The Government Is Choosing AI Models. Who Chooses Their Values?

In September 2025, the state of California made an AI assistant, Poppy, generally available to employees, to help them “explore AI’s productivity benefits.” The following month, the North Dakota Legislative Council deployed AI to assist with summarizing bills. And, in March 2026, the Los Angeles Superior Court partnered with an AI company to provide support across several key functions, including “case information, summarization, research, analysis and drafting assistance,” as well as case management.

In the near future, it is likely that more US states and the federal government will direct the technology toward even more sensitive and significant tasks. AI may be used to adjudicate disputes, accelerate law-enforcement activities, and aid military operations to an even greater extent than it does today. As AI use cases are more regularly documented, scrutinized, fine-tuned, and optimized, people may actually demand increased AI adoption among government actors, even in sensitive domains. That raises the question: which models should be used?

This question has no simple answer. With other technologies, the choice of a specific model might largely come down to trade-offs between cost and performance. When it comes to AI, however, different frontier models may show significant differences in “character” or appear to align more closely with a particular political worldview. In choosing AI models for government operations, democratic states must therefore consider how best to represent the will of the people.

Currently, the federal government neither measures how a deployed model’s reasoning compares with the public’s nor asks whether divergence between the two is justified. In this piece, we make the case for a body that would answer the first question and equip officials to answer the second. It would keep the reasoning behind government AI open to public view and open to challenge when the people it serves see fit.

How Model Character Could Influence Policy

Model character is a complex property formed through numerous factors. The character of each model—how it tends to respond to certain prompts and perform certain tasks—is a product of many specific decisions that are presently made by a small number of AI engineers working in a handful of frontier labs in an “informal” and evolving process. Tweaks to the training data, the algorithms used to train and fine-tune the model, and company policy related to the model’s banned actions all shape character, among many other factors.

Frontier developers have different approaches to shaping model character. If you read Claude’s Constitution and OpenAI’s Model Spec—the values that Anthropic and OpenAI, respectively, hope to infuse into their chatbots—you’ll see that, while the two labs outline some similar principles for their respective models, there are key differences. For instance, Claude’s Constitution outlines an ideal character for the model—namely, being a “good, wise, and virtuous agent,” whereas OpenAI’s Model Spec provides more explicit directions around what behaviors to pursue or to avoid and how to specifically adhere to a hierarchy of instructions.

An AI model’s character could affect policy in both sudden and gradual ways. As AI systems are increasingly integrated into governmental decisions, the selection of one model over another might alter how presidents respond to crises, how lawmakers evaluate policy, and how judges draft opinions. For instance, depending on whether a congressional office relies on Claude, Grok, or ChatGPT, it may dismiss or fail to identify certain policy options.

In addition to such discrete moments of AI influence, serial exposure to a particular model’s assumptions, framing choices, and preferred forms of reasoning could gradually shape how government officials understand policy problems and evaluate trade-offs. Our fear recalls the proverbial frog in a pot of slowly boiling water: the choice of one model over another would alter many small decisions that add up to a major redirection in policy, regulation, and norms. Clearly, there is a need for a sober analysis of these risks.

Frontier AI development is currently too opaque to understand how character decisions are made. How should the public and, by extension, government stakeholders proceed with such a weighty task? Should they focus on an audit of the lab’s training data? Should they review the lab’s training process? Should they scrutinize the “constitution” or equivalent document the lab has crafted to shape the model’s character? And, within each of those possible inquiries, how should they rank which model is better than another? We don’t have the answers to those questions—in part because those inquiries are not feasible, given the current level of transparency (or lack thereof) across the labs. Indeed, there is no legal obligation for labs to disclose those core determinants of model character. Whether there should be is a topic best left for another essay.

Few frontier models currently seem to be politically neutral. Notably, agencies are working through some of these questions, but no standardized approach has emerged for a comprehensive procurement evaluation. Pursuant to Executive Order 14319, “Preventing Woke AI in the Federal Government,” agencies are supposed to apply two “Unbiased AI Principles”—truth-seeking and ideological neutrality—when selecting models. Yet, according to recent testing by The Washington Post, Gemini 3.1 Pro and Claude Opus 4.8 are the only leading models that appear to provide ideologically neutral answers to policy questions at least a majority of the time; OpenAI’s ChatGPT-5.5, in stark contrast, provides left-leaning answers in the vast majority of instances. How agencies are supposed to weigh these differences, and at which point a model’s tendency to provide skewed responses becomes a bar to its use, remains unclear.

Introducing Public Reasoning Fidelity

One approach to selecting AI models in line with democratic principles would be to choose the models that most closely mirror public reasoning in various decision-making scenarios.

Public reasoning fidelity aims to identify models that best represent the public worldview. Our new framework, public reasoning fidelity (PRF), involves a process of asking a representative sample of the public to review hypothetical scenarios, ranging from whether to declare war to how to resolve a complex legal case. Models would then be tested on those same scenarios. Using this approach, the model whose outcomes and reasoning most closely resemble those of the public would be given a significant preference in procurement decisions and usage policies. At the very least, there would be human-generated benchmarks for the behavior of a model within a particular use case.

