The Quiet Rise of Synthetic Public Opinion in Government

Governments and public institutions are increasingly turning to AI anticipate how communities might respond to policy decisions.
From modeling how “rural voters” could react to climate legislation to predicting neighborhood responses to zoning reforms, AI systems are being positioned as stand-ins for public opinion. As these tools gain traction, a deeper question is coming into focus: who, exactly, do these systems represent?
The challenge, as discussed in an article at the Burnes Center for Social Change, is not whether AI can generate plausible answers, but whether it can faithfully reflect the diversity and distribution of real human views. As one concern frames it: “How do we know when a model is genuinely representing a population, rather than producing a fluent stereotype?”
A new framework
The Collective Intelligence Project (CIP) is attempting to address this issue through a developing methodology known as the Digital Twin Evaluation Framework, or DTEF. Led by researcher Evan Hadfield, the project borrows the concept of “digital twins” from engineering, where virtual replicas are used to model physical systems. In this context, however, the term refers to simulated versions of public attitudes, sometimes called “silicon samples.”
Rather than releasing a single technical paper, CIP has begun sharing the framework through a series of public posts outlining how such systems might be evaluated. The goal is not to build better opinion simulators per se, but to create a way to test whether AI models can accurately mirror the spread of opinions within a group, including minority and dissenting views.
At the heart of the DTEF is a shift away from asking whether a model produces an average or “typical” answer. Instead, it asks whether a model can capture the full distribution of opinions held by a demographic group.
How the evaluation works
The framework draws on CIP’s Global Dialogues dataset, which includes surveys and deliberative discussions about public attitudes toward AI. In a typical evaluation scenario, a real group of people responds to a hypothetical policy question. The AI model is then provided with demographic information about that group, along with examples of how they answered previous questions.
The model is asked to predict how the group will respond to the new question, not as a single answer but as a distribution of opinions. That prediction is then compared with the actual distribution of human responses.
“The DTEF tests whether a model can mirror real opinion patterns rather than rely on learned assumptions.”
The resulting performance scores are intended to show where a model performs well, where it breaks down, and which populations it struggles to represent. CIP suggests these insights could eventually help policymakers, developers, and civic organizations assess when synthetic data might be reliable and when it is likely to mislead.
Notably, the framework does not attempt to answer whether synthetic data should be used in policymaking, only how accurately it reflects real populations when it is used.
Unanswered questions
The emergence of DTEF highlights unresolved governance questions that extend beyond technical accuracy. One of the most pressing is the absence of a shared standard for what counts as “representative enough.” Institutions often evaluate AI systems on technical metrics, but rarely on whether their outputs align with real-world opinion patterns across different groups.
DTEF makes these gaps more visible, but it cannot determine when a synthetic public crosses the threshold from experimental tool to legitimate input for real decisions.
Another question is when synthetic input is appropriate at all. AI-generated public opinion is frequently framed as a way to reduce the cost and time associated with public engagement. Its appeal is clear: it is fast, inexpensive, and repeatable. Yet without clear limits, these systems risk displacing genuine participation rather than supplementing it.
If evaluations reveal that models perform unevenly across populations, policymakers are left without guidance. When should a model augment engagement? What verification should be required before use? And where should synthetic publics never stand in for real people?
“Synthetic publics will not fail loudly. They will fail confidently and persuasively.”
Legitimacy beyond accuracy
Even a highly accurate model raises deeper questions about democratic legitimacy. Silicon samples are designed to predict behavior, not to ensure fairness, inclusion, or accountability. In many policy contexts, legitimacy comes not from statistical representativeness but from giving voice to those most affected by decisions.
Representativeness, in this sense, is only one dimension of democratic input, and often not the most important one.
Why this matters for governance
Digital twins of public opinion are still emerging, but they hint at a new kind of representational infrastructure. Once agencies begin using them to test policies, allocate resources, or anticipate backlash, these systems can quietly become proxies for the public.
“The risk is drift: AI systems becoming default decision-makers because they are convenient, not because they are legitimate.”
Synthetic consultation may gradually crowd out slower, messier forms of real engagement, especially under budget constraints and political pressure. Because digital twins reflect the data they are trained on, they may also amplify the perspectives of those who are already digitally visible, while excluding communities most affected by policy outcomes.
Safeguarding democratic legitimacy will require clear guardrails. Communities must be able to see, contest, and correct how they are represented. Models should be validated against real population data, not internal benchmarks. Synthetic publics must not replace statutory public input, and their outputs should be treated as signals rather than substitutes for democratic voice.
As AI-based representations of human opinion move from experiments to infrastructure, policymakers face urgent questions: Who controls synthetic publics? Who benefits from their use? And what kind of democratic future is being built in our name?
If these terms are not set now, institutions may find themselves answering to synthetic publics rather than real ones.
Governor Kathy Hochul just made New York the second state in the nation to impose comprehensive AI safety regulations.
The post The Quiet Rise of Synthetic Public Opinion in Government appeared first on eWEEK.