Can Large Language Models in Social Science Predict Human Behaviour?

The Ultimate Life Simulator

Imagine playing a hyper-realistic life-simulation game where you test how your classmates would react to a new school uniform policy before announcing it. Researchers are trying to do exactly this with real human societies, using artificial intelligence to forecast human behaviour.

The Rise of Large Language Models in Social Science

Traditionally, social scientists spend months organising surveys to understand public opinion. This new study tested whether we can use AI to speed up this process. The researchers compiled an archive of 70 representative American survey experiments involving over 119,000 participants to see if AI could predict the outcomes.

What the Silicon Simulated

The team prompted GPT-4 to simulate how different demographic groups would respond to various situations. The study measured a strong correlation between the AI's predictions and actual human data, performing on par with crowdsourced human experts. However, the data also showed that the models consistently overestimated the size of the effects.

Why This Matters For Your Future

This technology could alter how we design public policies and social programmes. Instead of guessing which messages work, researchers might use AI to filter out weak ideas early. The study suggests several immediate applications:

• Pilot testing complex survey questions.
• Selecting the most promising interventions for real-world testing.
• Identifying past experiments that require urgent replication.

However, because the models exaggerate outcomes, they must assist rather than replace real human trials.