teleo-codex/domains/ai-alignment/ai-models-fail-local-alignment-providing-generic-responses-to-culturally-specific-contexts.md

type	domain	description	confidence	source	created	secondary_domains
claim	ai-alignment	Global AI models provide generic responses to culturally-specific contexts despite having relevant local information in training data	experimental	CIP Year in Review 2025, Sri Lanka elections and Samiksha evaluations	2026-03-11	collective-intelligence

AI models fail local alignment by providing generic responses to culturally-specific contexts despite having relevant training data

CIP's evaluation of AI models during Sri Lanka's elections revealed a specific failure mode: models provided generic, irrelevant responses despite the local context being available. This suggests that global models trained predominantly on Western data fail to activate or prioritize culturally-specific knowledge even when it exists in their training corpus.

This failure mode is distinct from lack of capability—the models had access to information about Sri Lankan politics but defaulted to generic responses rather than contextually appropriate ones. This reveals a structural misalignment between global model training and local deployment contexts. The problem is not that the knowledge is absent, but that the model's optimization process does not reliably surface or weight local context appropriately.

The finding is reinforced by Samiksha's evaluation of 25,000+ queries across 11 Indian languages, which required 100,000+ manual evaluations precisely because automated metrics could not capture cultural appropriateness. Domains tested included healthcare, agriculture, education, and legal contexts—all areas where local norms, practices, and values diverge materially from Western-centric training data. The requirement for human expert review to assess accuracy and safety indicates that standard evaluation metrics miss culturally-embedded alignment failures.

Evidence

• Sri Lanka elections: Models provided generic, irrelevant responses despite local context being available in training data
• Samiksha scale: 25,000+ queries across 11 Indian languages with 100,000+ manual evaluations required
• Domains tested: Healthcare, agriculture, education, legal contexts in Indian languages
• Evaluation requirement: Human expert review necessary to assess accuracy and safety, indicating automated metrics insufficient
• Implication: The failure is not capability but prioritization—models have the information but don't reliably use it

Implications

This failure mode suggests that scaling model size or training data alone will not solve alignment for diverse global populations. The models need mechanisms to recognize and prioritize local context, not just possess the information. This has direct implications for the no research group is building alignment through collective intelligence infrastructure despite the field converging on problems that require it claim—local alignment may require continuous community input rather than one-time training data inclusion.

---

Relevant Notes:

• pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state
• community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules
• persistent irreducible disagreement

Topics:

• domains/ai-alignment/_map
• foundations/collective-intelligence/_map