public data service’ (République numérique, 2019). It is responsible for making available its datasets with the greatest economic or social impact to facilitate their re-use. This legal framework, inherited from the European one, requires administrations to completely reconfigure their data value chains, leading to 15 ministerial data roadmaps published in 2021. In this way, the State, through its regulatory powers, is profoundly changing its actions, its instruments and its relationships with other actors, developing new partnerships to support the production and use of valuable data and models. Among these new forms of partnership, a third path has opened up between the public and private sectors with the notion of ‘digital commons’ that is gaining traction as a response to the privatisation of public services and as a means to share the costs of computational development and infrastructure – albeit controversially, as the Table 7 shows. The dynamics and openings between the two spaces again reveal potential for drawing new trajectories.

Table 7.

Notion of ‘commons’ as discussed in the fourth type.

A $\bullet$ Appropriation

B $\bullet$ Normative

‘The idea is indeed to do it in an open way, well…the notion of commons, I think it’s very important for AI. I think it’s not yet… We can say that we can do many things with Open Street Map, for example what we did with Open Solar Map. That’s one thing. But, there are many many things to do around it and I have the impression that the AI community is still a bit separated from the digital-commons community, because it’s a bit more business and a bit more research-oriented potentially. The main thing is still, as the Ministry of Economy and Finance says, to open up data to make it a strategic asset…’

‘We talk a lot about open data, which is a bit of a buzzword that actually means to favor the use of all data by certain people rather than others, by the richer, the more powerful, or by those who are more capable of developing technology…And the digital commons, which we see everywhere, particularly annoys me because there is a lack of definition of what it covers and the big problem behind it for me is a legal question…The commons is very beautiful, but when there is a problem, who decides, who is responsible? We see that these forms are very fragile, mainly because of the law and the economy that goes with it.’

Our analysis of the four types of soucis made visible problematisation fault lines, drawing different geometries from the same objects, and stressing political divergences.

Participatory Problematisation: A generative perspective

We now discuss our contribution to the fields of controversy and AI studies. In a way, our research invests participatory means to revisit a classic question anew: What and who contributes to problematise the development of technical objects – from accounting to framing their issues?

Thick accounts and power dynamics

Asking this question is especially relevant in the case of AI where big players largely control the controversiality of AI by spreading both fear and fascination narratives, while at the same time, presenting themselves as willing – and best equipped – to handle problems (Luccioni and Bengio, 2020). An empirical approach has the ability to desexceptionalise AI by resituating its developments in concrete histories and courses of actions. Our results suggest that rich and plural accounts of situated practices help renew and sharpen the types of questions that AI progressive constitution raises, thereby redistributing some grip to less prominent actors.

First, situated accounts (both from media narratives and from co-inquirers’ practices) avoid the reduction of AI issues to a set of off-the-shelf problematisations. Instead, they multiply the entities at stake and the situations in which they are embedded, adding layers of intertwined story lines. But in addition to a welcome ontological reconsideration of what constitutes AI, elicited soucis force a shift from binary and often sterile interpretations (good/bad, winners/losers, powerful/powerless) to careful reflections and nuanced statements.

Indeed, each souci rides the entire loop: It thickens by being told from orthogonal standpoints. Such thickness ‘unframes’ technical focuses that have been prominent in AI studies (Selbst et al., 2019; Ananny and Crawford, 2018; Amershi et al., 2014). Of course, some well-known and well-studied problems appear in a shallower way in this inquiry, but the soucis’ mouvement is not deepening but expanding. The soucis juxtapose practitioners’ accounts, making salient their diversity and frictions, thereby augmenting problems through their circulation along the loop and shedding light on unexpected or underrepresented categories of actors, entities and issues. ‘The only thing that is conceded to us [labour unions] regarding AI developments is the risk of algorithmic biases. They [the direction] are aware of it; they tell us every time. I think it’s more of a superficial discourse, systematically reintroducing algorithmic biases into the landscape rather than reflecting on other issues and on the shifts in the workflow chain. And they never explain to us what measures they will put in place to prevent biases.’ (D. union worker) This testimony shows how the well-known problem of biases becomes an alibi for project managers to maintain an asymmetric relation, depriving workers of their right to oppose. Contrasting standpoints and acknowledging circulations develop new layers of interpretation and open up dialogues among practitioners, in place of disqualifying or silencing some positions. Through these compositions, repertoires of arguments get richer and ways of problematising become sharper.

