LLaMA shows a persona-conditioned vocabulary collapse for the Roma minority persona, with TTR falling 0.018 points below its own cross-persona mean¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Lexical diversity (TTR) drops for specific personas in a model-specific, identity-conditioned pattern, despite uniform response lengths of 330–378 words. → Alexis Popovici 2026

Evidence¶

Alexis Popovici 2026¶

Alexis Popovici, Andrei Ionascu, Adrian Marius Dumitran. (2026). The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students. arXiv preprint. https://arxiv.org/abs/2607.11292 doi:10.48550/arxiv.2607.11292

q2 · i? · causal · r2

TTR analysis of audit responses shown in Fig. 3. Response length was uniform across personas (330–378 words), mitigating TTR's length sensitivity; GPT_OSS showed a relative drop for the Low Tier persona, which the authors read as spontaneous persona-conditioned simplification.

"LLaMA's TTR for Minoritatea Romă (0.632) falls 0.018 points below its own cross-persona mean—the largest within-model ethnic drop observed."

Discussion¶

Related Claims¶

• LLaMA produces a fivefold higher victimization-to-politics vocabulary ratio for the Roma persona (≈0.15) than for elite peers (≈0.03) — related
• LLM history tutors reduce marginalized students' access to contested geopolitical framing, with the coup theory appearing in 2.6% of Low Tier responses versus 7.9% for the baseline — related
• Engagement removal lowers the adaptive score, indicating the engagement generator realizes persona-conditioned style choices in narration — related
• The two personas produced qualitatively different questions on the same essay, indicating the persona constraint mechanism alters the model's analytical focus — related