Research map / Large language models / AI and society
Evaluating LLM-contaminated Crowdsourcing Data Without Ground Truth
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.
Proposes a training-free conditioned correlated-agreement peer prediction score that uses the requester's LLM-generated labels to detect LLM-reliant and low-effort crowdworkers on multiple-choice annotation data.
Problem
Crowdsourced annotation datasets are contaminated when workers use LLMs to answer multiple-choice labeling tasks, and existing LLM-text detectors need high-dimensional text, so they do not apply to structured labels; without ground truth, low-effort workers are hard to identify.
Approach
Extends the correlated agreement peer prediction mechanism by conditioning agreement scores on a signal Z (samples of LLM labels available to the requester), with theoretical information-monotonicity guarantees under a model of LLM collusion and a weaker condition for lazy-reporting strategies.
Key results
- On the preference alignment dataset with GPT-4-reliant agents, the conditioned CA mechanism reaches average AUC 0.85 (bottom 10% 0.77), versus 0.70 for standard CA and 0.68 for EM.
- On the hatefulness, offensiveness and toxicity datasets, the conditioned CA mechanism achieves average AUC 0.85, 0.94 and 0.91 respectively, the best among tested methods.
- Empirically, assumptions 1 and 2 generally do not hold, while assumptions 3 and 4 hold when conditioning on the same model's samples.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use conditioned CA instead of the strongest baseline, CA | Low-effort-agent detection with GPT-4-reliant agents · Hatefulness | AUC: 0.84 -> 0.85 | Helped |
| Use conditioned CA instead of the strongest baseline, CA | Low-effort-agent detection with GPT-4-reliant agents · Offensiveness | AUC: 0.92 -> 0.94 | Helped |
| Use conditioned CA instead of the strongest baseline, CA | Low-effort-agent detection with GPT-4-reliant agents · Toxicity | AUC: 0.89 -> 0.91 | Helped |
| Use conditioned CA instead of the strongest baseline, CA | Low-effort-agent detection with GPT-4-reliant agents · Preference Alignment | AUC: 0.70 -> 0.85 | Helped |
Surprising findings
- Even independently sampled responses from the same LLM remain correlated after conditioning on another independent sample, so the strong independence assumptions fail in practice.
- Classic EM and CA perform well with few LLM-reliant agents but their AUC degrades substantially as LLM-reliant fraction grows, and OA_Z has high variance.
Limitations
- Method is designed for subjective labeling tasks where humans are the gold standard; it may reward incorrect correlations if LLMs surpass humans.
- Does not beat all baselines in every setting; EM performs better when low-effort agents are diverse or random/biased.
- Only five language models tested, and some models refuse toxicity prompts.
Experimental setup
Datasets: Hatefulness/Offensiveness/Toxicity labeling dataset (3459 comments, MTurk) Preference alignment dataset (10461 questions, crowd and expert labels)
Benchmarks: Hatefulness/Offensiveness/Toxicity labeling dataset Preference alignment dataset
Metrics: AUC average AUC bottom 10% quantile AUC conditional total variation distance mutual information
Baselines: Output Agreement Correlated Agreement Expectation-Maximization (Dawid-Skene) Conditioned Agreement (OA_Z)
Builds on: correlated agreement mechanism Dawid-Skene model peer prediction
Closest papers by meaning
- Exploring and Mitigating Adversarial Manipulation of Voting-Based Leaderboards (ICML 2025 · similarity 0.85)
- Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors (ICLR 2025 · similarity 0.85)
- A Synthetic Dataset for Personal Attribute Inference (NeurIPS 2024 · similarity 0.85)
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.85)
- From Crowdsourced Data to High-quality Benchmarks: Arena-Hard and Benchbuilder Pipeline (ICML 2025 · similarity 0.85)
- Beyond Memorization: Violating Privacy via Inference with Large Language Models (ICLR 2024 · similarity 0.84)
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.