Skip to main content
Back

LLMs Are Quietly Poisoning the Human Data That Trains Them

A survey of 155 NLP researchers reveals that 44% have found LLM-generated text in their crowdsourced datasets, and most don't know how to stop it.

Paper, Full Access
  • 44% of surveyed NLP researchers have observed LLM-generated text in their crowdsourced datasets, per a new survey of 155 researchers.
  • Awareness without action: 93% of affected researchers anticipated the problem, but half were unsure what precautions to take.
  • Top detection methods are informal: unusual writing style and suspiciously fast task completion times.
  • Prior studies estimate 30-40% of crowdworkers use LLMs for text production tasks, even when asked not to.
  • Model collapse risk: LLM-generated crowdsourced data fed back into training pipelines can degrade future model diversity and quality.
  • No silver bullet exists; the paper recommends multi-layered defenses including task design, behavioral monitoring, and explicit policies.

Crowdsourcing has been the backbone of NLP data collection for over a decade. Platforms like Amazon Mechanical Turk and Prolific powered the creation of foundational datasets and benchmarks that trained the very models now threatening to replace the workers who built them. A new paper asks the uncomfortable question: can crowdsourcing survive the LLM era?

The widespread use of LLMs as writing tools challenges the validity of crowdsourced data, as crowdworkers may outsource tasks to models. To better understand how this is being addressed, the authors surveyed 155 researchers in NLP and related disciplines about their experiences collecting free-text responses via crowdsourcing. The results paint a picture of a community that sees the problem clearly but is struggling to act on it.

The Scale of the Problem

Prior work estimated that 30-40% of crowdworkers rely on LLMs for text production, and a separate study found that 34% of Prolific participants self-report using LLMs for open-ended questions. This isn't a fringe behavior. It's becoming a default workflow for a significant share of the crowd labor market.

Platforms like Amazon Mechanical Turk and Prolific are widely used by researchers for tasks such as data labeling and survey administration. However, the integrity of this data collection process is increasingly threatened by the very AI models it seeks to improve. The irony is hard to miss: the human feedback pipelines that trained ChatGPT are now being filled by ChatGPT.

Process flowchart showing how human-written, LLM-assisted, and LLM-generated responses all flow into crowdsourced datasets, creating data integrity concerns

What the Survey Found

44% of respondents reported observing LLM usage in their crowdsourced data. While 93% of them had anticipated this, half were unsure what precautions to take. That gap between awareness and action is the central finding. Researchers know the problem is real, but the field hasn't converged on reliable solutions.