theseus: extract from 2025-12-00-cip-year-in-review-democratic-alignment.md

- Source: inbox/archive/2025-12-00-cip-year-in-review-democratic-alignment.md
- Domain: ai-alignment
- Extracted by: headless extraction cron (worker 2)

Pentagon-Agent: Theseus <HEADLESS>

domains/ai-alignment/58-percent-believe-ai-could-decide-better-than-elected-representatives-creating-ambiguity-about-democratic-alignment-goals.md

@@ -0,0 +1,39 @@
+---
+type: claim
+domain: ai-alignment
+description: "Majority willingness to defer to AI over human representatives creates ambiguity about whether democratic alignment targets human authority or AI optimization"
+confidence: experimental
+source: "CIP Year in Review 2025, Global Dialogues findings"
+created: 2026-03-11
+secondary_domains: [collective-intelligence]
+---
+
+# 58% of Global Dialogues participants believe AI could make superior decisions versus local elected representatives, creating ambiguity about whether democratic alignment targets human authority or AI optimization
+
+CIP's Global Dialogues found that 58% of participants believed AI could make superior decisions compared to local elected representatives. This finding is deeply ambiguous: it could indicate trust in AI-augmented democratic processes, or willingness to cede decision authority to AI systems.
+
+If the latter interpretation is correct, it undermines the human-in-the-loop thesis at scale. Democratic alignment assumes humans want to retain decision authority while using AI as a tool. But if a majority believes AI should make decisions instead of humans, the alignment target shifts from "AI that helps humans decide" to "AI that decides on behalf of humans."
+
+The 28% who agreed "AI should override established rules if calculating better outcomes" reinforces this ambiguity. This is not a fringe position — it's more than one in four participants endorsing consequentialist AI authority over rule-of-law constraints. The 47% who felt chatbot interactions increased their belief certainty suggests AI influence on human judgment formation itself.
+
+The critical question is whether these responses reflect:
+1. Frustration with current representatives (AI as protest vote)
+2. Genuine belief in AI superiority (AI as technocratic authority)
+3. Misunderstanding of what "AI decision-making" means in practice
+
+Without disambiguation, democratic alignment infrastructure may be building toward a goal (human authority) that the majority does not actually want.
+
+## Evidence
+- 58% believed AI could make superior decisions vs. local elected representatives (CIP Global Dialogues, 10,000+ participants, 70+ countries)
+- 28% agreed AI should override established rules if calculating better outcomes
+- 47% felt chatbot interactions increased their belief certainty
+
+## Limitations
+The survey question framing is not provided in the source. "Could make superior decisions" is ambiguous — superior in what sense? Faster? More informed? More aligned with participant values? The interpretation depends heavily on how the question was asked. Without access to the survey instrument, we cannot determine whether responses reflect genuine preference for AI authority or misunderstanding of the question. This is a single survey from a single organization, so confidence is experimental.
+
+---
+
+Relevant Notes:
+- [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]] — but may be building toward AI authority rather than human authority
+- [[AI alignment is a coordination problem not a technical problem]] — coordination around what goal?
+- [[safe AI development requires building alignment mechanisms before scaling capability]] — but what if the alignment target is AI decision authority?

domains/ai-alignment/ai-models-fail-local-alignment-when-trained-globally-sri-lanka-election-responses-were-generic-despite-local-context.md

