[Paper Review] Can social media provide early warning of retraction? Evidence from critical tweets identified by human annotation and large language models

Er‐Te Zheng, Hui‐Zhen Fu+2 more|arXiv (Cornell University)|Mar 25, 2024

Artificial Intelligence in Healthcare and EducationMedicine3 citations

TL;DR

This study investigates whether social media, specifically critical tweets, can serve as an early warning system for retracted scientific articles. By combining human-annotated critical tweets (8.3% of retracted articles vs. 1.5% of non-retracted ones) with LLM-based detection (GPT-4o mini, Gemini 2.0 Flash-Lite, Claude 3.5 Haiku), it finds that human-annotated signals are significantly more reliable than LLMs alone, suggesting a human-AI collaborative approach is essential for scalable and accurate research integrity monitoring.

ABSTRACT

Timely detection of problematic research is essential for safeguarding scientific integrity. To explore whether social media commentary can serve as an early indicator of potentially problematic articles, this study analysed 3,815 tweets referencing 604 retracted articles and 3,373 tweets referencing 668 comparable non-retracted articles. Tweets critical of the articles were identified through both human annotation and large language models (LLMs). Human annotation revealed that 8.3% of retracted articles were associated with at least one critical tweet prior to retraction, compared to only 1.5% of non-retracted articles, highlighting the potential of tweets as early warning signals of retraction. However, critical tweets identified by LLMs (GPT-4o mini, Gemini 2.0 Flash-Lite, and Claude 3.5 Haiku) only partially aligned with human annotation, suggesting that fully automated monitoring of post-publication discourse should be applied with caution. A human-AI collaborative approach may offer a more reliable and scalable alternative, with human expertise helping to filter out tweets critical of issues unrelated to the research integrity of the articles. Overall, this study provides insights into how social media signals, combined with generative AI technologies, may support efforts to strengthen research integrity.

Motivation & Objective

• To assess whether critical social media commentary on scientific articles can serve as an early warning signal for retractions.
• To compare the effectiveness of human-annotated critical tweets versus large language model (LLM)-detected tweets in identifying articles later retracted.
• To evaluate the reliability and alignment of LLMs in detecting research integrity-related criticism compared to human judgment.
• To propose a human-AI collaborative framework for scalable and accurate monitoring of post-publication discourse for research integrity.

Proposed method

• Collected 3,815 tweets referencing 604 retracted articles and 3,373 tweets referencing 668 non-retracted articles.
• Used human annotation to identify tweets critical of research integrity, establishing a gold-standard dataset.
• Employed three LLMs—GPT-4o mini, Gemini 2.0 Flash-Lite, and Claude 3.5 Haiku—to classify tweets as critical or non-critical.
• Compared LLM predictions against human-annotated labels to assess alignment and reliability.
• Calculated precision, recall, and F1 scores to evaluate LLM performance in detecting integrity-related criticism.
• Proposed a human-AI collaborative model where human oversight filters out non-integrity-related criticism from LLM outputs.

Experimental results

Research questions

• RQ1 Can critical tweets on social media serve as an early indicator of scientific articles that are later retracted?
• RQ2 How well do large language models detect research integrity-related criticism in tweets compared to human-annotated standards?
• RQ3 What is the level of alignment between human-annotated critical tweets and LLM-generated classifications of such tweets?
• RQ4 To what extent can LLMs alone be trusted for automated monitoring of post-publication discourse for research integrity?
• RQ5 What role can a human-AI collaborative approach play in improving the scalability and accuracy of early warning systems for retractions?

Key findings

• 8.3% of retracted articles were associated with at least one human-annotated critical tweet prior to retraction, compared to only 1.5% of non-retracted articles.
• The human-annotated dataset showed a strong signal: critical commentary preceded retraction in a significant minority of cases.
• LLM-detection of critical tweets showed only partial alignment with human-annotated labels, indicating limited reliability for full automation.
• GPT-4o mini, Gemini 2.0 Flash-Lite, and Claude 3.5 Haiku all exhibited notable false positive and false negative rates in identifying integrity-related criticism.
• The study concludes that fully automated LLM monitoring should be applied with caution due to misclassification risks.
• A human-AI collaborative approach is recommended, where human expertise filters and validates LLM outputs to improve accuracy and relevance.