Click the star to add/remove an item to/from your individual schedule.
You need to be logged in to avail of this functionality.
Log in
- Convenors:
-
Nicolas Chartier-Edwards
(Institut National de la Recherche Scientifique)
Jonathan Roberge (National Institute of Scientific Research, Canada)
Etienne Grenier (Institut national de la recherche scientifique)
Send message to Convenors
- Discussants:
-
Nicolas Chartier-Edwards
(Institut National de la Recherche Scientifique)
Michael Castelle (University of Warwick)
Etienne Grenier (Institut national de la recherche scientifique)
Maud Barret Bertelloni (Sciences Po Université Technologique de Compiègne)
- Format:
- Traditional Open Panel
- Location:
- HG-02A37
- Sessions:
- Tuesday 16 July, -, -
Time zone: Europe/Amsterdam
Short Abstract:
The hype around neural networks reorganised the field of AI research in terms of both epistemic culture and political economy. What are the emergent controversies of the contemporary AI paradigm and how newly developed STS methodologies can benefit the rising field of critical AI studies?
Long Abstract:
There exists, in Science and Technology Studies (STS), what could be qualified as a first wave corpus of work on research in artificial intelligence (Woolgar, 1985; Collins, 1992; Forsythe, 1993; Suchman & Trigg, 1993). However, the paradigm shift from symbolic to connectionist AI — and in turn to multibillion-parameters large language models (LLMs) — has reorganized the field in terms of both epistemic culture and political economy, each of which has strongly transformed AI research practices in different and/or overlapping ways.
As new trends in synthetic data and automated dataset labeling jeopardizes the very idea of “ground truth” so crucial to traditional supervised machine learning, the dominant contemporary AI discourse is one which seeks to monopolize resources within corporations and depends on an industry-dominated gatekeeping system that redefines what is valuable in terms of research. However, this gives rise to controversies as non-profits, minority actors, and other independent research communities try to make their own statements in the field. How can we understand the alternative discourses unfortunately submerged by the narrative of the major AI research institutions?
This reactivates the need for a “strong” STS programme that studies the mutations of research practices in the field of AI: Where are contemporary AI models made, how do they come into existence and come to be deployed and dwell in societies? The recent Shaping AI international research consortium in Canada and UK/Europe has studied the connectionist paradigm through the lens of controversy analysis, but in a context of controversy attenuation, how can both classical STS methods and new digital methods open the supposed black box of neural network research? Is it possible to foster public engagement while marginalized and/or dissonant technoscientific narratives are buried so deeply that they appear resistant to classical controversy analysis?
Accepted papers:
Session 1 Tuesday 16 July, 2024, -Paper short abstract:
‘AI doomerism’ deals with control, namely that i) discourses and dramatic performances about the existential risk of losing control to machines simultaneously ii) relates to how control plays out within the research field of AI in terms of who’s capable of steering its entire development.
Paper long abstract:
The release of ChatGPT came with a resurgence of ‘AI doomerism’. “Risks of extinction” now deals with control: namely that i) discourses and dramatic performances about the existential risk of losing control to machines simultaneously ii) relates to how control plays out within the research field of AI in terms of who’s capable of steering its entire development. My argument is threefold. First, I depict Bengio’s take on Xrisk, for it is a vocal attempt at re-thinking social engineering. From there, the second part of the argument shows how AI doomerism has been met with resistances, e.g. the position of LeCun who has argued that doom is “preposterous” as well as “political”. Which brings the third part of the argument: the profoundly conflictual nature of the current debate shows how “AI is a registry of power” (Crawford, 2021), even if —or precisely because—such a reality is obfuscated in the debate. In fact, this debate is not two-, but three-sided as a lesser-known faction tries to have its voice heard, i.e. the DAIR-Parrot community, represented, among others, by Gebru and Bender. Their position is at times portrayed as insisting on ‘short-term’ biases, but this is only one aspect of a broader critique addressing the issue of control. The problem, in turn, is that such subtler and sociologically accurate perspectives is what make them unable to shape global conversations on AI; something which also raises the question of the very possibility of Critical AI Studies as a field of inquiries.
Paper short abstract:
The Stochastic Parrot controversy highlighted the emergence of flawed hermeneutical machines fostering the automation of interpretation. Interviews conducted with technoscientific field specialists reveal the fractures left in the wake of the deployment of LLMs outside their labs.