@@ -0,0 +1,31 @@
+---
+type: claim
+domain: ai-alignment
+description: "Global model training creates systematic failure mode where models cannot provide locally-relevant responses to context-specific queries"
+confidence: experimental
+source: "CIP Year in Review 2025, Weval Sri Lanka elections evaluation"
+created: 2026-03-11
+secondary_domains: [ai-safety]
+---
+
+# Global model training creates systematic failure mode where AI models provide generic responses to local context-specific queries, as evidenced by Sri Lanka election evaluation
+
+CIP's Weval evaluation of AI models during Sri Lanka elections found that models provided generic, irrelevant responses despite being given local context. This reveals a specific failure mode: global training creates models that cannot align to local contexts even when explicitly prompted.
+
+This is distinct from general capability failures. The models were not unable to respond — they responded with generic political advice that would apply anywhere, failing to engage with the specific electoral dynamics, candidates, or issues in Sri Lanka. The failure is one of alignment granularity: the model's training optimized for global applicability at the cost of local relevance.
+
+The implication is that democratic alignment at scale may require region-specific training or fine-tuning, not just global deliberation. A model aligned to aggregate global preferences may systematically fail populations whose contexts differ from training distribution centroids.
+
+## Evidence
+- Weval Sri Lanka elections evaluation: Models provided generic responses despite local electoral context
+- This occurred despite CIP's global deliberation framework being active
+- The failure mode is systematic (generic responses) not random (hallucination or refusal)
+
+## Limitations
+Single-country evaluation limits generalizability. We don't know if this is specific to Sri Lanka, to elections, or to the models tested. The source doesn't specify which models were evaluated, what prompts were used, or whether this failure mode appears in other regional evaluations (e.g., Samiksha in India). Confidence is experimental because this is a single case study.
+
+---
+
+Relevant Notes:
+- [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]] — but may not solve local context failures
+- [[pluralistic alignment must accommodate irreducibly diverse values simultaneously rather than converging on a single aligned state]] — local context as irreducible diversity

domains/ai-alignment/community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules.md

@@ -19,6 +19,12 @@ Since [[democratic alignment assemblies produce constitutions as effective as ex
 
 Since [[collective intelligence requires diversity as a structural precondition not a moral preference]], community-centred norm elicitation is a concrete mechanism for ensuring the structural diversity that collective alignment requires. Without it, alignment defaults to the values of whichever demographic builds the systems.
 
+
+### Additional Evidence (confirm)
+*Source: [[2025-12-00-cip-year-in-review-democratic-alignment]] | Added: 2026-03-12 | Extractor: anthropic/claude-sonnet-4.5*
+
+CIP's Weval political neutrality evaluation generated 400 prompts and 107 evaluation criteria from 1,000 participants, achieving 70%+ consensus across political groups. Samiksha conducted 25,000+ queries across 11 Indian languages with 100,000+ manual evaluations in healthcare, agriculture, education, and legal domains. Both programs demonstrate that community-centered evaluation at scale surfaces context-specific alignment targets that global model training misses — as evidenced by models providing generic responses to Sri Lanka election queries despite local context.
+
 ---
 
 Relevant Notes:

domains/ai-alignment/democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations.md

@@ -19,6 +19,12 @@ However, this remains one-shot constitution-setting, not continuous alignment. T
 
 Since [[collective intelligence requires diversity as a structural precondition not a moral preference]], democratic assemblies structurally ensure the diversity that expert panels cannot guarantee. Since [[the alignment problem dissolves when human values are continuously woven into the system rather than specified in advance]], the next step beyond assemblies is continuous participatory alignment, not periodic constitution-setting.
 
+
+### Additional Evidence (extend)
+*Source: [[2025-12-00-cip-year-in-review-democratic-alignment]] | Added: 2026-03-12 | Extractor: anthropic/claude-sonnet-4.5*
+
+CIP's Global Dialogues scaled democratic alignment to 10,000+ participants across 70+ countries in 2025, achieving 70%+ cross-partisan consensus on AI evaluation criteria. This represents a 100x scale-up from previous experiments while maintaining consensus properties. Frontier labs (Meta, Cohere, Anthropic) and governments (India, Taiwan, Sri Lanka) adopted the frameworks, indicating the approach has crossed the credibility threshold for institutional use. However, the gap between evaluation adoption and deployment impact remains unclear — labs using these tools does not necessarily mean findings changed what was deployed.
+
 ---
 
 Relevant Notes:

domains/ai-alignment/democratic-ai-alignment-scaled-to-10000-participants-across-70-countries-achieving-cross-partisan-consensus.md

