The conditional superiority of fast silicon sampling

Imagine you are trying to predict how a crowd of people will react to a new law, a weird movie, or a spicy food trend. In the past, you had to go out, knock on doors, and ask thousands of real humans. But now, we have "Large Language Models" (LLMs). Think of these as super-smart, digital brains that have read almost everything on the internet. They are so good at mimicking human conversation that they can pretend to be people. This has led researchers to a wild idea: instead of asking real people, why not just ask the AI? This is called "silicon sampling." It's like asking a room full of digital clones to fill out a survey. But here's the big question: if we ask these digital clones to answer quickly, in a big rush, will their answers be any good? Or do we need to ask them one question at a time, slowly and carefully, to get the truth? This paper dives into that exact question, testing whether speed kills accuracy when it comes to simulating human opinions.

The researchers in this study decided to put this idea to the test using a specific group: people from Singapore. They took a real survey that had been given to over 2,000 actual Singaporeans about things like immigration and ethical rules (like whether it's okay to file taxes incorrectly or be rude). Then, they used a very advanced AI model (OpenAI's GPT-5.4) to create two different sets of "digital clones." One set was created using the "slow" method: the AI was asked one question at a time, with a fresh start for every single question, just like a human taking a test. The other set used the "fast" method: the AI was given the entire list of questions in one giant, complex prompt and asked to spit out all the answers at once.

The results were a mix of "not bad" and "definitely not perfect." First, the bad news: the digital clones weren't perfect copies of real humans. While they were pretty good at guessing the average opinion (for example, if most real people thought something was "okay," the AI also thought it was "okay"), they failed at capturing the variety of opinions. Real humans have a wide range of feelings; some people love a thing, some hate it, and some are in the middle. The AI, however, tended to flatten everything out, making everyone sound like they felt the exact same way. It also messed up the relationships between different opinions. For instance, in real life, people who hate immigration might also hate something else in a specific pattern. The AI's digital clones didn't quite get that pattern right; the "shape" of their opinions looked distorted compared to real people.

However, here comes the twist that makes the paper exciting. The researchers found that the "fast" method was actually just as good as, and sometimes even slightly better than, the "slow" method. This is surprising because everyone assumed that asking the AI a giant list of questions at once would confuse it or make it less responsive. But the data showed that the fast digital clones were just as faithful to the real data as the slow ones. In fact, the fast method was a massive time-saver. It took about 3 seconds to generate a full set of answers for one person, compared to 20 seconds for the slow method. It also used about 64% fewer computer resources (tokens) and cost half as much money.

So, what's the verdict? The paper concludes that while silicon sampling is still a method that needs to be used with "great caution" because it doesn't perfectly capture the messy, varied nature of real human thought, you don't need to be slow to get the best results you can get. If you are going to use AI to simulate human opinions, you might as well do it fast. It saves time and money without making the answers worse. But the authors warn us not to get too excited yet: these digital clones are still just exploratory tools. They are great for testing ideas or seeing how a model behaves, but they shouldn't replace real human surveys when we need to know exactly what a population thinks, especially in a diverse place like Singapore. The digital clones are getting better at the basics, but they still haven't quite learned how to be truly human.