Memória discursiva computacional: entre a consignação e o silenciamento do feminismo nas textualidades e visualidades algorítmicas em modelos de linguagem de larga escala

Abstract

This thesis aims to propose, develop, and demonstrate the concept of computational discursive memory through an analysis of the linguistic and discursive functioning of large language models (LLMs), with particular attention to ChatGPT and Gemini. In a context in which these models have become central actors in the production of textual and imagistic utterances, it is urgent to investigate how they operate, non-neutrally, upon discursive archives, and to what extent they reproduce, silence, or divert meanings. The study begins from the hypothesis that the language generated by artificial intelligence systems, although grounded in mathematical–statistical techniques, configures a form of discursive memory that is analogous, though not identical, to that defined within the tradition of materialist Discourse Analysis (Pêcheux, 1995, 1997, 2006). This memory, rather than merely storing data, functions as a condition of possibility for saying, enabling the emergence of articulated effects among historically inscribed utterances. By examining whether it is possible to speak of a machine memory that operates discursively, the thesis offers an innovative approach at the intersection of linguistics, computer science, and critical technology studies. The increasing adoption of artificial intelligence systems across diverse social domains has naturalized the idea that machines “understand” and “respond” linguistically. Yet this operation is grounded in massive volumes of linguistic data that bear historically constituted discursive formations and function as archives. In this sense, artificial intelligence does not “create” in a vacuum but draws upon what has already been said. The research problem, then, lies in investigating and analyzing: can memory in LLMs be conceptualized as discursive? If so, how this concept contributes to understanding artificial language as a symbolic and ideological process. The justification is grounded in the recognition that these models operate on pre-existing, socially situated utterances, systematically recombining sequences of words that reactivate historically sedimented meanings and discursive positions. By challenging the traditional conception of memory in computer science, restricted to technical storage and retrieval, the thesis proposes a critical, political, and epistemic expansion of the concept. Furthermore, as a researcher, woman, situated within a research group dedicated to body, language, and technologies, the choice of feminism as a thematic and analytical axis for testing the textual and imagistic meaning effects generated by these LLMs is justified both epistemologically and through personal and political implication. From this enunciative position, the study seeks to show how the erasure of certain forms of existence and resistance in artificial intelligence generated outputs exposes the limits, and dangers, of computational discursive memory. The research adopts a qualitative, critical, and interpretivist approach, articulating discourse analysis with experimentation in natural language and image generation. The study draws on LLMs available on public platforms, using prompts designed to challenge their ability to generate ideologically marked meanings. The corpus comprises texts and images generated from these prompts, whose analyzed data reveal consistent patterns of reproducing hegemonic discourses, erasing antagonisms, and producing symbolic sanitization. This suggests that generative artificial intelligence relies on utterances aligned with dominant discourses present in training data, thereby avoiding dissent. In the domain of image generation, the models consistently produced white, young, slim, Westernized bodies. The confrontation between these outputs and the theory of discursive memory reveals that the architecture of LLMs favors the repetition of historically dominant meanings: by operating on statistical patterns, the models produce ideological stabilization effects, reinforcing naturalized discourses without intention or interpretation. The notion of computational discursive memory therefore does not operate as a metaphor, but as an analytical concept that makes it possible to critically examine how statistical regularities resulting from the model architectures generate effects that can be reinscribed within the discursive field. When such patterns are put into circulation in dialogue with human subjects, they acquire symbolic density, for they are appropriated, challenged, displaced, and reinscribed in discursive practices that produce meaning effects, including those of subjectivation, erasure, or resistance. The theoretical contribution of the thesis lies in updating the concept of discursive memory for a new empirical domain, namely, algorithmic language technologies. Its political contribution lies in exposing the erasures and neutralizations these technologies enact, often under the guise of technical neutrality. Its methodological contribution lies in integrating discourse analysis, technology studies, and critical experimentation with artificial intelligence systems.