Contenu du sommaire : Contrôler l'intelligence artificielle ?

By investigating the debates around the issues of AI regulation, we found that definitional problems were at the heart of normative conflicts on the means of subjecting AI to “social control”, whether technical, ethical, legal, or political. By taking the varied significations of AI as a common thread of analysis, this article contributes to the understanding of normative tensions over its control. We propose a mapping of places, actors and approaches that shows how the debates around the control of AI are structured according to four differentiated normative arenas: transhumanist speculations on the dangers of superintelligence and the problem of controlling its alignment with human values; the self-accountability of researchers developing a science entirely devoted to the technical certification of machines; denunciations of the harmful effects of AI systems on fundamental rights and the control of power rebalancing; finally, European regulation of the market through the control of security in relation to AI products and services.

Radiology was one of the first medical sectors to be directly concerned by the arrival of devices labelled “artificial intelligence” for image processing. This evolution is largely invisible insofar as it concerns tools intended for professionals. It raises the question of the conditions of regulation of these innovations, in relation to both the existing rules of the health system and the necessary adaptations for these new technologies. Based on a survey on radiologists, the industry, and representatives of the health authorities, this article shows that this regulation is largely delegated to the players in the field. Current regulation is the product of the normative work of radiologists, who are the users of medical devices defending the conditions of their profession, and the companies that develop them and have to rely on doctors' collaboration. The public and political debate on the regulation of AI in the health field is put second to the practical considerations encountered by the stakeholders. The latter revolve around the delimitation of the radiologists' profession and the competition between the traditional manufacturers of imaging devices and the new entrants bringing in digital innovations.

This study examines if and how a high standard of judicial protection can apply to litigants in cases dealing with harm caused by autonomous and opaque decisions made by Machine Learning (ML) systems. Based on a critical analysis of the procedural means offered by the EU's Product Liability Directive (85/374), this study makes two normative claims. First, that a fault liability (rather than a strict liability) model should govern the evidence adducing and assessment, regarding the proof of accountability in cases of AI-related harm. This model appears to be more procedurally fair given that programmers, users and developers of ML systems would have the possibility to rebut the—currently irrefutable—presumption of human agency, by proving that in causing harm, an algorithm had acted alone. Second, the study argues that, for the purpose of compensating AI-related harm, two criteria should be applied: on the one hand, for ML systems presenting so-called notorious risks of harm (i.e. types of harm that have already occurred in practice), the duty of compensation should lie with the programmer, user or deployer who accepted the risk of such harm materializing. On the other hand, cases of harm that have not yet occurred in practice could qualify as force majeure and could, therefore, be compensated through insurance schemes set up for the purpose of not leaving victims without compensation, when it is plausibly proven that such a harm was authored by an ML system without any human intervention.

An open letter published in 2015, signed by Elon Musk, Stephen Hawking, Steve Wozniak and Noam Chomsky, warned of the existential risks facing humanity as a result of new developments in AI. This letter constitutes a milestone in the reactivation of a set of narrative and normative resources aimed at bringing a critical debate on the progress of AI into the public arena. Successive waves of promises and warnings about artificial intelligence have occupied public debate in a spectacular way. Using NLP methods applied to a large corpus from the English-language press, this article shows that the focus on algorithms and AI has been growing steadily in the media sphere over the past five years. The corpus is structured around two semantic spaces that constitute two predominant regimes of critique, one based on the injustices produced by algorithms, and the other on fears of AI's and robots' autonomy. A comparative analysis of these spaces shows that they mobilize distinct technological and human agents, issues, and time frames. By developing a critical discourse on the damage caused by these technologies, the media sphere contributes not only to shaping public opinion but also to defining the forms of acceptability of these calculating agents.

The sale of Montreal-based Element AI to Service Now Inc. in November 2020 took the Quebec and Canadian artificial intelligence ecosystem by surprise. How could the start-up achieve narwhal status—that of a Canadian company with a capitalization of more than a billion dollars—and be celebrated by the government, the media and the business community, only to be bought out a few years later “for a song”? In this article, we present the case of Element AI, including both its rise and its fall, as an ideal-typical “cyberization of power” in which regulation is facilitative, remote and able to perceive control and communication as two ends of a single feedback loop. While the quick emergence of Element AI is characterized by its quest for “super-credibility”, expressed through partnerships with authoritative actors and institutions, its collapse is the sign of a desynchronization which met with rebukes and contradictions, even from the government. We have much to learn from this shift from justification to criticism since it points towards a vacuum within the Montreal AI ecosystem, as well as its difficulty in projecting itself even into a near future.

Based on the case of a food delivery platform, this article reveals the ethical problems posed by the application of algorithmic systems and explores the actions taken by platform workers to ensure that their rights are respected in the difficult political context of China. The author draws on a seven-month field survey in three Chinese cities and an analysis of official Chinese government documents. Outside the platform system, delivery workers facing these work-related ethical issues develop informal solidarity against the algorithmic system and the companies. They build communication spaces through social media to organize actions, and they find strategies of resistance, such as turning off their mobile phone positioning system (Diaoxian) at work and, in case of bad weather, launching strikes to counter the system. Sometimes the delivery workers' protest shifts from collective to individual action. These informal practices of the workers, who associate the algorithmic system with a “socio-technical controversy”, are a first step towards government regulation of algorithmic systems in China.

How do individuals classify and organize digital cultural goods? Where physical goods are concerned, classifying DVDs, video games or CDs by size, format, genre, author, or order of consumption corresponds to forms of engagement that support the consumption of cultural goods and give them meaning. When it comes to digital goods and there is no longer any classification to be established by oneself, in what ways are the different levels of engagement manifested in practice?We draw on more than 60 semi-structured interviews and on data from a questionnaire administered to over 2,000 respondents using the quota method. The article shows how individuals set up attentional architectures to arrange cultural goods in stocks and flows. These practices are changing and are increasingly moving towards idiosyncratic forms of classification where the individual is the main criterion of classification practices. This is more a question of arranging digital goods according to the personal management of one's consumption than according to institutional criteria (genre, date, alphabetical order).

Although little is known about the second-hand book market, which is therefore difficult to quantify, it plays a fundamental role in the circulation of printed cultural goods. Since the late 1990s, its retailers and operating rules have been destabilized by the platformization of this market space, materialized in the rise of new entrants and the use of computer technologies for profit-making. Based on a long-term qualitative survey of supply-side actors, this article examines the evolution of retailers' merchandizing skills by looking in depth at one of the essential stages in the creation of profit-making opportunities: the construction of a selection of products. The two processes that run through this operation—its decommodification and digitization—call into question the prevailing conditions of circulation of these second-hand cultural goods.