@@ -0,0 +1,33 @@
+---
+type: claim
+domain: ai-alignment
+description: "Democratic alignment infrastructure can scale to 10,000+ participants across 70+ countries while maintaining 70%+ cross-partisan consensus on evaluation criteria"
+confidence: likely
+source: "CIP Year in Review 2025, Global Dialogues program"
+created: 2026-03-11
+secondary_domains: [collective-intelligence, mechanisms]
+---
+
+# Democratic AI alignment scaled to 10,000+ participants across 70+ countries achieving 70%+ cross-partisan consensus on evaluation criteria
+
+CIP's Global Dialogues program in 2025 achieved 10,000+ participants across 70+ countries in 6 deliberative dialogues, demonstrating that democratic alignment infrastructure can operate at scale while maintaining meaningful consensus. The Weval political neutrality evaluation generated 400 prompts and 107 evaluation criteria from 1,000 participants, achieving 70%+ consensus across political groups.
+
+This represents a 100x scale-up from previous democratic alignment experiments while maintaining the consensus properties that make the approach viable. The cross-partisan consensus threshold (70%+) is particularly significant because it demonstrates that diverse populations can agree on AI evaluation criteria despite political polarization.
+
+The scale achievement matters because it moves democratic alignment from experimental proof-of-concept to operational infrastructure. Frontier lab adoption (Meta, Cohere, Anthropic) and government incorporation (India, Taiwan, Sri Lanka) indicate the approach has crossed the credibility threshold for institutional use.
+
+## Evidence
+- CIP Global Dialogues: 10,000+ participants, 70+ countries, 6 deliberative dialogues (2025)
+- Weval political neutrality: 1,000 participants, 400 prompts, 107 criteria, 70%+ cross-partisan consensus
+- Frontier lab partners: Meta, Cohere, Anthropic, UK/US AI Safety Institutes
+- Government adoption: India, Taiwan, Sri Lanka incorporated findings into policy
+
+## Limitations
+The gap between evaluation adoption and deployment impact remains unclear. Labs using these tools as evaluation frameworks does not necessarily mean the findings changed what was deployed. The source provides no evidence that Weval/Samiksha results altered product decisions or deployment behavior.
+
+---
+
+Relevant Notes:
+- [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]] — extended to 10,000+ scale
+- [[community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules]] — confirmed at global scale
+- [[AI alignment is a coordination problem not a technical problem]] — democratic infrastructure as coordination mechanism

domains/ai-alignment/no research group is building alignment through collective intelligence infrastructure despite the field converging on problems that require it.md

@@ -17,6 +17,12 @@ This gap is remarkable because the field's own findings point toward collective
 
 The alignment field has converged on a problem they cannot solve with their current paradigm (single-model alignment), and the alternative paradigm (collective alignment through distributed architecture) has barely been explored. This is the opening for the TeleoHumanity thesis -- not as philosophical speculation but as practical infrastructure that addresses problems the alignment community has identified but cannot solve within their current framework.
 
+
+### Additional Evidence (challenge)
+*Source: [[2025-12-00-cip-year-in-review-democratic-alignment]] | Added: 2026-03-12 | Extractor: anthropic/claude-sonnet-4.5*
+
+CIP is building alignment through collective intelligence infrastructure at scale. Global Dialogues reached 10,000+ participants across 70+ countries. Weval and Samiksha provide evaluation frameworks adopted by Meta, Cohere, Anthropic, and governments in India, Taiwan, and Sri Lanka. CIP's 2026 plans explicitly aim to make Global Dialogues 'standing global infrastructure' for democratic AI alignment. While gaps remain (evaluation vs. deployment impact), CIP is operationalizing collective intelligence for alignment, not just theorizing it.
+
 ---
 
 Relevant Notes:

domains/ai-alignment/samiksha-is-most-comprehensive-ai-evaluation-in-indian-contexts-with-25000-queries-across-11-languages.md

