Woman Is To Nurse As Man Is To... Nurse Too? : Uncovering Structural Bias in GPT-2

Sammanfattning

Large Language Models like GPT-2 are not merely passive mirrors reflecting societal biases; they actively reconstruct and amplify such biases within their internal semantic spaces. This thesis critically interrogates how occupational-gender biases become structurally internalized in GPT-2’s embedding space, uncovering not superficial stereotypes but deeply embedded social hierarchies encoded through computational means. By analyzing the semantic positioning of various occupations via cosine similarity, gender axis projection, and unsupervised semantic clustering, this research uncovers a structural bias more profound than simple stereotypes. The study’s key finding is that GPT-2 operates with a powerful, implicitly male-coded prototype for the conceptual categories of “occupation” and “professional.” This creates a gravitational pull where even female-dominated roles like “nurse” or “librarian” are semantically distorted towards a masculine center. Furthermore, this thesis demonstrates that the model’s internal organization of occupations is not primarily driven by gender, but by latent hierarchies of power, prestige, and institutional authority. High-status, credentialed professions form a semantically central cluster, mirroring societal power structures. This work in Digital Humanities challenges prevailing assumptions about neutrality in AI, highlighting the necessity of addressing deeply rooted computational structures, not merely surface outputs, to foster genuinely equitable artificial intelligence systems.