Paper long abstract:
Computer scientists Bender and Gebru called out Large Langue Models (LLMs) as stochastic parrots. Pulverizing records in state-of-the-art language comprehension benchmarks, LLMs have cracked human syntax while lacking a deep semantic understanding of their constructs. The emergence of this class of hermeneutic machines built on statistical inference (Roberge & Lebrun, 2017) announces the automation of interpretation. Meaning is now disconnected from its formal manifestation, as training data is only concerned with the latter. Now confronted with the growing acceptance of invalid statistical relationships in the name of practicality and the neglect of a deep theoretical understanding of the problems it pretends to solve (Jones, 2004), the AI research community is faced with a sensemaking crisis. If “we have been led down the garden path (Bender, Gebru et al., 2021),” where are we standing now and how can sense be made in this strange landscape?
Through a series of interviews conducted with computer science experts distributed across Canada, the Shaping AI initiative researchers collected data that could offer potential answers. Following the foundational ethnographic work accomplished in the AI research communities (Forsythe, 1993; Hoffman, 2017), we caught a glimpse of how these experts reflect upon their research practices. Our preliminary results suggest the existence within the scientific community of what could be construed as an hermeneutic malaise fuelled by the neglected issue of sensemaking in AI systems. We argue that sensemaking is the core element that must be addressed to avoid the further development of this malaise into a full-blown crisis.
Paper short abstract:
AI as a techno-cultural object is the site of conflictual interpretations on the material implications of AI research. Results of the Shaping 21st Century AI project reveals a latent discontent in the Canadian research community around a form of industrial gatekeeping translating in GPU hoarding.
Paper long abstract:
The 21st-century paradigm of machine learning (ML) artificial intelligence (AI) (Roberge & Castelle, 2021), ignited by the crowning of neural networks (NN) in the image recognition contest of ImageNet, prompted a certain re-engineering of research practices in the field computer science. The “paradigmation” of ML and NN unleashed new possibilities for model scaling based on “increased computation” through graphical processing units (GPUs) (Sutton, 2019, p.2), coupled with massive “resource allocations” (Roberge & Castelle, 2021) and data harvesting (Cardon & al., 2018). This upscaling in the means of computation contributed to the publicization of the now (un)famous conversational large language model (LLM) Chat-GPT.
Shifting from the outputs of LLMs to demystify (Roberge & al., 2020) and study AI’s “thingness” (Suchman, 2023), we ask ourselves not “can language models be too big?” (Bender & al., 2021) but rather “how large models transforms computer science’s research practices”? The Shaping AI Canadian team aimed to revigorated the study of AI laboratory ethnographies (Forsythe, 1999; Hoffman, 2017; Seaver, 2022) through 20 interviews and uncovered, in a serendipitous way, the existence of a “shadow controversy” amongst labs located in the non-funded Canadian research circuits.
The issue seems to oscillate in the inability to access sufficient computing power to do research and the current hype on large models that contributes to a devaluation of fundamental research and toy problem experimentation. GPU concentration the industry is denounced as a gatekeeping that shapes the legitimity of “good and bad” AI research.
Paper short abstract:
What do the problems of AI look like when accounted for from multiple practices? We present two methods for engaging AI practitioners in the study of AI controversiality, and we discuss the implications of these participatory approaches for the STS agenda, in terms of epistemology and methods.
Paper long abstract:
This study addresses how the production of accounts shapes the problems arising with AI technologies. As STS scholars have long shown, framings of technological problems and controversies play a key role in technological development. In the case of AI, we argue that there is a lack of participation in its problematisation, fostering a disempowering feeling towards its developments. We therefore sought to engage people participating in AI development in its account. What issues emerge when studying AI problems from a plurality of situated practices? What methods can this type of collective effort develop, aiming to what effects?
As part of the French Shaping AI project, we developed participatory approaches to engage “AI practitioners” in two different co-enquiries. The first gathered 25 French AI practitioners, grounding AI issues in the sensitive and practical dimension that sustains them (Ricci, Crépel and Gourlet, 2024). This allowed to elicit 17 problems across practices. The second addresses the problems arising in the applicative context of “Foncier Innovant”, a project to automate the land register at the French Ministry of Economics. Here, practitioners’ issues concern the transformation of cadastral work and its organization, the lack of administrative transparency, externalization, maintenance, matters of governance and technology evaluation in public administration.
These two experiments, their different scales and modes of enquiry, illustrate different ways in which STS can develop methods to diversify problems associated with AI. We argue that, weaving AI problems into the fabric of its deployments, such a participatory approach holds potential to re-politicise AI.
Paper short abstract:
Despite abundant Explainable AI (xAI) literature, controversy is notably absent. Our study categorizes 12,000+ xAI papers, uncovering heterogeneity along technical, empirical and ontological dimensions. We urge a renewal of STS methods to build a critical discourse on this field.
