Source-linked AI summary
Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks
Veniamin Veselovsky, Manoel Horta Ribeiro, Robert West
TL;DR
Crowd workers may use LLMs to increase productivity, threatening the use of crowdsourced responses as human gold-standard data. The paper reruns an MTurk abstract-summarization task and combines synthetic-text classification with keystroke validation. It estimates that 33–46% of submitted summaries involved LLM use, while cautioning that generalization beyond summarization is uncertain.
Problem
Crowd workers’ LLM use could undermine crowdsourcing as a source of human gold-standard data for validating LLMs and conducting research.
Method
The study reruns an MTurk abstract-summarization task and combines keystroke detection with a bespoke synthetic-text classifier trained on human and ChatGPT-generated text.
Results
33–46% of crowd-worker summaries were estimated to have been produced with LLM assistance.
Takeaways & Limitations
The findings call for platforms, researchers, and crowd workers to develop ways to ensure that crowdsourced human data remain human.
Takeaways & Limitations
The study examines one text-summarization task, so generalization to other tasks, especially those challenging for LLMs, remains uncertain.
Abstract
from arXiv · showhide
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
1 Introduction
Crowdsourcing is widely used for human-generated data, but crowd workers may use LLMs to increase productivity, undermining the intended human gold standard. This study develops a detection-based case study estimating LLM use in MTurk abstract summarization.
- Crowdsourcing platforms support data creation, annotation, summarization, surveys, and experiments across research and industry.
- LLMs can annotate data, simulate human behavior, and generate high-quality text, making human gold-standard data important for validation.
- Crowd workers may use LLMs to increase productivity and income, reducing the utility of crowdsourced data as an intended human gold standard.
- The study combines keystroke detection and synthetic-text classification to estimate LLM use in an MTurk abstract-summarization task.
- 33–46% of summaries submitted by crowd workers were estimated to have been produced with LLM assistance.
- The authors conclude that crowdsourced text is already produced to a large extent by machines and call for ways to keep human data human.
2 Related work
Related work establishes crowdsourcing as infrastructure for research and industry, while prior studies examine its use, quality, and worker conditions. Parallel research shows that LLMs can generate data, proxy human populations, and produce text, motivating methods to detect synthetic data.
- Research using crowdsourcing: MTurk helped shift machine-learning dataset creation and user research toward crowd work performed by humans through an apparent artificial-intelligence interface.
- Research about crowdsourcing: Crowdsourcing research studies efficient task design, annotation quality, and the demographics and socioeconomic conditions of platform workers.
- LLM-generated data: LLMs have been used as “silicon samples” to proxy human subpopulations in controlled generation, surveys, and experiments.
- LLM-generated data: Prior work reports strong LLM performance for human labeling and high-quality text generation, while also noting factual inaccuracy in generated text.
3 Methods
The study reruns a shortened MTurk abstract-summarization task and detects synthetic text with a task-specific classifier trained on human and ChatGPT-generated examples. Keystroke logging provides post-hoc validation of the estimates.
- Task of choice: abstract summarization: Workers condensed New England Journal of Medicine abstracts into concise summaries, reducing the original multi-round task to one approximately 100-word summarization step.The original task involved 16 abstracts across vaccination, breast cancer, cardiovascular disease, and nutrition.
- Task of choice: abstract summarization: The task was selected because summarization is laborious for humans, easy for LLMs, and requires genuinely human text for studying information loss.
- Data collection: 46 de-duplicated summaries from 44 workers were collected, alongside JavaScript-recorded keystrokes including copy-and-paste actions.
- Detecting synthetic text: Out-of-the-box detectors performed poorly, detecting only six of 10 ChatGPT-synthesized summaries with GPTZero.
- Detecting synthetic text: The authors therefore fine-tuned a bespoke ChatGPT detector using human MTurk text and ChatGPT summaries generated from the task instruction.
- Detecting synthetic text: The classifier was evaluated with abstract-level and summary-level train/test splits to assess performance on pooled and unseen abstracts.
- Post-hoc validation: Keystroke-based heuristics were used after classification to validate whether subsets of summaries were synthetically or human generated.
4 Results
The detector performed strongly across both evaluation splits, and estimates indicated that 33–46% of crowdsourced summaries were produced with LLM assistance. Keystroke and overlap analyses provided post-hoc support for distinguishing synthetic from human-written summaries.
- Synthetic-text detector: 99% accuracy and 99% macro-F1 were achieved in the summary-level split, where summaries from all abstracts were pooled before splitting.The result suggests the detector identified fingerprint-like artifacts in ChatGPT summaries.
- Synthetic-text detector: 97% accuracy and 97% macro-F1 were achieved in the abstract-level split, indicating artifacts generalized to abstracts unseen during training.The authors describe these artifacts as universal across the abstracts in this task.
- LLM-use prevalence: 46% of summaries were classified as synthetic at logit threshold 0, while 33% were classified as synthetic at threshold 4.These corresponded to 21 of 46 summaries (95% CI: [31%, 61%]) and 15 of 46 summaries (95% CI: [20%, 45%]), respectively.
- Post-hoc validation: 89% of summaries involved pasting, while all five summaries written entirely in the text box were classified as human-written.The latter result suggested a low false-positive rate under the assumption that summaries without pasting were human-generated.
- Post-hoc validation: Among 13 summaries with less than 10% overlap with the original abstract, 10 (76%) were classified as synthetic.Synthetic-classified summaries generally had small overlap, supporting the interpretation that pasted text often came from LLM outputs rather than the original abstracts.
5 Discussion
The discussion argues that widespread LLM use by crowd workers threatens the human character of crowdsourced text data and may make human data harder to obtain. It also identifies possible shifts toward human filtering of model outputs while limiting generalization beyond summarization and other LLM-friendly tasks.
- Implications: Crowd workers’ LLM use raises concerns about the dilution of the human factor in crowdsourced text data.The paper connects this concern to the richness, uniqueness, and diversity attributed to human-generated data.
- Implications: The popularization and increasing capabilities of LLMs may make acquiring human data harder.The paper frames this as a concern about the future availability of genuinely human-produced data.
- Implications: Crowd workers may increasingly serve as human filters that identify when language models succeed or fail rather than provide de novo annotations.The discussion presents this as a possible shift in the value provided by crowd work, not its end.
- Limitations: The findings’ generalization is uncertain because the study examined only one text-summarization task, especially for tasks that substantially challenge LLMs.The authors call for research across tasks and over time as LLMs become more widespread.
6 Ethical considerations
The study used keystroke collection to validate its results, but broader tracking could potentially infringe on user privacy if not handled appropriately.
- Keystroke tracking was restricted to interactions with the edit box and copy-pastes on the page.
- More expansive keystroke tracking could potentially create user-privacy problems.