@@ -0,0 +1,32 @@
+---
+type: claim
+domain: ai-alignment
+description: "Samiksha represents unprecedented scale for multilingual, multi-domain AI evaluation in non-English contexts"
+confidence: likely
+source: "CIP Year in Review 2025, Samiksha program"
+created: 2026-03-11
+secondary_domains: [collective-intelligence]
+---
+
+# Samiksha conducted 25,000+ queries across 11 Indian languages with 100,000+ manual evaluations, representing the most comprehensive multilingual AI evaluation in non-English contexts
+
+CIP's Samiksha program conducted 25,000+ queries across 11 Indian languages with 100,000+ manual evaluations, covering healthcare, agriculture, education, and legal domains. CIP describes this as "the most comprehensive evaluation of AI in Indian contexts," and the scale supports that claim — no comparable multilingual, multi-domain evaluation exists in public literature.
+
+The significance is methodological and political. Methodologically, it demonstrates that rigorous AI evaluation can be conducted in non-English, non-Western contexts at scale. Politically, it provides evidence for Indian policymakers that AI systems trained primarily on English/Western data may not serve Indian populations adequately.
+
+The 100,000+ manual evaluations indicate human-in-the-loop assessment at scale, not automated metrics. This matters because automated evaluation metrics (BLEU, ROUGE, perplexity) are known to correlate poorly with actual utility in domain-specific, multilingual contexts. Medical review was included for healthcare accuracy and safety assessment, indicating domain-expert validation.
+
+## Evidence
+- Samiksha: 25,000+ queries, 11 Indian languages, 100,000+ manual evaluations
+- Domains: healthcare, agriculture, education, legal
+- Medical review included for healthcare accuracy and safety assessment
+- Indian government incorporated findings (specific policy changes not detailed in source)
+
+## Limitations
+The source does not provide specific findings from Samiksha — only scale metrics and domain coverage. We don't know what the evaluation revealed about model performance, what failure modes were identified, or how Indian government policy changed in response. The claim is about the evaluation's comprehensiveness and methodology, not its results. Confidence is 'likely' based on scale and institutional adoption, but the lack of detailed findings limits how much we can infer about impact.
+
+---
+
+Relevant Notes:
+- [[community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules]] — Samiksha as community-centered evaluation at scale
+- [[global model training creates systematic failure mode where AI models provide generic responses to local context-specific queries, as evidenced by Sri Lanka election evaluation]] — Samiksha may reveal similar failures in Indian contexts

inbox/archive/2025-12-00-cip-year-in-review-democratic-alignment.md

@@ -7,9 +7,15 @@ date: 2025-12-01
 domain: ai-alignment
 secondary_domains: [collective-intelligence, mechanisms]
 format: article
-status: unprocessed
+status: processed
 priority: medium
 tags: [cip, democratic-alignment, global-dialogues, weval, samiksha, digital-twin, frontier-lab-adoption]
+processed_by: theseus
+processed_date: 2026-03-11
+claims_extracted: ["democratic-ai-alignment-scaled-to-10000-participants-across-70-countries-achieving-cross-partisan-consensus.md", "ai-models-fail-local-alignment-when-trained-globally-sri-lanka-election-responses-were-generic-despite-local-context.md", "samiksha-is-most-comprehensive-ai-evaluation-in-indian-contexts-with-25000-queries-across-11-languages.md", "58-percent-believe-ai-could-decide-better-than-elected-representatives-creating-ambiguity-about-democratic-alignment-goals.md"]
+enrichments_applied: ["democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations.md", "community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules.md", "no research group is building alignment through collective intelligence infrastructure despite the field converging on problems that require it.md"]
+extraction_model: "anthropic/claude-sonnet-4.5"
+extraction_notes: "Four claims extracted focusing on: (1) democratic alignment scaling achievement, (2) local alignment failure mode, (3) Samiksha evaluation comprehensiveness, (4) ambiguity in public willingness to defer to AI. Three enrichments applied to existing claims about democratic alignment and collective intelligence infrastructure. The 58% finding about AI vs. elected representatives is particularly significant as it creates ambiguity about whether democratic alignment should preserve human authority or enable AI authority. CIP entity updated with 2025 achievements and 2026 plans."
 ---
 
 ## Content
@@ -59,3 +65,13 @@ CIP's comprehensive 2025 results and 2026 plans.
 PRIMARY CONNECTION: [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]]
 WHY ARCHIVED: Scale-up evidence for democratic alignment + frontier lab adoption evidence
 EXTRACTION HINT: The 70%+ cross-partisan consensus and the evaluation-to-deployment gap are both extractable
+
+
+## Key Facts
+- CIP Global Dialogues: 10,000+ participants, 70+ countries, 6 deliberative dialogues (2025)
+- Weval political neutrality: 1,000 participants, 400 prompts, 107 evaluation criteria generated
+- Samiksha: 25,000+ queries, 11 Indian languages, 100,000+ manual evaluations
+- 13.7% reported concerning/reality-distorting AI interactions affecting someone they know
+- 47% felt chatbot interactions increased their belief certainty
+- Frontier lab partners: Meta, Cohere, Anthropic, UK/US AI Safety Institutes
+- Government adoption: India, Taiwan, Sri Lanka