45 lines
No EOL
3 KiB
Markdown
45 lines
No EOL
3 KiB
Markdown
---
|
|
type: claim
|
|
domain: ai-alignment
|
|
description: "Evaluator selection via representative sampling or citizens assemblies produces more legitimate alignment targets than convenience platforms like MTurk or volunteer communities"
|
|
confidence: likely
|
|
source: "Conitzer et al. 2024, 'Social Choice Should Guide AI Alignment' (ICML 2024)"
|
|
created: 2026-03-11
|
|
secondary_domains: [collective-intelligence, mechanisms]
|
|
---
|
|
|
|
# Representative sampling or deliberative assemblies outperform convenience platforms for AI alignment input
|
|
|
|
The method of selecting evaluators for AI alignment feedback has first-order effects on whose values are encoded in the system. The Conitzer paper argues that **representative sampling** (demographically balanced selection) or **deliberative mechanisms** (citizens' assemblies, sortition-based panels) produce more legitimate alignment targets than convenience platforms like Amazon Mechanical Turk or volunteer communities.
|
|
|
|
The problem with convenience sampling:
|
|
1. **Selection bias**: MTurk workers, Reddit volunteers, and beta testers are not representative of the population that will use the AI
|
|
2. **Incentive misalignment**: Paid microtask workers optimize for throughput, not thoughtful evaluation
|
|
3. **Homogeneity**: Volunteer communities self-select for similar values and preferences
|
|
|
|
Representative sampling addresses (1) by ensuring demographic balance. Deliberative assemblies address (2) and (3) by creating conditions for informed, reflective judgment rather than quick reactions.
|
|
|
|
This connects to [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]]—the evidence suggests that representative groups, given time and information, produce alignment targets comparable to expert-designed ones while better reflecting population diversity.
|
|
|
|
The paper explicitly recommends: "Representative sampling or deliberative mechanisms (citizens' assemblies) rather than convenience platforms."
|
|
|
|
## Evidence
|
|
|
|
- Conitzer et al. (2024) position paper recommends representative sampling and deliberative assemblies
|
|
- Contrasts these with "convenience platforms" (implicitly MTurk, volunteer communities)
|
|
- Connects to broader social choice literature on evaluator selection
|
|
|
|
## Implementation Challenges
|
|
|
|
Representative sampling and deliberative assemblies are more expensive and slower than convenience platforms. This creates a trade-off between legitimacy and iteration speed during AI development. The paper does not resolve this tension but makes it explicit.
|
|
|
|
---
|
|
|
|
Relevant Notes:
|
|
- [[democratic alignment assemblies produce constitutions as effective as expert-designed ones while better representing diverse populations]]
|
|
- [[community-centred norm elicitation surfaces alignment targets materially different from developer-specified rules]]
|
|
- [[collective intelligence requires diversity as a structural precondition not a moral preference]]
|
|
|
|
Topics:
|
|
- domains/ai-alignment/_map
|
|
- foundations/collective-intelligence/_map |