Paper long abstract:
The upcoming successes of deep learning based systems in critical fields (medicine, military, public services) is conducive to serious concerns on the interpretability and accountability of its outcomes. Therefore, the research production on “Explainable AI” (xAI) should raise considerable scientific controversy and social debate.
In contrast, this communication emphasizes the actual almost non-existence of controversy emerging from the development of the xAI literature. Even though, in 2016, DARPA’s “Explainable AI program” was followed by a sudden appearance of scientific publications on xAI, those generally framed xAI as a technical problem rather than an epistemological and political one.
Exploring this paradox between an abundant literature on xAI and an absence of controversy, we intend to open the black box of self-appointed AI-explainers. Our presentation thus urges a renewal of STS methodologies to establish a critical typology of xAI techniques. Our methodology was twofold: we first systematically categorized 12,000+ papers in the xAI research field, then proceeded to an analysis of the mathematical content of a representatively diversified sample. As a first result, we show that xAI methods come considerably diversified. We summarize this diversity in a 3-dimensional typology: technical dimension (what kind of calculation is used?), empirical dimension (what is being looked at?) and ontological dimension (what makes the explanation right?) standpoints.
The heterogeneity of those techniques not only illustrates disciplinary specificities, but also shows that the research field on xAI progresses rather autonomously and opportunistically with primary objectives to fuel market strategies and answer the institutional demand for explanation
Paper short abstract:
We survey the current state of open source AI and show that open data, models, and products are a critical first step towards enabling future critical AI research such as probing bias in actual training data (Birhane et al 2023), as well as new data-driven STS methodologies.
Paper long abstract:
When Open AI pulled the plug on a dozen of their language models in January 2024, they rendered more than thousand research papers that used the models unreproducible, devalued, and devoid of core scientific principles (Liesenfeld et al 2023). The incident is a forewarning of what AI research may look like in a world of industry-led, proprietary, closed source technology. The solution is a healthy open source AI ecosystem. Our survey of open source generative AI tracks efforts to build more open, transparent and accountable alternative to the likes of ChatGPT and DALL-E (Liesenfeld and Dingemanse 2024). In this contribution we show in two case studies how open technologies can spark new controversies and enable new research methods (Llama2 vs Big Science Workshop's BloomZ, and DALL-E versus Stable Diffusion).
We argue that open data, models, and products are a critical first step towards enabling future critical AI research such as probing bias in actual training data (Birhane et al 2023), as well as new data-driven methodologies such as controversy mapping or EMCA studies on AI (Mlynar et al 2024).
References
Birhane, A et al, 2023. Into the LAIONs Den: Investigating Hate in Multimodal Datasets.
Mlynář, Jakub, et al. forthcoming. AI within situated action: A scoping review of ethnomethodological and conversation analytic studies.
Liesenfeld, A. et al, 2023. Opening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generators.
Liesenfeld, A., Lopez, A. and Dingemanse, M., under review. A European Open Source Generative AI Index under the EU AI Act.
Paper short abstract:
This paper will explore the controversy around “anthropomorphic” AI as staged by in expert discourses and compare this to how communities of practice stage their experiences of anthropomorphic-AI crises. It will explore a praxeological approach as a way to “unfreeze” controversies around AI.
Paper long abstract:
Anthropomorphic AI has been controversial in the computer science community since the Turing test (Natale 2021). Since the advent of LLMs, experts across academia and journalism have criticized, problematized, and occasionally advocated the anthropomorphization of AIs (Li and Suh 2021). These discourses concentrate on the public understanding about AI’s technical workings, cognitive-behavioral outcomes, and sometimes even the merit of ascribing human-like qualities to machines. While experts stress the stakes of anthropomorphization in ethical AI, there has been little consensus on what it means in situated practice (cf. Marres 2020). This paper will use digital methods to conduct two comparative controversy mappings. First, it will map the expert discourse around the anthropomorphic AI controversy across popular science, popular journalism, and academia. It will ask: what risks, challenges, and stakes do experts stage around anthropomorphic AI? What solutions do they propose? Then, it will compare this controversy map with existing research conducted on a specific crisis of anthropomorphic AI – the Replika AI sexualized content crises. This controversy map stages user, popular culture, and corporate staging of the risk of a situated instance of anthropomorphic AI. By comparing these two findings, this paper will demonstrate the distance or alignment between the expert voices prioritized in popular AI debates and the communities of practice actually engaging with it. It will demonstrate how a praxeological (Burkhardt et al. 2022) approach can help unfreeze (Dandurand, McKelvey, and Roberge 2023) controversies about AI and challenge AI’s “uncontroversial ‘thingness’ ” (Suchman 2023) by situating it in its ethnographic context.