The new shape of erasure : When name absence is no longer the criterion to detect biasand hallucination in LLM’s generated narratives of history ofscience

Sammanfattning

Large language models (LLMs) do not merely retrieve information – they narrate it, curate it. When asked about the history of scientific discovery, they produce accounts that are grammatically coherent, epistemically authoritative, and, as this thesis demonstrates, structurally biased in ways that reproduce historical injustices without announcing them as such. This study investigates that process through a close empirical analysis of how four closed major LLMs construct narratives about the discovery of nuclear fission and the important contributions of Lise Meitner and Otto Hahn – a case in which the gendered misattribution of scientific credit is among the most extensively documented in 20th-century history of science. The thesis develops an auditing framework including original two-branch analytical framework combining Epistemic Heuristic Audit, which examines the linguistic architecture of AI-generated narratives through sixteen operationalised rules, with Ground Truth Retrieval, which evaluates the evidentiary grounding of individual claims against a stratified corpus of authoritative scholarly and popular sources using BM25 lexical retrieval. Across 556 atomic claims extracted from 40 generated-responses, the analysis finds that 68.9% combine high epistemic certainty with insufficient evidentiary support. What the analysis reveals is not that Meitner has been erased, she is mentioned. However, the form of her presence in these generated-responses is consistently different from the form that establishes someone as the author of a discovery. That difference accumulates in the places where the question of who owns a piece of knowledge is most directly at stake. When the two analytical branches are read together, they tell the same story from different approach directions. The fact that they converge as often as they do is difficult to account for without concluding that something systematic is happening. Theorised through the Matthew Effect, the Matilda Effect, and Fricker's account of epistemic injustice, the findings suggest that LLMs’ systems do not generate bias so much as inherit and amplify it. They are encoding historical documentations into narratives delivered with the unqualified confidence of established fact. The contribution of this thesis lies not only in its empirical findings but in the analytical framework it proposes: a replicable method for making visible the epistemic structures that AI systems reproduce when they tell us who did what, and how it directly influences the way people understand, acknowledge the marginalized scientific actors.