For seventy years, social science ran on one premise: to know what people think, you ask people. Recruit them, pay them, survey them. It was slow and costly.
In 2026, it’s quietly being replaced by something that never tires, never demands a payout, and never has to exist at all — a chatbot pretending to be a person. The strangest part: the real threat isn’t the one everyone panicked about first.
The invasion that didn’t come
The headline fear was bots flooding surveys, corrupting years of behavioral data. So when Prolific‘s researchers, led by Andrew Gordon, tested nearly 4,800 real responses across a dozen platforms for AI-generated text, the expectation was grim.
The result: AI text showed up in fewer than 1% of responses on most platforms. Mechanical Turk was the outlier, at roughly 16%. “Human data quality is the biggest problem in online research,” Gordon concluded not AI.
A separate five-year investigation agreed. The “bots” ruining crowdsourced psychology data are usually humans — fraud rings gaming panels for cash. Researchers Shalom Jaffe, Aaron Moss, and colleagues named it plainly in their paper’s title: the bots ruining social science are not bots at all.
That panic turned out to be a mirror pointed at an older problem. It also cleared the way for something bigger: researchers inviting AI in on purpose.
Hiring the impostor as a research assistant
A growing wing of social science now manufactures synthetic respondents deliberately — “silicon sampling.” Feed a model a demographic profile, ask it to answer as that person would, and you get a reply in milliseconds instead of days.
A 2026 study by Ashokkumar, Hewitt, Ghezae, and Willer found large language models could forecast experiment results with real accuracy — previewing findings before anyone funds the human trial. Political scientist James Druckman, in Nature, called it a plausible way to triage a field where good experiments are “time-consuming and costly.”
It’s a genuinely useful instrument — a flight simulator for hypotheses. The trouble starts when the simulator gets mistake for the sky, a risk not unlike the opacity research describe in why nobody really understands how models like GPT-6 Astra behave.
What Do You Lose When You Ask the Mirror What It Likes
A 2026 paper matched 277,000 synthetic responses from GPT-4, Claude, and DeepSeek against a real national arts survey. Real Americans liked about 24% of seventeen music genres offered. AI surrogates liked nearly twice that — around 46%, especially for niche genres.
Researchers called it “stylized omnivorousness”: a model trained on plausible text defaults to agreeableness. The relationships between tastes — how preferences cluster by age, class, and race — were “complete lost among silicon samples.” The model mimicked an opinion, not the social fabric behind it.
Psychologist Raluca Rilla saw the sharpest version of this when a chatbot, mistaken for a participant, answered a survey field with: “I don’t experience confusion in the same way humans do.” Unintentionally, the most honest answer a synthetic subject has ever given.
The hall of mirrors, formalized
Social science spent decades pouring human text into language models. Now it’s using those same models to simulate the humans it studies, faster and cheaper. Each pass launders a little more contradiction out of “the average person,” replacing it with the average sentence a model expects.
Nature asked whether AI will “ruin the social sciences — or revolutionize them.” So far, it’s doing both. The outcome hinges on something unglamorous: treating a silicon sample as a hypothesis sketch, never as data.
A field built to understand what makes people distinct now has a tool that’s brilliant at producing a plausible person and bad at producing a particular one. The bots didn’t storm the survey — social science opened the door and asked them to fill in for the humans who made the whole enterprise slow, expensive, and, it’s now clear, irreplaceable.
Related: How AI Is Changing Entry-Level Tech Careers in 2026
