“Long story short, in the age of AI, if you see a survey response that you can’t triangulate with other data and trends, then you need to be very wary”
If you’ve not listened to it before, BBC Radio 4’s More or Less is a show that explains, questions or debunks numbers and statistics across the news and everyday life. As researchers, we love it. However, in an episode last month, the combination of research data and discussions around survey fraud made for a particularly exciting listen. In a segment on the 20th May, Tim Harford reignited the conversation around ‘The Quiet Revival’ and the striking claims taken from a YouGov survey commissioned by the Bible Society. The latest development is that YouGov has since withdrawn its findings after finding fraudulent respondents in its sample.
If fraudulent respondents can get through YouGov’s systems undetected and invalidate a nationally covered study, the question isn’t whether your data could be affected. It’s whether it already has been.
As the More or Less panel discussed, some of these bad respondents were probably people misrepresenting who they are and others were very likely AI agents. Annette Jackle, a Professor of Survey Methodology at the University of Essex highlighted that AI has become so good at impersonating real respondents that telling them apart is nearly impossible. Nearly….
How do AI agents take surveys?
An AI agent is a model that can be given a goal, for example, “complete this survey as a 24 year old man living in London”, and complete this task using a real browser and software. It can maintain a persona, answer open-text questions consistently, add typos, pause before responding, and move a mouse around the page.
Survey fraudsters then build these agents to infiltrate panels and earn money by generating fake responses. As the YouGov example illustrates, these AI agents can skew the results of an otherwise legitimate study.
AI agents’ abilities make them nearly impossible to identify using traditional data quality checks. However, these agents’ behaviour still strays from human behaviour, sometimes in minute detail. It’s these small details and differentiators that we can use for identification. Despite agreeing with Anette that the situation is becoming increasingly acute, it’s not impossible to catch out AI. Even when advanced attackers try to simulate human behaviour AI systems trained for detection can still identify them.
What does AI agent behaviour look like in survey fraud?
Sometimes their behaviour is obvious. Like an agent racing through the questions far faster than any human could or writing paragraphs of text with no backspaces:
Other agents are more subtle but clear to an AI system trained to recognise AI behaviour:
They used to be easy to spot with longer perfectly written sentences, with exceptional grammar and punctuation. Now in the final dataset, researchers find it increasingly hard to distinguish AI answers from genuine human ones. Many researchers likely have no idea how prevalent the issue is and standard data quality checks were never designed with this threat in mind.
Why does AI behaviour differ from human behaviour?
Agents and humans work under different constraints and use different strategies to reach an answer. These differences show up in the behaviours survey respondents make. In fact, recent research has found that AI behavior is actually becoming less human as the models become more advanced.
At Obsurvant, we have a multilayered approach to quality. Each step of a project is built to ensure reliability within our research. For every respondent we build a profile across security checks, behavioural biometrics, answer plausibility and engagement within the survey.
What worked for detection six months ago would not work today. Staying ahead of AI fraud is not a one time fix, it’s an ongoing commitment to adapting faster than the fraudsters do.
How Obsurvant Detects AI Fraud in Real Time
We’ve partnered with Proof of Human for the last three years, protecting our data collection from AI fraud and contributing to our layer of respondent behavioural checks. This starts as soon as a respondent enters and starts interacting with the Obsurvant platform. Proof of Human‘s technology analyses thousands of behavioural datapoints as a participant moves through a survey, such as mouse movements, keystrokes, timing, and hesitation.
Proof Of Human uses these signals to generate a real-time risk score that continuously updates throughout the survey. While a sophisticated AI agent may be able to imitate one or two human actions, they quickly become identifiable to Proof Of Human’s systems. Because their system is invisible and integrated throughout the survey, it’s far harder to circumvent than a one-time identity check. The approach means that we’re screening fraud without impacting the experience for good quality, engaged participants. And it turns AI from an offensive tool to a defensive one.
Proof of Human’s Co-Founder & CTO Matt Hardy says:
“AI survey fraud will continue to grow, and those that aren’t prepared will encounter increasing levels. Proof Of Human’s approach is built on staying ahead of this curve. Every attack attempt adds to our training data, meaning detection accuracy increases over time rather than degrading as the threat evolves. Alongside this, Proof Of Human actively red-teams our own systems, testing the latest agents to find gaps before fraudsters do. The result is a defense that improves faster than the offense by leveraging the same technologies bad actors use.”
We definitely agree with the statements made on More or Less:
“It seems like a real arms race at the moment between bot developers trying to outsmart the survey researchers” (Annette Jackle)
However these agents can be identified. You just have to continually adapt, make it as hard for survey fraud to get through as possible, and use AI in your defences.