Our Investigation: What we did
To learn about workers using server farms and how effective TurkPrime’s solution is at blocking them, we conducted a study which included a series of data quality checks. Our study examined response characteristics and data quality among workers known to operate through server farms (i.e., farmers) and workers known to operate without server farms (i.e., non-farmers). We refer to workers completing HITs from server farms as ‘farmers’ instead of ‘bots,’ because the term ‘farmer’ is agnostic as to whether these workers are humans, bots, or humans using some form of automation.
In addition to measures of data quality, we included several “bot-detection” measures in our study. Across these measures we found no evidence of bots—not one worker failed the re-captcha or honeypot. We did, however, find substantial evidence that farmers may be foreign workers operating outside the US. Farmers failed a basic English proficiency screener that 96% of non-farmers passed. In addition, while 95% of non-farmers referred to a pictured vegetable as an eggplant, 89% of farmers called the vegetable a ‘brinjal.’
Consistent with the idea that farmers are foreign workers with low English proficiency, we found they do well on tasks that involve short instructions written in basic English. For example, farmers responded correctly to anchoring questions about the height of Mt. Everest and although they had a smaller effect size on the Mt. Everest anchoring task the manipulation was still effective for them. This evidence in combination with farmers’ apparent knowledge about the height of Mt. Everest—a mountain in Nepal, just to the north of India—strongly implies that these workers are humans based in India.