SINGAPORE: Governments and organisations have long relied on surveys and focus groups to gather and understand public sentiment. Now, artificial intelligence may be changing that – for good and bad.
Last year, researchers from Stanford University simulated the personalities of more than 1,000 individuals based on in-depth interviews, and found that these AI profiles could simulate responses to survey questions with an 85 per cent accuracy.
This is a development worth taking seriously, as it suggests that AI may become a useful addition to the policymaking and research toolkit, especially for testing new policy ideas and possible public reaction to those new ideas.
AI-generated polling also recently made headlines, albeit for the wrong reason. In March, US media company Axios ran an article referencing “findings” by AI startup Aaru on public trust in doctors and nurses. The “findings” were later found to be computer simulations. No real people were surveyed.
CNA Games
Show More Show Less
This prompted criticism about how readers could be misled and the ethics of what is known as “silicon sampling”, the use of large language models to simulate survey responses from thousands of digital "personas".
To be sure, the use of AI in understanding public sentiment is not a bad thing. In the age of information overload, there is real value in anything that helps policymakers and researchers spot concerns earlier and respond more thoughtfully.
Used well, AI can process large amounts of information quickly, detect patterns that might otherwise be missed and flag issues which deserve closer attention. They can also be useful for rapid pulse-checking when policymakers want an early sense of possible public reactions.
It may also gradually become cheaper and save time compared to traditional polling.
However, it does not tell the whole story. AI is good at producing neat answers, but public opinion is often less tidy than that.
People’s views are shaped by context, emotions and everyday experiences. Someone may support a policy but worry about how it will be carried out. They may sound supportive at first, only to raise different concerns when asked to reflect more deeply.
This is why speaking and listening to people matters, especially when the issues are complex or closely tied to everyday life.
In Singapore, public opinion is often gathered through more than one channel. For example, the government’s feedback unit REACH runs in-person and online “Listening Points” surveys to gather views on national issues. It also organises dialogues and focus groups on hot-button topics, such as cost of living and job security. The Forward SG exercise went even further, engaging more than 200,000 Singaporeans through partnerships, surveys, roadshows and digital platforms.
As traditional forms of public engagement, these exercises do more than collect opinions. They help uncover why people feel the way they do, trade-offs they are willing to accept and the concerns that may lie beneath an initial answer.
If AI were to play a role in future exercises of this kind, it might help identify patterns or recurring concerns more quickly. But a quick read of likely reactions may still miss hesitation, ambivalence, or the reasons behind a person’s unease.
That matters because policies are rarely judged only by their broad intent. They are also judged by whether people feel consulted and whether their concerns are being acknowledged.
Minister for National Development Desmond Lee attends the ministry's first Forward Singapore engagement session on Sep 25, 2022. (Photo: Ministry for National Development)
Another concern is over-reliance. As these AI tools improve, so does the temptation to rely on them more heavily. Once organisations learn that AI can simulate public responses, the temptation is predictable: fewer interviews and focus groups, as well as less time spent on the ground. The danger is that a tool meant to support public understanding slowly becomes a shortcut around it.
That would be a mistake. Some of the important things in public opinion do not always show up in the first answer or the most common response. They emerge in follow-up questions and reveal what people are reluctant to say plainly. This is where AI has its limits.
There is also a broader question of trust.
AI outputs can look confident and data-driven, even when underlying profiles are incomplete, outdated or shaped by biased input. Some voices may be overrepresented, while others are missing. Users may not always know what data went into the model, how the result was produced or how much confidence it deserves.
When AI tools are used to build “synthetic publics” – personas drawn from interviews or survey data that purport to represent what communities think – those communities may find themselves represented through averages and aggregates, rather than their own living voices.
Consider how it feels to be told: “We did not ask for your views directly, but the model predicts what you would probably say.” Many would reasonably ask whether they had really been heard at all.
Where sensitive personal data is involved, questions of consent, privacy and accountability matter too. People are more likely to accept new tools if they believe they are being used responsibly and transparently. Where public trust is weak, even a technically impressive tool may struggle to gain legitimacy.
The question here is not whether AI should be used at all in gathering public opinion. It has its merits so the key is how to use it well. Done responsibly, these tools can sharpen policymaking without weakening the human judgment that gives engagement its meaning.
That means keeping in mind a few principles. First, AI should be used to support, not replace, efforts to understand what people think and feel. Second, simulated outputs should be treated as starting points, not final answers or even public opinion itself.
If such tools are to be adopted, guardrails are needed. Human judgment should remain central, especially where the social stakes are high. Public trust should be treated as part of the design problem, not something to be dealt with after.
AI may well make our understanding of public views more responsive, wide-ranging and thoughtful. However, if AI is to help us understand society better, it must remain grounded in a simple idea: Technology can help us listen better, but it should not become a substitute for listening itself.
Shane Pereira is Research Associate and Elvin Xing is Research Fellow at the Institute of Policy Studies Social Lab.
Continue reading...
