The panel described the current situation as an arms race between the capabilities of AI and available detection technologies. Traditional safeguards such as CAPTCHA, once an effective tool against automated responses, are becoming rapidly outdated. The panellists drew an important distinction between genuinely malicious actors deploying swarms of bots to exploit paid survey panels for financial gain, and ordinary participants who paste their open-ended answers into ChatGPT out of convenience. Both forms contaminate data, but each calls for different countermeasures. Tactical solutions such as “poison pill” sentences, tracking metadata like keystrokes and response times, and restricting copy-paste functionality offer only partial relief. The panel agreed that these are temporary measures and not structural solutions.
A more philosophical question that arose during the discussion was whether it still matters who or what completes a survey, if an AI-generated answer is indistinguishable from a human one. The panel answered that question with a clear yes. Large language models are trained on existing human data but do not reflect the unique lived experiences of individual people. They tend to exaggerate effect sizes, over-confirm hypotheses, and systematically misrepresent minority groups. When non-native speakers use AI to polish their open-ended answers, something is lost: the unfiltered tone, sentiment and individual voice that qualitative researchers rely on to extract richer insights from data. The panel also voiced a broader concern: as AI-generated content floods the internet, training data for future models becomes increasingly homogeneous, raising the risk of an epistemic feedback loop that could further distort scientific knowledge at scale.
Large-scale online samples remain indispensable for certain research designs; it is simply not possible to bring two thousand nationally representative respondents into a laboratory. At the same time, the speed and scale that online surveys enable, once considered their greatest advantages, now amplify specific vulnerabilities. The panel questioned the accelerated production of academic output, with publications accumulating at a rapid pace and reviewers increasingly turning to AI themselves. According to the panel, responses are needed at multiple levels: researchers must critically examine their data and report detection methods transparently, journals should enforce integrity standards comparable to open-science practices, and commercial survey panel providers must take responsibility for the authenticity of their respondents. Auditing for AI-generated responses is not an administrative burden, but a core requirement of rigorous scientific methodology.
The panel concluded on a note of cautious optimism. AI is also a powerful amplifier of research, for instance by automating text and video analysis and enabling large-scale questionnaire generation. The challenge for the scientific community is therefore to harness the benefits of AI without undermining the integrity of the data on which research depends.
