Weak Programmers Need Not Apply, LLMs Welcome! Survey Screening in the AI Era

Quantitative research on software development practice often depends on the results of anonymous online surveys. However, it is difficult to verify that the respondents to such surveys are genuine developers, especially if survey responses are paid for, incentivising respondents to game the system and attracting the attention of bots and professional survey farms. One solution is to include screening questions to assess respondents’ coding skills, but this introduces the risk of unfairly excluding considered responses from developers who spent time and energy on the survey. We reviewed the responses of 86 developers who were excluded via screening from analysis of a large unpaid developer survey (n=1,048). We estimated that up to 86% of the developers who were excluded were genuine developers. The advent of large language models (LLMs) casts further doubt on the use of screening questions. Investigating LLM ability, we found that five sample LLMs could answer most widely-used screening questions. Powerful LLM-based survey-taking tools now exist. We researched current survey screening techniques and found that they are susceptible to fraud via LLM use, bots and survey farms. Recruitment strategy may be a better screening technique. We recommend that if survey respondents are compensated only known developers should be invited. Surveys that are widely distributed, e.g. on social media, should not compensate respondents. Instead, researchers should focus on good survey design and sustained, imaginative recruitment strategies.