Large-scale ASR systems such as Whisper achieve competitive aggregate Word Error Rate (WER) on multilingual benchmarks, but this aggregate conceals systematic disparities across speaker populations. We evaluate Whisper large-v3 on 276 recordings from the Corpus Oral y Sonoro del Español Rural (COSER), a dialectological archive of elderly rural speakers across all Spanish provinces. WER is computed separately for Informants and Interviewers within each recording, revealing that mixed-role evaluation underestimates Informant WER in the majority of provinces, with the largest corrections in southern areas.Negative Binomial regression with cluster-robust standar errors shows that Andalusia and Extremadura generate significantly more Informant errors than the Castilian heartland (Andalusia IRR = 1.20, p < 0.001; Extremadura IRR = 1.24, p = 0.020), while no geographic predictor reaches significance for Interviewers sharing the same recording environment. Male Informants generate 12.5% more errors than females after geographic adjustment (p < 0.001), consistent with differential vernacular retention in traditional rural communities. The geographic pattern aligns with established dialectological classifications of Peninsular Spanish. These results demonstrate that role-disaggregated evaluation is a necessary methodological prerequisite for fairness audits of ASR systems applied to sociolinguistically diverse corpora: aggregate benchmarks systematically suppress disparities that are borne disproportionately by the most underrepresented speaker populations, and their use in isolation constitutes both an allocative harm and a measurement failure

We evaluate whether open-source LLMs can produce proficiency-graded English adaptations of entries from the Diccionario de colombianismos (DiCol), a Colombian Spanish lexicographic resource used in language teaching. Three 7–8B instruction-tuned models—Llama 3.1, Qwen2.5, and Mistral—generate Beginner, Intermediate, and Advanced translations for all 8,252 definitions using structured zero-shot prompts identical across levels except for the target CEFR band. Automated metrics show that Intermediate targeting collapses (73–83% classified as Advanced by vocabulary, 𝜒² > 705, p < .001) and that Advanced outputs expand 4.9–8.2× relative to the source. Expert annotation of a 360-entry stratified sample (𝜅 = 0.61–0.68) identifies hallucination in 19% of entries (Fleiss’ 𝜅 = 0.77 for cultural preservation categories, 97% unanimity among flagged cases). Hallucination concentrates in the Advanced condition (81%, 𝜒² = 86.6, p < .001) and is associated with higher expansion (U = 16,662, p < .001, r = 0.68), manifesting primarily as generic elaboration and, in a smaller proportion, as Colombia-stereotyping and pragmatic polarity inversion. We discuss these findings through the lens of (CITATION)’s domestication framework and describe the observed pattern as algorithmic domestication.