Last year, researchers from Stanford University simulated the personalities of more than 1,000 individuals based on in-depth interviews, and found that these AI profiles could simulate responses to survey questions with an 85 per cent accuracy.
This is a development worth taking seriously, as it suggests that AI may become a useful addition to the policymaking and research toolkit, especially for testing new policy ideas and possible public reaction to those new ideas.
AI-generated polling also recently made headlines, albeit for the wrong reason. In March, US media company Axios ran an article referencing “findings” by AI startup Aaru on public trust in doctors and nurses. The “findings” were later found to be computer simulations. No real people were surveyed.
CNA Games
Show More Show Less
This prompted criticism about how readers could be misled and the ethics of what is known as “silicon sampling”, the use of large language models to simulate survey responses from thousands of digital "personas".
A USEFUL TOOL
To be sure, the use of AI in understanding public sentiment is not a bad thing. In the age of information overload, there is real value in anything that helps policymakers and researchers spot concerns earlier and respond more thoughtfully.
Used well, AI can process large amounts of information quickly, detect patterns that might otherwise be missed and flag issues which deserve closer attention. They can also be useful for rapid pulse-checking when policymakers want an early sense of possible public reactions.
It may also gradually become cheaper and save time compared to traditional polling.
NOT A SUBSTITUTE FOR LISTENING
However, it does not tell the whole story. AI is good at producing neat answers, but public opinion is often less tidy than that.
People’s views are shaped by context, emotions and everyday experiences. Someone may support a policy but worry about how it will be carried out. They may sound supportive at first, only to raise different concerns when asked to reflect more deeply.
Related:
This is why speaking and listening to people matters, especially when the issues are complex or closely tied to everyday life.
In Singapore, public opinion is often gathered through more than one channel. For example, the government’s feedback unit REACH runs in-person and online “Listening Points” surveys to gather views on national issues. It also organises dialogues and focus groups on hot-button topics, such as cost of living and job security. The Forward SG exercise went even further, engaging more than 200,000 Singaporeans through partnerships, surveys, roadshows and digital platforms.
As traditional forms of public engagement, these exercises do more than collect opinions. They help uncover why people feel the way they do, trade-offs they are willing to accept and the concerns that may lie beneath an initial answer.
If AI were to play a role in future exercises of this kind, it might help identify patterns or recurring concerns more quickly. But a quick read of likely reactions may still miss hesitation, ambivalence, or the reasons behind a person’s unease.
That matters because policies are rarely judged only by their broad intent. They are also judged by whether people feel consulted and whether their concerns are being acknowledged.
Minister for National Development Desmond Lee attends the ministry's first Forward Singapore engagement session on Sep 25, 2022. (Photo: Ministry for National Development)
ISSUES IN OVER-RELIANCE AND TRUST
Another concern is over-reliance. As these AI tools improve, so does the temptation to rely on them more heavily. Once organisations learn that AI can simulate public responses, the temptation is predictable: fewer interviews and focus groups, as well as less time spent on the ground. The danger is that a tool meant to support public understanding slowly becomes a shortcut around it.
That would be a mistake. Some of the important things in public opinion do not always show up in the first answer or the most common response. They emerge in follow-up questions and reveal what people are reluctant to say plainly. This is where AI has its limits.
There is also a broader question of trust.
AI outputs can look confident and data-driven, even when underlying profiles are incomplete, outdated or shaped by biased input. Some voices may be overrepresented, while others are missing. Users may not always know what data went into the model, how the result was produced or how much confidence it deserves.
When AI tools are used to build “synthetic publics” – personas drawn from interviews or survey data that purport to represent what communities think – those communities may find themselves represented through averages and aggregates, rather than their own living voices.
Related:
Consider how it feels to be told: “We did not ask for your views directly, but the model predicts what you would probably say.” Many would reasonably ask whether they had really been heard at all.
Where sensitive personal data is involved, questions of consent, privacy and accountability matter too. People are more likely to accept new tools if they believe they are being used responsibly and transparently. Where public trust is weak, even a technically impressive tool may struggle to gain legitimacy.
USING AI RESPONSIBLY
The question here is not whether AI should be used at all in gathering public opinion. It has its merits so the key is how to use it well. Done responsibly, these tools can sharpen policymaking without weakening the human judgment that gives engagement its meaning.
That means keeping in mind a few principles. First, AI should be used to support, not replace, efforts to understand what people think and feel. Second, simulated outputs should be treated as starting points, not final answers or even public opinion itself.
If such tools are to be adopted, guardrails are needed. Human judgment should remain central, especially where the social stakes are high. Public trust should be treated as part of the design problem, not something to be dealt with after.
AI may well make our understanding of public views more responsive, wide-ranging and thoughtful. However, if AI is to help us understand society better, it must remain grounded in a simple idea: Technology can help us listen better, but it should not become a substitute for listening itself.
Shane Pereira is Research Associate and Elvin Xing is Research Fellow at the Institute of Policy Studies Social Lab.
Continue reading...