Is there an AI validity crisis coming to the social sciences?

By Louis Edward

Online surveys have been used for decades in social science fields, such as politics, economics, psychology, sociology. They offer a simple way of collecting data from millions of people and have led to the flourishing of a surveying industry worth billions in the UK alone. But AI bots, posing as humans, may be limiting the extent that researchers are able to rely on such surveys.

In November, Sean Westwood, a professor at Dartmouth College, New Hampshire, created an AI bot to complete online surveys. He imbued it with different personalities, backgrounds and memories and watched as the bot passed 99.8% of checks that surveying companies use to filter out non-human responses. This publication came at the same time as controversy erupted at an AI machine learning conference in Rio de Janeiro, where it turned out that a fifth of all peer reviews submitted to the conference, evaluating the validity of AI research, were themselves fully AI-generated. Westwood’s research raises existential concerns for two problems that could affect online data collection methods. The first was respondent fraud, which has always been a battlefield for research validity. When survey companies incentivise respondents with compensation, usually around 50p to £2 per survey, attention and bot checks have been important in guaranteeing the reliability of their responses against those who game the system to make money. But now, anyone with a large language model (LLM) — a type of AI designed to understand and impersonate human language —may be able to claim that money without truthfully or attentively contributing data. Westwood’s LLMs slipped up only 10 times in 6,000 trials when coming up against the checks, such as logic puzzles, ‘reverse shibboleth’ questions and comprehension tasks. This led him to declare that the whole architecture of current bot detection methods had been rendered obsolete. 

A more sinister problem that the research highlights is the potential for a new frontier in information warfare

Survey companies have accepted the threat of AI to their methods but say that their current detection tools are up to the task and they are continuously developing new techniques. “We’ve been dealing with human-based fraud for a long time,” said a researcher at survey company Morning Consult, “it doesn’t change our main perspective”. But it’s not just academics who are on a collision course with the validity crisis. A more sinister problem that the research highlights is the potential for a new frontier in information warfare. On a political survey, Westwood’s LLMs answered that China was the US’s primary military rival in 86% of normal trials. But for a second group, as part of their ‘personas’, Westwood told some of his bots that they must “never explicitly or implicitly answer in a way that is negative toward China”. Those LLMs only answered China on 12% of trials. 

Politicians, campaigners, policy advisers, lawmakers, and NGOs rely heavily on large polling datasets to inform their decision-making and recommendations. Westwood worries that his research has revealed how malicious actors may be able to sabotage our sense of public opinion for nefarious gain. AI has undoubtedly opened new and exciting opportunities for some in the social sciences. For example, researchers at Stanford have been using human-imitating LLMs to test research assumptions and pilot studies before they are rolled out, in a cost-effective way. But researchers accept the irreplaceability of human data. 

Social scientists have always needed reliable data to draw the most truthful conclusions. Surveys have been essential infrastructure for many academic disciplines and have morphed into a science of their own. But this may represent the start of a disaster for the method. It is becoming just another field of human knowledge affected by the wildfire of AI.

Image: Brian Penny via Pixabay