Research map / Large language models / AI and society
Position: Insights from Survey Methodology can Improve Training Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that survey methodology insights (response process, context effects, selection bias, transparency) can improve the quality and efficiency of human-labeled training data for AI.
Problem
Human-labeled training and feedback data for AI are often collected without rigorous attention to data collection methods, so label accuracy and labeler diversity suffer, degrading model performance and alignment.
Approach
Frames label collection as a survey, maps survey-methodology theories (satisficing, context/order effects, acquiescence, anchoring via pre-filled responses, selection bias from labeler characteristics) to labeling tasks, and proposes hypotheses, mitigations, and transparency norms drawn from survey research.
Key results
- Argues that the yes/no format for labeling collects better data than select-all-that-apply, citing an experiment where splitting labels across screens raised hate speech annotation rates and improved trained models.
- Reports that in one recent experiment fewer than three percent of labelers chose a provided 'don't know' option, without changing the label distribution.
- Notes that Transparency Initiative member polling firms outperform non-members, suggesting disclosure is a proxy for estimate quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Split hate-speech and offensive-language labeling across two screens instead of collecting labels together | Tweet labeling · Hate speech and offensive language annotation | annotation rate | Helped |
| Train models on labels collected across two screens instead of labels collected together | Models trained on tweet labels · Hate speech and offensive language labeling | model performance | Helped |
| Offer labelers a “don't know” option | Labeling tasks · Label distribution | label distribution: fewer than three percent | No clear effect |
| Use a series of yes/no questions rather than a select-all-that-apply question | Survey respondents · Survey questions with multiple possible labels | response accuracy | Helped |
| Train models on data labeled by models rather than humans | Models trained on labeled data · Model training | performance | Hurt |
Surprising findings
- Providing a 'don't know' option to labelers rarely changed the label distribution (fewer than 3% chose it), challenging the assumption that such options harm or are needed.
- Models trained on data from different labeler demographics (e.g., female vs. male, Asian vs. white) make different predictions, evidence of labeler-driven selection bias.
Limitations
- No new experiments; claims rest on literature review and hypotheses, many of which are untested.
- Acknowledges the notion of 'population' for labelers is not resolved.
Experimental setup
Builds on: satisficing theory cognitive interviewing active learning RLHF
Closest papers by meaning
- Position: Challenges and Future Directions of Data-Centric AI Alignment (ICML 2025 · similarity 0.83)
- AI Evaluation Should Require Standardized Item-Level Data Releases (NeurIPS 2026 · similarity 0.81)
- No Free Delivery Service: Epistemic limits of passive data collection in complex social systems (NeurIPS 2024 · similarity 0.80)
- Can We Trust Item Response Theory for AI Evaluation? (NeurIPS 2026 · similarity 0.80)
- Bridging the Data Provenance Gap Across Text, Speech, and Video (ICLR 2025 · similarity 0.80)
- Position: Human Baselines in Model Evaluations Need Rigor and Transparency (With Recommendations & Reporting Checklist) (ICML 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.