홍수 난 광화문서 유튜브 중계?…AI 명화 패러디로 꼬집은 ‘욕망의 한국’

It was during the survey days, twenty years ago, that the first lesson was learned: "Respondents are people." Each small change to the survey question involved days of effort and tweaks for phone response rates, which had been declining. The concept of "surveying people" was ingrained in our minds, and we lived accordingly. Several years ago, the "sentic survey" began to appear, using AI to create "virtual respondents."

As I delved into this topic, researching "sentic survey: AI survey" this year, I wanted to understand how to apply this concept to actual research. However, I could not come up with a clear answer. Even experts found it interesting, but the question arose, "Can we trust the answers given by people?" Meanwhile, this "sentic survey" was chosen to examine seniors' perceptions of AI.

It was a deep dive into a gap that couldn't be explained by quantitative surveys. The results were fascinating but not easily explainable. One question remained unanswered: why do people respond this way? It would be difficult to find out, as contacting participants repeatedly is challenging. The most crucial aspect was the identity of those 3% who answered.

Those with time, interest, willingness to communicate with strangers, and not being shy in front of unfamiliar people were the only ones likely to participate in AI-related interviews. Even those who had never used a smartphone, only used a smartphone for texting, those who struggled financially, and those without support were difficult to reach.

Those who should be reached first would be left behind due to the policy. The principle of asking people directly meant that those needing answers were not reached. This was why the "sentic survey" was chosen. The purpose was clear: to find a target for further research. The first principle was to create a virtual respondent based on the results of the initial quantitative survey.

The second principle was to have the questions created by people. Thus, the task began with classifying respondents into five types based on their AI perception and exposure. The 5 types were: △ Have used AI and expect a lot, △ Have used it but don't find it impressive, △ Are interested in trying it out, △ Are cautious about using it but want to learn, and △ Have never used it and are not interested.

Out of the 532 participants, 20 individuals matching these categories were selected. Their answers were entered into the "Claude" project and 90-110 questions were asked over 80 minutes. The guidelines were set to: "Do not change the answers from the survey. If the content is different from the survey, indicate it. Do not hide differences between answers.

Do not summarize answers." This process allowed the AI to fill in the gaps between answers. This method was used to clarify the respondent's perspective. For example, someone who dislikes being the only person talking about AI was identified. The person living alone would respond, "Because there are people, I would say that. I'm not here."

Another respondent, facing economic difficulties, said, "Emergency services are delayed because others are watching. Is it because I look old? It's the problem of dying." The AI showed a sense of logical control during the conversation. The virtual respondents were grouped, and focus group interviews (FGI) were attempted by bringing multiple people together.

However, the AI declined, stating that it was not possible to create such situations due to the lack of evidence in senior surveys. Of course, there are limitations. The answers from the AI contain biases from the texts it has learned. While we cannot blindly trust the results, they can still serve as a useful foundation for creating policy frameworks.

This was the first "sentic survey" conducted by a recognized public opinion research institute. I hope that this investigation will be given critical attention.