The effect of responsible AI principles on the effectiveness of a culturally-aware large language model in preserving intangible cultural heritage: Evidence from Turkana Traditional Medicine Knowledge

Responsible AI (RAI) principles are widely acknowledged as necessary for trustworthy Large Language Model (LLM) deployment, yet they are typically stated as values rather than embedded as testable, architectural mechanisms, a gap with direct consequences for LLM systems entrusted with preserving Intangible Cultural Heritage (ICH). This paper reports an empirical evaluation of the effect of RAI principles, operationalised through Fairness, Accountability, and Transparency, on the effectiveness of a culturally-aware LLM applied to Turkana Traditional Medicine Knowledge (TMK) in Kenya. A sequential mixed-methods design drew on 107 respondents and 400 simulation outputs generated across four LLM configurations, of which two are the focus of this paper. RAI correlated moderately strongly with LLM Effectiveness (r = .444, p < .001) and significantly predicted it in regression (R² = .226, p < .001), an effect driven specifically by Transparency (β = .460, p = .019) rather than Fairness or Accountability. Under controlled simulation, however, RAI implemented independently of community knowledge grounding produced no significant effectiveness gain on any of five indicators, and a significant decline relative to baseline on BLEU-4 and ROUGE-L. The findings indicate that RAI principles are positively associated with perceived LLM effectiveness, but are architecturally insufficient on their own to improve measured effectiveness, with direct implications for how Responsible AI is operationalised in LLM systems for indigenous knowledge preservation. This paper isolates RAI's standalone contribution specifically, distinct from any interaction with community knowledge grounding, which is examined separately in a companion paper.