Why you shouldn’t trust data collected on MTurk
Abstract
Several prior studies have used advanced methodological techniques to demonstrate that there is an issue with the quality of data that can be collected on Amazon’s Mechanical Turk (MTurk). The goal of the present project was to provide an accessible demonstration of this issue. We administered 27 semantic antonyms—pairs of items that assess clearly contradictory content (e.g., “I talk a lot” and “I rarely talk”)—to samples drawn from Connect (N1 = 100), Prolific (N2 = 100), and MTurk (N3 = 400; N4 = 600). Despite most of these item pairs being negatively correlated on Connect and Prolific, over 96% were positively correlated on MTurk. This issue could not be remedied by screening the data using common attention check measures nor by recruiting only “high-productivity” and “high-reputation” participants. These findings provide clear evidence that data collected on MTurk simply cannot be trusted.
Study specs
27 semantic antonym pairs were administered to participants from Connect (N=100), Prolific (N=100), and MTurk (N=400, N=600) to examine response quality and correlation patterns.
- Authors
- CS Kay
- Institution
- Stanford University
- Sample Size
- N=1,200
- Study Type
- Experimental Study
- Year
- 2025
- Human Data Platform
- Prolific
- Source
- View Source Google Scholar
Measured Outcomes
The correlation of responses to semantic antonym pairs as an indicator of data quality across different survey platforms.
Peer Review & Critical Discussion
Potential Selection Bias in 2023 Cohort
The participant pool shows a concerning overrepresentation of users from high-income demographics. Looking at Table 3, we can see that 78% of respondents had annual incomes above $75k, which significantly limits the generalizability of these findings to broader populations.
Non-naive Participants Issue
I've noticed a methodological concern regarding participant naivety. Given that Prolific users often complete multiple studies, there's a real risk that participants had prior exposure to similar experimental paradigms, which could confound the results.
RLHF Applicability to This Study Design
The implications for RLHF training pipelines are understated. If we accept the authors' conclusions about preference stability, this has direct consequences for how we should structure reward model training. The temporal decay effect described in Section 4.2 is particularly relevant.
Verify your expertise to join discussion
Create an account and verify your credentials to participate in peer discussions.