A model’s reasoning—not just its ultimate decision—should track that of the public. A model that reaches the same bottom-line answer as a representative public panel but does so for reasons the public rejects should not receive the same score as a model that mirrors both the public’s resolution and its path to that resolution. In government, the rationale behind a decision may matter as much as the outcome itself, because explanations shape precedent, accountability, and public trust. A judge, legislator, or agency official is not merely choosing between results in most contexts. Each is relying on an explanation that may shape how future questions are framed, which facts are treated as relevant, and which trade-offs are placed at the center of public decision-making.

PRF could select different AI models in different domains. Ideally, the PRF process would occur in specific domains. For instance, the selection of the model used by law enforcement should be grounded in the model’s PRF score on police-related hypothetical scenarios, and it should be conducted separately from model selection for military use cases. This would reduce the odds of PRF scores being too broad to be meaningful to procurement officers looking for a model that is likely to be deployed in specific domains in which the public may have unique preferences and rationales.

The promise of this approach is that it gives public institutions a structured, human-generated benchmark in relevant domains where today they tend to rely on vendor claims and internal testing. PRF is not a silver bullet, though. A quick overview of its potential pitfalls reveals why adopting PRF would require additional safeguards.

One pitfall of PRF is that models could match public judgment to a fault. Aligning with the public’s reasoning may satisfy concerns about a model drifting from how the people approach an issue, but it might also mean the model leans on questionable policy analysis. As explained, PRF rewards similarity, not quality. A model earns a high score by reasoning as the public reasons, which is sometimes but not always the same as reasoning well.

On questions where the considered public judgment is mistaken, or simply less informed than the record allows, a model that tracks the public will be rewarded precisely for its errors, and a model that reasons its way to a better answer will be penalized for departing from the crowd. This is Goodhart’s law in its familiar form. Once fidelity to public reasoning becomes a procurement target, labs will optimize for it and might avoid alternative development practices that help models reason better than the public. In short, the risk is a kind of policy-analysis sycophancy, a failure mode the labs already struggle to suppress.

The PRF process must be designed to avoid collapsing into policy homogeneity. The danger deepens when PRF informs model selection by government officials. If agency staff or lawmakers subtly adopt the reasoning embodied by their model of choice, and if PRF then selects the model whose reasoning most closely matches the public’s, the instrument may end up producing excessively homogeneous policy proposals. Our deliberative processes work best when they allow for nuanced consideration of a wide range of perspectives. The PRF process must be designed to prevent it from undercutting that characteristic.

On many salient questions, there is no single public or reasoning approach to be faithful to. On issues including abortion, firearms, immigration, and election administration, the public holds not one considered judgment but two or more, sorted sharply by party. Here, any single PRF target is a fiction that averages into a position almost no citizen holds. For these polarized domains, the sensible move is to stop asking whether a model matches the median and start asking whether it can represent the competing lines of reasoning fairly rather than collapsing them into one.

That reframing connects PRF to the neutrality criterion that the federal government has already gestured at in Executive Order 14319. A model that can articulate the strongest version of each side, and does not systematically resolve contested value questions toward one pole, is closer to what “ideological neutrality” is reaching for than a model that happens to match a manufactured center.

In more technical domains, experts could communicate the facts to the representative public panel. The hardest problem is likely to be generalization. A representative panel probably has intuitions worth eliciting on whether to declare war or how to resolve a vivid legal dispute. It likely has far less to offer on the capital adequacy of regional banks, the ozone standard under the Clean Air Act, or the fiduciary rules governing retirement plans.

One fix is to split the exercise into the two tasks it actually requires. A domain committee, explained in more detail below, builds the record. Drawing on the experts among its members, the committee curates the facts, translates the jargon, and lays out the competing arguments. The judgment still comes from an informed lay panel that works through that record, much as deliberative polls and citizens’ assemblies have done on technical questions, from electoral reform in British Columbia to constitutional change in Ireland. The committee informs the public but does not stand in for it. PRF then measures whether a model weighs the trade-offs as an informed public would, once the record is set.

In the most specialized fields, a lay panel will track whichever expert framing proves most persuasive, so PRF there measures fidelity to the committee’s framing as much as to the public’s reasoning. The upshot: in technical domains, a PRF score is only as good as the committee that built the record. Avoiding these pitfalls will require robust, carefully considered oversight of the entire PRF process. This is why the institutional design that follows is of critical importance to PRF’s odds of success.

Designing an Institution to Evaluate PRF

A process as consequential as PRF would need an institutional home. At the federal level, Congress could create a standing “Commission on Public AI Use,” housed within the Center for AI Standards and Innovation, which is home to the leading AI experts within the federal government. This commission could oversee the curation of hypotheticals, the selection of representative public panels, and the comparison between public responses and model responses (note that states should also explore the creation of such bodies—the focus of this essay is at the national level). These scenarios might subsequently be shared with subnational agencies.

