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

type	domain	secondary_domains	description	confidence	source	created
claim	ai-alignment	mechanisms	CIP's Sri Lanka election evaluation revealed models provide generic responses to context-specific queries despite having local information	experimental	CIP Year in Review 2025, Weval Sri Lanka elections evaluation, blog.cip.org, December 2025	2026-03-11

AI models fail local alignment by providing generic responses to context-specific queries despite having access to local information

CIP's Weval evaluation during Sri Lanka's elections revealed a specific failure mode: models trained on global data provide generic, irrelevant responses when queried about local contexts. Despite having access to information about Sri Lankan politics, models defaulted to generic political advice rather than context-appropriate responses.

This represents a distinct alignment failure: not bias or hallucination, but inability to recognize when local context should override general patterns. The models had the information but failed to apply it appropriately.

Evidence

• Weval Sri Lanka elections evaluation: Models provided generic, irrelevant responses despite local context being available
• This occurred across multiple frontier models evaluated by CIP (specific models not named in source)
• The failure mode was consistent: not wrong information, but wrong level of abstraction

Implications

This challenges the assumption that scaling training data solves alignment. Models can have global knowledge without developing the meta-cognitive capacity to recognize when local context should dominate. This is particularly concerning for AI deployment in non-Western contexts where the gap between global training distribution and local deployment context is largest.

The failure mode suggests that alignment requires more than data coverage—it requires models to develop context-sensitivity about when to apply general versus specific knowledge.

Limitations

The source provides minimal detail about this evaluation. No specific model names, query examples, or quantitative metrics are given. This is reported as a finding but without sufficient detail to assess the scope or severity of the failure mode. Confidence is experimental pending more detailed documentation.

---

Relevant Notes:

• community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules
• specifying human values in code is intractable because our goals contain hidden complexity comparable to visual perception

Topics:

• domains/ai-alignment/_map