Running a fake human on your survey just got a lot cheaper.
A new study compared nine AI agent setups - from fully open, locally run models to closed commercial ones - and had each autonomously fill out a survey built with multiple detection traps. The free, open-weight agents ran with no usage fees and matched the commercial agents on performance. Open and commercial agents tripped different detection checks, and no single check caught every agent. Open-ended text answers turned out to be the strongest signal for telling agent from human.
That matters because survey research already leans on the assumption that respondents are people, not scripts. If anyone with a laptop and an open-weight model can generate competitive synthetic responses for free, the cost barrier that used to limit this kind of pollution is gone. The researchers' answer is to stack multiple detection methods instead of relying on one, with open-text analysis doing the heaviest lifting.
Detection tools built for a world of expensive, closed models were never going to survive contact with free ones.