Evaluating LLM-Contaminated Crowdsourcing Data Without Ground Truth

TL;DR

The paper addresses LLM contamination in crowdsourcing by proposing a training-free scoring mechanism that extends peer prediction through conditioning on a principal-provided cheap signal Z. The conditioned Correlated Agreement (CA) mechanism learns a data-driven scoring function and assigns bonus/penalty rewards to workers based on how their responses correlate with peers when conditioned on Z, yielding a score that approximates the conditional TVD mutual information $I_TVD(X_i; X_j | Z)$ under truthful reporting. The authors derive sufficient and (under some regimes) necessary conditions for information monotonicity, including weaker conditions for lazy-reporting strategies, and validate the approach on real subjective labeling tasks with multiple LLMs. Empirical results show robust detection of low-effort agents across mixed crowds, outperforming several baselines in scenarios with substantial LLM-assisted responses, while acknowledging limitations and opportunities for extension to open-ended tasks and adaptive sampling.

Abstract

The recent success of generative AI highlights the crucial role of high-quality human feedback in building trustworthy AI systems. However, the increasing use of large language models (LLMs) by crowdsourcing workers poses a significant challenge: datasets intended to reflect human input may be compromised by LLM-generated responses. Existing LLM detection approaches often rely on high-dimensional training data such as text, making them unsuitable for annotation tasks like multiple-choice labeling. In this work, we investigate the potential of peer prediction -- a mechanism that evaluates the information within workers' responses without using ground truth -- to mitigate LLM-assisted cheating in crowdsourcing with a focus on annotation tasks. Our approach quantifies the correlations between worker answers while conditioning on (a subset of) LLM-generated labels available to the requester. Building on prior research, we propose a training-free scoring mechanism with theoretical guarantees under a crowdsourcing model that accounts for LLM collusion. We establish conditions under which our method is effective and empirically demonstrate its robustness in detecting low-effort cheating on real-world crowdsourcing datasets.

Figures (7)

• Figure 1: The causal relationship among key variables.
• Figure 2: The TVD mutual information between $X_i$ and $X_j$ (for "Human", "Normal worker", and "expert") or $Z_i$ and $X_j$ (for remaining bars) conditioned on $Z$, i.e. $I_{\text{TVD}}(X_i; X_j \mid Z)$ or $I_{\text{TVD}}(Z_i; X_j \mid Z)$. LLM names on the x-axis denote the models used to sample $Z$ while each bar in the same group represents the model used to sample $Z_i$. We use "Random Agents" as a reference, who randomly selects an answer to each question according to the prior.
• Figure 3: (a) Distribution of the conditioned CA scores with $(\alpha_{llm}, \alpha_r, \alpha_b)=(0.1,0.05,0.05)$. (b) AUC scores of various methods as the fraction of LLM-reliant agents increases. In both panels, GPT-4 generates samples for $Z_i$, $Z_j$, and $Z$ using the preference alignment dataset.
• Figure 4: The area under the ROC curve (AUC) for various methods while varying the fraction of LLM-reliant agents. These are the analogous figures for \ref{['fig:auc_pref']} for the remaining LLMs on the toxicity labeling dataset.
• Figure 5: The area under the ROC curve (AUC) for various methods while varying the fraction of questions the principal can sample.

Theorems & Definitions (15)

• Definition 3.1
• Definition 3.2
• Theorem 4.3
• Proposition 4.4
• Proposition 4.6
• Remark 4.7
• proof : Proof of \ref{['thm:sufficiency']}
• Lemma D.1
• Lemma D.2
• proof : Proof of \ref{['lem:ZandX']}