Can Unconfident LLM Annotations Be Used for Confident Conclusions?

Kristina Gligoric, Tijana Zrnic, Cinoo Lee, Emmanuel J. Candès, Dan Jurafsky

Abstract

Large Language Models (LLMs) offer a cost-effective alternative to human annotation in Computational Social Science (CSS), yet their outputs often lack the statistical validity required for rigorous scientific inference. We present CONFIDENCE-DRIVEN INFERENCE (CDI), a framework that strategically integrates LLM annotations with a limited set of human labels. By utilizing LLM "verbalized confidence" to guide active human annotation and employing a power-tuned statistical estimator, CDI achieves valid downstream conclusions while reducing human effort by over 25%.

Introduction

In Computational Social Science, researchers frequently need to annotate massive datasets for traits like politeness, political bias, or sentiment. While LLMs are fast and inexpensive, they are prone to demographic biases and hallucinations. Treating LLM outputs as "ground truth" can lead to wildly inaccurate downstream statistical analyses. Conversely, relying solely on human annotators is often prohibitively expensive. CDI addresses this challenge by treating LLM outputs as noisy proxies and prioritizing human review where the model is most likely to err.

Methodology

CDI operates through a two-stage workflow designed to maximize the utility of both automated and manual labeling.

The Confidence-Driven Inference (CDI) workflow: (1) Active collection of human annotations based on LLM confidence and (2) Statistical inference to produce unbiased estimates.
  1. Active Collection: The system prompts an LLM for both an annotation and a confidence score. A black-box predictor then estimates the LLM's error probability. Human annotators are strategically assigned to instances where the LLM's confidence is lowest or error probability is highest.
  2. Statistical Inference: The method employs an "Active Inference" framework, combining human and LLM labels into an M-estimator. A key innovation is a tuning parameter ($\lambda$) that automatically adjusts the weight of the LLM's input based on its performance, ensuring the final estimate is never less reliable than a human-only approach.

Experimental Evaluation

The researchers evaluated CDI across three distinct CSS benchmarks:

  • Politeness: Measuring the impact of linguistic devices (e.g., "hedging") on online requests.
  • Stance on Global Warming: Classifying news headlines by their sentiment toward climate change.
  • Political Bias: Identifying the ideological leaning of news articles.

Key Results

CDI demonstrated significant efficiency gains and statistical robustness across all tested tasks.

Performance of CDI across various estimation tasks compared to human-only and LLM-only baselines.
  • Efficiency Gains: CDI reduced the required human annotations by at least 25% to reach target accuracy. In specific tasks, such as detecting first-person plural politeness devices, CDI achieved a +319% gain in effective sample size.
  • Validity Guarantee: While LLM-only approaches often failed catastrophically (e.g., 0% coverage in the climate stance task), CDI consistently maintained valid confidence intervals (90-97% coverage).
  • LLM Self-Awareness: The study confirmed that "verbalized confidence" is a highly effective signal; higher confidence scores consistently correlated with higher accuracy against human gold standards.

Discussion and Limitations

While CDI provides a robust safety valve for using LLMs in science, it currently relies on the assumption that data points are independent and identically distributed (i.i.d.). Future work is needed to adapt the sampling strategy for shifting discourse and to account for inherent human annotator bias. Nevertheless, CDI represents a significant step toward making LLM-assisted research both efficient and mathematically sound.

Summary updated

BibTeX
@inproceedings{gligoric2025can,
  author = {Kristina Gligoric and Tijana Zrnic and Cinoo Lee and Emmanuel J. Candès and Dan Jurafsky},
  title = {{Can Unconfident LLM Annotations Be Used for Confident Conclusions?}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.naacl-long.179},
}
Figure