Écologies des pratiques LLMs
How to understand, describe and account for the system of practices around LLMs ?
The project Ecologies of LLMs’ Practices (EL2MP) is dedicated to creating critical situations and instances for studying the frictions that emerge from the interaction between LLMs and society.
Public discussions about generative models such as Large Language Models (LLMs) often subsume them as vague as Artificial Intelligence or reduce them to mere interfaces like conversational agents. This oversimplification frequently results in polarized perspectives. On one hand, there are enthusiastic claims about their potential to catalyze transformative shifts across various domains. Conversely, a chorus of critical voices expresses concerns about their associated risks and harms. Moreover, these discussions tend to treat LLMs in isolation, detached from the intricate network in which they operate and the diverse contexts in which they are intended to be utilized. It is crucial to recognize that, akin to any other object, LLM systems don't exist or develop in a vacuum. Instead, they form a complex and specific network of technological, legal, institutional, human, and infrastructural components. Questions regarding reliability, authority, adversarial use, copyright, safety, and trust, among other concerns associated with LLMs, must be approached within the intricate networks they operate in rather than from a detached and abstract standpoint.
To address these questions, the project Ecologies of LLMs’ Practices (EL2MP) is dedicated to creating critical situations and instances for describing and studying the frictions that emerge from the interaction between LLMs and individuals, disciplinary and professional groups, and, ultimately, society as a whole. The project's objective is to meticulously examine mistakes and failures, and discern the harms and risks associated with the development, use, and outcomes of LLMs in various situated and specific contexts. It will interrogate how users relate to LLMs’ outcomes in terms of:
- Hesitation, e.g. whether they view them as harmful results or as sources of creative inspiration, whether they categorise them as errors or report them as bugs;
- Iteration, e.g. strategies and actions for designing prompts or conducting result verifications.
- Negotiation, e.g. how they are integrated into routine tasks and responsibilities within various professional domains to satisfy specific and situated expectations and requirements, especially moral ones.
The project will build upon a series of experimental re-enactments of disciplinary and professional situations in which LLMs are currently used or have potential applications.
Leveraging on the unique context offered by SciencesPo and its dedication to promoting innovative intersections of social sciences research and experimental pedagogy, the project will engage final-year students from a range of master's programs at SciencesPo (e.g. Environmental Policy, Economic Law, Journalism, Communications, Media and Creative Industries, Urban Planning, and Sociology) These students will participate as both contributors and co-inquirers in the workshops, collaboratively to produce artefacts belonging to their different disciplines and practices (e.g. drafting a newspaper article, composing a policy report, or authoring a scientific paper).