[Paper Review] Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks

Veniamin Veselovsky, Manoel Horta Ribeiro+1 more|arXiv (Cornell University)|Jun 13, 2023

Mobile Crowdsensing and CrowdsourcingComputer Science71 citations

TL;DR

The paper develops a detection method to estimate how often crowd workers on MTurk use LLMs to produce text, and finds 33–46% of summaries were LLM-generated in a specific abstract-summarization task.

ABSTRACT

Large language models (LLMs) are remarkable data annotators. They can be used to generate high-fidelity supervised training data, as well as survey and experimental data. With the widespread adoption of LLMs, human gold--standard annotations are key to understanding the capabilities of LLMs and the validity of their results. However, crowdsourcing, an important, inexpensive way to obtain human annotations, may itself be impacted by LLMs, as crowd workers have financial incentives to use LLMs to increase their productivity and income. To investigate this concern, we conducted a case study on the prevalence of LLM usage by crowd workers. We reran an abstract summarization task from the literature on Amazon Mechanical Turk and, through a combination of keystroke detection and synthetic text classification, estimate that 33-46% of crowd workers used LLMs when completing the task. Although generalization to other, less LLM-friendly tasks is unclear, our results call for platforms, researchers, and crowd workers to find new ways to ensure that human data remain human, perhaps using the methodology proposed here as a stepping stone. Code/data: https://github.com/epfl-dlab/GPTurk

Motivation & Objective

• Quantify the prevalence of LLM usage among crowd workers in a text summarization task.
• Develop a low-cost, task-specific detector to distinguish human-written vs. LLM-generated summaries.
• Validate the detector using keystroke data and overlap analysis to corroborate results.
• Assess implications for the reliability of human-generated data in crowdsourcing contexts.

Proposed method

• Fine-tune an e5-base classifier to distinguish ChatGPT-generated vs real summaries using a contrastive-pretraining foundation.
• Train on real human summaries and synthetic ChatGPT-generated summaries derived from MTurk task prompts.
• Evaluate across two data-split settings: abstract-level and summary-level, examining generalization across abstracts and summaries.
• Apply the detector to 46 new summaries to estimate LLM usage among crowd workers, reporting prevalence across threshold variants.
• Use post-hoc keystroke-based validation to assess classifier reliability and false-positive rates.

Figure 1: Illustration of our approach for quantifying the prevalence of LLM usage among crowd workers solving a text summarization task. First, we use truly human-written MTurk responses and synthetic LLM-written responses to train a task-specific synthetic-vs.-real classifier. Second, we use this

Experimental results

Research questions

• RQ1 What is the prevalence of LLM usage among crowd workers performing an MTurk abstract-summarization task?
• RQ2 Can a task-specific, low-computation detector reliably distinguish synthetic from human-written summaries in this setting?
• RQ3 Do keystroke and overlap analyses corroborate the detector’s classifications of synthetic vs human text?

Key findings

• The synthetic-vs-real classifier achieved 99% accuracy and 99% macro-F1 in the summary-level split.
• The same classifier achieved 97% accuracy and 97% macro-F1 in the abstract-level split, indicating good generalization.
• Using a logit threshold of 0 (50% probability), 46% (21/46) of summaries were classified as synthetic; with a higher threshold of 4 (98% probability), 33% (15/46) were classified as synthetic.
• Across thresholds, the estimated prevalence of LLM-generated summaries remained between 33% and 46%.
• Post-hoc analysis showed most workers pasted text, but the classifier labeled those with high paste usage differently, suggesting genuine LLM-generated outputs rather than mere copying.
• Less overlap with the original abstract in synthetic classifications supports the presence of LLM-generated content.

Figure 2: Depiction of the MTurk task studied in this paper, where crowd workers were asked to condense research abstracts from the New England Journal of Medicine into summaries about 100 words long.