Secondly, echoing the feminist saying ‘the personal is political’, we argue that a focus on AI practices offers both a multi-scalar and generative understanding of AI developments, while departing – and because they depart – from intimate situations. More political in nature, the elicitation of valuation operations with concerned persons, which track back the genesis of problems within concrete operations, situations and milieux, as practitioners experiment, evaluate and structure their conducts, individually and collectively. Hence, they help identify the intricate elements that constructs situated (dis)empowerement, from inter-personal relations to large economic or social forces. Such approach offers a vantage point to account for both critical and transformative processes that construct AI as an object. Therefore, the soucis help keeping track of the permanent reconfigurations of relationships among actors (especially between the State, industries, and markets) and the progressive establishment of renewed values (political, economic, or social), while reconstructing how some maintain power asymmetries and control AI developments. Politicising AI experiences (Jaton and Vinck, 2023) and reclaiming AI problematisation are both needed to reweave the public problems of computational technologies into the fabric of political and social structures. Accounting from practices contributes to disarming the rhetoric of scales, which often operates as a delegitimisation tool. Departing from experiences authorise anyone to reclaim their ability to investigate and account for multi-scalar dynamics at play in AI developments.

Here, we need to acknowledge that the demonstration of thick accounts’ potential reaches one main limitation: the richness and density of each souci makes it hard to synthetise without loosing its force, which lies in a subtle work of composition. This paper presents a problem space to grasp the fruitful agonism of AI problematisation, but the video materials (available online) are more truthful to the spirit of our collective entreprise.

Now, about the generative dimension. The problem space we drew is not case-specific but bridges and contrasts pluri-situated experiences. Collective inquires, when designed to take advantage of these contrasts, have the potential to both reveal and reshape the network of perspectives that organises the very experiences of the problems at stake – thus, redefining them. Being attentive to others’ standpoints, co-inquirers could engage in unexpected alliances, and thereby generate public problems that stand a chance to matter in socio-technical unfolding. This leads us to our future work: ‘infrastructuring’ (Karasti, 2014; Le Dantec and DiSalvo, 2013) a public of concerned practitioners over time to go beyond short-term engagement and sustain vigilance, update AI problematisations and alert on specific developments.

A shared and pressing participatory concern

Despite the fault lines of AI problematisation, there is a common concern that emerged from all co-inquirers’ perspectives: the lack of ‘democracy’ or ‘participation’ in AI developments. It is a transveral souci, the only one we could not divide and relate to specific spaces of practice. Rather, it accounts for unfruitful attempts to escape the binary vision described before, i.e. seeing AI developments primarily as a regulatory issue or as a testing issue (‘should we’ vs. ‘can we’?).

A case reported by one co-inquirer offers a striking illustration: In 2020, the Ministry of Justice announced the experimentation of an algorithm for assessing bodily injuries based on the automated processing of judicial decisions (‘DataJust’ project). It was first authorised by a decree from the Prime Minister and the Minister of Justice, but caused an outcry among legal professionals. These critiques remained unheard, and led to several associations filing an appeal before the Council of State, pointing out the violation of privacy laws among other issues. The Council of State validated the decree anyway, but the experimentation project was eventually abandoned by the Ministry itself due to a lack of technical and human resources to implement it.

Problematic indeed in terms of participation, this situation does not necessarily lead to similar focuses. Some are worried about tests being conducted into society without any form of consent or consultation, while others advocate for greater means to conduct meaningful tests (instead of limited or bugged POCs). Some rather put into question regulatory mechanisms themselves that eventually, even when experiments are blocked at first, always extend further the legality of systems threatening fundamental rights. Although they point to different mechanisms, all bring to the fore pressing concerns about the loss of genuine democratic processes, i.e., the possibility to have a say regarding technological innovations and their heavy infrastructures. References are made to other technical innovations for which explicit and consensual refusals have hit a wall; for instance, the case of 5G network in France, which development was rejected by the Convention Citoyenne pour le Climat but nonetheless approved by the Government, reinforcing a general distrust in public institutions and decision makers.

If experiments are presented by some as a risk mitigation tool, others have documented their lack of evaluation. The opaque but important presence of large international consulting firms is seen as a concerning loss of public agency and adds to this institutional distrust. Pushing the development of computational technologies forward in an accelerated pace, these companies exert a major influence on AI trajectories, as they act altogether as lobbyists, strategic advisors to both public and private actors, educators, and solution providers (let alone the widespread phenomenon of high-ranking civil servants joining them).

In line with previous work that analysed social acceptability as one crucial objective for actors who oversee the development of technologies (Angeli Aguiton, 2014), our co-inquiry documents that any form of reservation, resistance, or protest is invisibilised and disqualified as radical by dominant public discourses and public authorities. Even legitimate demands for information are very often dismissed. In France, we witness an increasing inefficiency of traditional counter-narratives and counter-actions (Chateauraynaud, 2022), reinforced by a severe repression of militants, and this general feeling of loosing grip is very much present when it comes to AI developments (see for instance, the unprecedented approval of facial recognition in France).

In face of such a big challenge, we are humble about the potential effects of our participatory endeavour. But, this novel participatory experience in France has given some room to discuss anew the development of computational technologies among the practitioners that we are. It calls for a field of open counter-inquiries that will hopefully ripple, starting within the situations co-inquirers inhabit and that ultimately shape AI.