Recognising, Anticipating, and Mitigating LLM Pollution of Online Behavioural Research

Raluca Rilla, Tobias Werner, Hiromu Yakura, Iyad Rahwan, Anne-Marie NussbergerarXiv 2025 · arXiv

TL;DR

Online behavioral research is facing an existential threat from "LLM Pollution"—the contamination of human data by large language models used by study participants. This paper provides a conceptual framework defining how LLMs distort human subject data and outlines a multi-layered defense strategy to safeguard the integrity of online behavioral science.

What Problem Does This Paper Address?

Behavioral science relies heavily on crowdsourcing platforms like Prolific and Amazon Mechanical Turk (MTurk) to gather fast, diverse human data for cognitive and psychological studies. However, the ubiquity of highly capable LLMs has broken the promise of "100% human" responses.

Participants are increasingly outsourcing parts or all of their study tasks to AI. This "LLM Pollution" homogenizes data, reduces natural human variance, introduces systematic cultural biases inherent in AI training data, and threatens the core epistemic foundation of human-subject research.

Methods

Rather than introducing a new algorithm, this paper provides a conceptual perspective and a practical, multi-layered mitigation framework. The authors categorize LLM Pollution into three distinct variants:

Figure 1: Three Variants of LLM Pollution. Partial LLM Mediation refers to cases where partici- pants use LLMs for translation, idea generation, or performance gains. Full LLM Delegation involves the use of agentic tools or plugins to automate participation entirely. LLM Spillover refers to normative or behavioural shifts prompted by participants’ beliefs about LLM involvement, regardless of whether any is present. These variants can interact and reinforce one another, compounding their threat to research validity and inference. Illustration by H. Jahani.

To combat these variants, the authors synthesize a suite of multi-layered defense mechanisms across the study lifecycle:

  • Preventative Measures: Norm signaling (explicitly asking participants not to use AI), multimodal instructions (using images, audio, or video to prevent easy copy-pasting), and LLM-specific comprehension checks (tasks that contemporary LLMs fail, such as specific visual illusions or modified Theory of Mind questions).
  • Post-hoc Detection: Invisible "honeypot" questions (text hidden via CSS that only scraping bots read and answer), fine-grained behavioral logging (tracking keystroke dynamics, mouse paths, and tab-switching), and commercial AI text detectors.

Datasets

The paper does not introduce a new machine learning benchmark dataset. Instead, it presents an illustrative case study of participants recruited via Prolific to demonstrate the real-world incidence of LLM pollution and evaluate the effectiveness of various defensive measures. It also synthesizes empirical data from recent adjacent studies (e.g., Veselovsky et al., 2025; Zhang et al., 2025).

Key Insights and Ideas

  • The Automation Continuum: LLM pollution isn't a binary "human vs. bot" issue. It exists on a spectrum from Partial Mediation (using AI for translation or phrasing support) to Full Delegation (employing autonomous browser agents like OpenAI's ChatGPT Agent to complete entire studies).
  • LLM Spillover (The Psychological Echo): Even when participants do not use LLMs, their data can still be polluted by the mere expectation of AI. For instance, humans might intentionally introduce typos ("strategic signaling") to prove to researchers they aren't a bot, or alter their economic game choices if they suspect their partner is an AI.
  • The Mitigation Catch-22: Aggressive, highly visible bot-detection measures can backfire. They risk alienating honest human participants, increasing attrition, and inadvertently triggering the very "Spillover" reactivity effects researchers want to avoid.

Key Results

The authors' case study highlights the immediacy of the threat:

  • In an initial pilot study lacking bot protections, 45% of participants engaged in copying and pasting on an open-ended question page, with many responses clearly showing LLM-generated traits (e.g., stating "I don't experience confusion in the same way humans do").
  • Implementing a layered defense in the main study significantly reduced the threat but revealed gaps: 0.2% failed reCAPTCHA v2, 2.7% were flagged by reCAPTCHA v3, 1.6% fell into the "honeypot" trap, and 4.7% attempted to copy/paste.
  • Despite these layers, a few identical, nonsensical responses still slipped through all filters undetected, proving that no single tool is a silver bullet.

The full spectrum of analyzed measures and their associated risks is detailed in the paper's summary table:

List of possible measures for detecting and mitigating LLM Pollution

Limitations and Future Work

The authors openly acknowledge that technical mitigations are an escalating arms race. As LLMs rapidly improve in vision-language alignment and reasoning, current "LLM-proof" comprehension checks (like optical illusions) will quickly become obsolete. Furthermore, fine-grained behavioral tracking (like keystroke dynamics) presents ethical and legal hurdles and risks unfairly penalizing participants who use legitimate assistive technologies.

Future work must move beyond individual researcher fixes toward platform-level accountability (e.g., stricter terms of service, automated infrastructure) and community-wide sharing of best practices. Ultimately, researchers may need to ask a deeper question: as AI becomes deeply embedded in daily life, at what point does LLM-assisted behavior stop being "pollution" and instead become the new ecological baseline of human cognition?

Summary updated

Abstract

Online behavioural research faces an emerging threat as participants increasingly turn to large language models (LLMs) for advice, translation, or task delegation: LLM Pollution. We identify three interacting variants through which LLM Pollution threatens the validity and integrity of online behavioural research. First, Partial LLM Mediation occurs when participants make selective use of LLMs for specific aspects of a task, such as translation or wording support, leading researchers to (mis)interpret LLM-shaped outputs as human ones. Second, Full LLM Delegation arises when agentic LLMs complete studies with little to no human oversight, undermining the central premise of human-subject research at a more foundational level. Third, LLM Spillover signifies human participants altering their behaviour as they begin to anticipate LLM presence in online studies, even when none are involved. While Partial Mediation and Full Delegation form a continuum of increasing automation, LLM Spillover reflects second-order reactivity effects. Together, these variants interact and generate cascading distortions that compromise sample authenticity, introduce biases that are difficult to detect post hoc, and ultimately undermine the epistemic grounding of online research on human cognition and behaviour. Crucially, the threat of LLM Pollution is already co-evolving with advances in generative AI, creating an escalating methodological arms race. To address this, we propose a multi-layered response spanning researcher practices, platform accountability, and community efforts. As the challenge evolves, coordinated adaptation will be essential to safeguard methodological integrity and preserve the validity of online behavioural research.
BibTeX
@misc{rilla2025recognising,
  author = {Raluca Rilla and Tobias Werner and Hiromu Yakura and Iyad Rahwan and Anne-Marie Nussberger},
  title = {{Recognising, Anticipating, and Mitigating LLM Pollution of Online Behavioural Research}},
  year = {2025},
  eprint = {2508.01390},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.01390},
}
Figure