When LLMs Disagree with Human Experts: Understanding LLM Annotation Failures in Nutrition Misinformation through Hierarchical Error Analysis using Seed Oil Narratives

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Nutrition, Fitness & Lifestyle Medicine · Depth: Expert, quick

Summary

A study evaluated the effectiveness of five open-source large language models (LLMs) as annotators for domain-specific health misinformation on social media, focusing on seed oil narratives. Researchers used a dataset of 169 Instagram posts, where expert nutritionists provided gold-standard labels, identifying 71% as positive for misinformation. A new hierarchical error taxonomy was introduced to categorize LLM misclassifications by direction, mechanism, and contributing factors, offering interpretable insights into model failures. The analysis revealed systematic error patterns, including LLMs misinterpreting nuanced claims and exhibiting overconfidence in predictions. This highlights conditions where LLM annotations diverge from expert judgment, underscoring the need for robust evaluation frameworks for LLM-based annotators in health and nutrition domains.

Key takeaway

For machine learning engineers developing LLM-based annotation systems for health or nutrition content, you must implement rigorous evaluation workflows. Your LLM's classifications of nuanced claims, like those in seed oil narratives, may exhibit systematic errors and overconfidence, diverging significantly from expert judgment. Prioritize careful evaluation of model robustness, potentially employing a hierarchical error taxonomy to understand specific failure mechanisms. This approach ensures higher quality datasets and more reliable automated annotation in sensitive, domain-specific applications.

Key insights

Large language models frequently misinterpret nuanced health misinformation, leading to annotation failures compared to human experts.

Principles

• LLMs misinterpret nuanced domain-specific claims.
• LLMs exhibit overconfidence in their predictions.
• Systematic error patterns exist in LLM annotations.

Method

Evaluate LLMs as annotators using expert-labeled datasets and a hierarchical error taxonomy categorizing misclassifications by direction, mechanism, and contributing factors.

In practice

• Develop robust evaluation frameworks for LLM-based annotators.
• Inform LLM developers about common annotation failure modes.

Topics

• Large Language Models
• LLM Annotation
• Nutrition Misinformation
• Error Analysis
• Social Media Data

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Food Noise & False Safety: A Systematic Evaluation of How LLMs Fail to Adapt to Eating Disorder Queries with Clinician Feedback — LLMs uncritically adapt to eating disorder queries, increasing unsafe response likelihood with specific linguistic cues.
• Understanding LLMs in Title-Abstract Screening: From Disagreements to Recommendations — LLM failures in systematic review screening are rooted in identifiable semantic and inference issues, not merely quantitative disagreement.
• The 1.6 Billion People Large Language Models Still Can’t Understand — Large language models fail profoundly in many non-English languages, going beyond mere performance degradation to confident incorrectness.
• The Wrong Kind of Right: Quantifying and Localizing Misfired Alignment in LLMs — LLM alignment can misfire, causing models to reject evidence-supported conclusions, particularly concerning stereotypes.
• The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows — The LLM fallacy describes misattributing AI-assisted output to one's own competence, distorting self-perception.
• LLM as a Meta-Judge: Synthetic Data for NLP Evaluation Metric Validation — LLMs can generate synthetic data to validate NLP evaluation metrics, replacing costly human annotations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.