Deloitte at #SMM4H 2024: Can GPT-4 Detect COVID-19 Tweets Annotated by Itself?

Harika Abburi, Nirmala Pudota, Balaji Veeramani, Edward Bowen, Sanmitra BhattacharyaACL 2024

TL;DR

This paper explores whether GPT-4 can distinguish between tweets it annotated itself and those annotated by humans, specifically for COVID-19 symptom data. While GPT-4 showed promise, achieving a respectable 0.51 accuracy in a shared task, the results highlight the significant challenge of detecting AI-generated annotations, especially in nuanced domains like medical tweets.

What Problem Does This Paper Address?

The rise of powerful Large Language Models (LLMs) like GPT-4 has revolutionized text generation, making them invaluable for tasks like data annotation. However, LLMs can "hallucinate" or generate plausible but incorrect information, and their use in data annotation raises concerns about inherent biases and the reliability of the annotations. This paper tackles the critical problem of detecting whether data annotations were produced by an AI model (specifically GPT-4) or a human expert. This is crucial for ensuring the integrity of datasets used in downstream applications, particularly in sensitive fields like healthcare.

Methods

The core of this research lies in using GPT-4 itself as a zero-shot detector. Instead of fine-tuning a separate model, the authors directly prompt GPT-4 to classify whether a given tweet's annotation was made by a human or an AI. They experimented with various prompts, ultimately finding that simpler prompts yielded better results. The chosen prompt frames GPT-4 as a data annotation expert tasked with discerning the origin of a COVID-19 symptom tweet's annotation. The model is asked to output a single word: "human" or "generated." To ensure consistent results, the temperature parameter was set to 0.

Datasets

The study utilized a dataset provided by the SMM4H (Social Media Mining for Health Research and Applications) shared task. This dataset consists of tweets detailing COVID-19 symptoms written in Latin American Spanish. It includes both human-annotated and GPT-4-annotated data. Specifically, 3,682 tweets were available for training and 2,110 for testing. The paper also references other zero-shot detection models (Binoculars and Fast-DetectGPT) for comparative evaluation, implying their use of standard benchmarks or datasets relevant to AI-generated text detection.

Key Insights and Ideas

  • GPT-4 as its own detective: The paper innovatively uses GPT-4 to identify its own annotations, bypassing the need for external models or extensive fine-tuning.
  • Simplicity is key for prompts: Complex prompts can lead GPT-4 astray; simpler, direct instructions are more effective for this specific detection task.
  • Medical tweets are a tough nut to crack: The nuanced and often subtle nature of medical information in short tweets presents a significant challenge for AI detection.
  • The "AI annotation detection" problem is hard: Even with a powerful model like GPT-4, achieving high accuracy in distinguishing AI from human annotations is difficult, with top scores hovering around 0.5.

Key Results

The authors' GPT-4 based approach achieved an accuracy of 0.51 in the SMM4H shared task, securing second place. While this is a respectable performance, it highlights the difficulty of the task. Across the three evaluation sets (Set 1, Set 2, Set 3), GPT-4 consistently outperformed baseline zero-shot detection models like Binoculars and Fast-DetectGPT in terms of accuracy (Acc) and F1-macro (Fmac), though the absolute scores remained modest (e.g., Acc around 0.49, Fmac around 0.40-0.42).

Limitations and Future Work

The paper acknowledges that the achieved accuracy of 0.51, while placing them second in the shared task, underscores the significant challenge of AI annotation detection, especially for short texts in the medical domain. For future work, the authors aim to improve performance by incorporating the reasoning behind GPT-4's predictions as additional input. This could help the model better differentiate between human and AI annotations by understanding the underlying logic.

Summary updated

BibTeX
@inproceedings{abburi2024deloitte,
  author = {Harika Abburi and Nirmala Pudota and Balaji Veeramani and Edward Bowen and Sanmitra Bhattacharya},
  title = {{Deloitte at \#SMM4H 2024: Can GPT-4 Detect COVID-19 Tweets Annotated by Itself?}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure