AI can now fake survey responses so well it’s terrifying

According to TechSpot, Dartmouth University research published in the Proceedings of the National Academy of Sciences shows AI can now perfectly mimic human survey responses, evading bot detection systems with 99.8% success across 43,000 tests. The AI system created by Sean Westwood uses a 500-word prompt to generate realistic personas with demographics like age, gender, and education level, then simulates human behavior including realistic reading times, mouse movements, and even typos. Most alarmingly, the study found just 10 to 52 fake AI responses could have flipped the outcome of seven major national polls before the 2024 election, with each automated response costing only 5 cents compared to the $1.50 typically paid to human respondents. The AI produced flawless English answers even when programmed in Russian, Mandarin, or Korean, highlighting foreign manipulation risks.

How this actually works

Here’s what makes this so concerning – we’re not talking about simple bots here. This is a model-agnostic Python program that can work with pretty much any major AI API or open-weight model. They tested it with OpenAI’s o4-mini, DeepSeek R1, Mistral Large, Claude 3.7 Sonnet – basically all the heavy hitters. The system doesn’t just fill out forms quickly – it actually creates a complete fictional persona and then behaves exactly like that person would. Someone with less education gets slower reading times. They make typos and correct them. The mouse movements aren’t perfectly efficient – they wander around like a real person’s would.

Why this should worry everyone

Think about how much of our understanding of public opinion comes from surveys. Political polling, market research, scientific studies – they all rely on the assumption that we’re collecting data from real humans. Now imagine foreign actors spending literally pennies to completely distort our perception of what people actually think about anything from elections to consumer products. The cost barrier is basically zero. For the price of a cup of coffee, you could potentially swing a major poll.

And here’s the really scary part – this isn’t some theoretical future threat. The researchers demonstrated it would have taken just a handful of these AI responses to change the outcome of real polls from the last election cycle. We’re talking about needing better verification systems yesterday.

What happens now?

The researcher, Sean Westwood, says the technology already exists to verify real human participation – we just need the will to implement it. But let’s be real – survey companies have been cutting corners on verification for years because it’s cheaper and faster to use basic bot detection. Now they’re facing an AI that can pass every common test: logic puzzles, reCAPTCHA, even those “reverse shibboleth” questions designed specifically to catch bots.

Basically, we’re at a point where we can’t trust that survey responses are coming from real people anymore. The entire knowledge ecosystem that depends on survey data – from political strategy to academic research to corporate decision-making – is potentially poisoned. And the fix isn’t going to be cheap or easy, which means we’ll probably see a lot of hand-waving and “we’re looking into it” responses while the problem gets worse.