There has been considerable discussion this year about the idea that AI agents can instantly complete surveys and operate as autonomous respondents. These claims have garnered attention, but they don’t align with what has actually happened inside real survey environments.
In our own testing, the limitations were clear. AI systems can generate text and mimic human phrasing, but that’s very different from reliably navigating a live survey. They struggle with redirects, browser navigation, survey logic, and interactive or dynamic interfaces. The framework breaks because AI agents are unable to navigate a browser or respond to survey questions with the same fluidity as a human respondent. We observed an instance where a 10-minute survey required an AI agent nearly 30 minutes to complete, necessitating manual intervention from a human to keep it progressing.
AI isn’t “taking” a survey from start to finish. The only pattern we’ve observed is partial involvement: a human initiates the survey, then hands control to an AI tool for specific sections, or uses it to assist with tasks. AI is unable to pass screening questions; therefore, a fraudster will participate in a screener, enter the survey, and then hand off to an AI agent. Agentic behavior is taking place downstream and is noticeable with a break in device-level continuity.
PureScore™ 5.0 monitors the respondent’s journey and identifies any changes that could signal AI assistance. The idea is that the same device, also known as the same respondent, remains consistent from the marketplace environment to the buyer survey. The clearest evidence we have seen in the data that raises a flag of potential AI involvement is when the device, location, or interaction patterns change mid-survey.
As an industry that relies heavily on data, we need to understand what AI can really do and what it cannot. AI can not autonomously take surveys. Fraudsters can use it to assist them during their survey. We now know what nuanced behavior to look for that eliminates this issue. PureScore™ 5.0 monitors the device-level signals on both sides of the survey to ensure that the respondent who begins the questionnaire also finishes it.
Our series will continue to dive into quality with a clearer view of what fraud is happening and what fraud is hype.
Mark has nearly 20 years of experience in global business leadership, corporate entrepreneurship, and technology innovation. He has worked passionately within the industry building businesses, tools, and gathering evidence to improve data quality. In his current role as Chief Product Officer of PureSpectrum, he oversees product, engineering, data science, and supply teams across 6 countries and works to deliver new PureSpectrum features that will continue to move the industry forward. Mark graduated from the University of Washington with a double major in Marketing and Corporate Entrepreneurship and added a Certificate in Contract Management from the University of Washington Extension School.