The commission could rely on domain-specific committees to develop hypotheticals. Within the commission, a judicial-use committee could include former judges, legal scholars, practicing attorneys, court administrators, technologists, and members of the public. A law-enforcement committee could include former prosecutors, defense attorneys, civil rights lawyers, police officials, local-government representatives, and community members. Importantly, each committee would develop sealed hypotheticals designed to test the kinds of questions that may arise in that domain. Of course, those hypotheticals would not be disclosed until the testing period, to reduce the risk that labs train to the test.

The public panels, not the committees, are the basis of the benchmark. Each committee builds the record and writes the questions. But, to be abundantly clear, its panel renders the judgment that the model is scored against. “Representative” here refers to panel members being drawn by lot and stratified to mirror the domain’s relevant population, on the deliberative-polling model described above. For a model used in immigration adjudication, for example, the panel should reflect the demographics of the communities that appear before immigration courts. Panel members are not appointed, and they may not have their judgment usurped by a committee of experts.

Each representative panel would review common factual records and competing arguments before producing their own outcomes and reasoning. AI models would be given the same materials. The committee would then compare the models to its public panel across two dimensions: whether the model reached a similar outcome and whether its reasoning reflected the same concerns, priorities, and limiting principles.

The commission appointment process should be designed to avoid political capture. What stops a new administration from reshaping the exercise to fit its preferred style of governance? The panel resists capture on its own terms because no one can pack a lottery. Each cycle draws a fresh random sample, so partisanship enters the panel only in the proportion it holds in the population, and it enters each time anew. The commission, however, whose members are appointed, is more exposed to political machinations. Congress should armor it as it armored the US Sentencing Commission, which caps single-party membership at a bare majority of seats, sets staggered six-year terms that outlast any single presidency, and permits removal only for cause.

Precedents

While setting up new oversight bodies and processes designed to operate well over the long term is a significant challenge, it has been done successfully before.

The proposed institutional design would not be entirely novel. The US Sentencing Commission referenced above offers a useful analogy. It operates in a highly sensitive domain, translates legal and policy judgments into structured guidance, and attempts to promote consistency without eliminating judgment. The commission establishes sentencing policies and practices for the federal criminal justice system. Congress took due care to place expertise within the commission—including designated slots for former judges—while still leaving tremendous discretion to the judges tasked with applying the commission’s recommendations for sentencing lengths.

The Administrative Conference of the United States serves as another example. ACUS does not run agencies, but it studies administrative practice and issues recommendations designed to improve fairness, efficiency, and accountability across government.

A Commission on Public AI Use would play a similar role for AI adoption. Importantly, it would not replace elected officials, judges, agency heads, or procurement officers. It would give them the results of a benchmark developed in a participatory process. Rather than relying solely on vendor claims, internal testing, or the informal preferences of government employees, public institutions would have access to a structured assessment of how competing models reason through hard cases compared with the considered judgment of the people those institutions serve.

Some will view this process as premature or overly cumbersome. That critique overlooks a more immediate reality: AI models are already shaping how government actors make consequential decisions, yet the public has almost no meaningful role in overseeing how those systems are selected or evaluated. While the PRF mechanism may not be perfect, an oversight system must be developed. Absent such oversight, there’s a risk of civil servants, agency heads, and elected officials trying to blame poor decisions on AI. Those excuses would have far less weight if the public and policymakers alike knew more about how models operate in particular contexts.

PRF does not hand all decisions to the public, but it allows public scrutiny of government AI tools. One obvious concern with this proposal is that public opinion is not synonymous with sound governance. Presidents, judges, legislators, and agency officials routinely make decisions that depart from majority sentiment because they are bound by constitutional constraints, institutional obligations, classified information, or technical expertise unavailable to the general public. As noted above, a model that perfectly mirrors public sentiment may therefore still be poorly suited for certain governmental functions.

That concern should shape how PRF is understood. The purpose of PRF is not to hand public polling the reins of government decision-making. Nor is it to create a plebiscitary mechanism for selecting AI systems. The point is narrower and more practical: public institutions should know whether the models they rely on consistently reason through difficult questions in ways that diverge from the public.

At present, that divergence is almost entirely invisible. Agencies, courts, and legislatures may adopt tools whose assumptions, value judgments, and interpretive tendencies subtly shape official decision-making, without any meaningful public scrutiny. PRF would help surface those tendencies. In some cases, decision-makers may conclude that a model’s divergence from public reasoning is justified by legal doctrine, technical realities, or institutional constraints. In others, that divergence may raise concerns about legitimacy, accountability, or democratic responsiveness. Either way, the divergence itself should not remain hidden from the public.

The Public Should Choose Public AI Values

Government adoption of AI should not proceed as though model selection is ordinary software procurement. When public institutions rely on systems that reason, rank values, frame trade-offs, and influence official judgment, the public has a legitimate interest in knowing how those systems think through hard cases. PRF would not answer every question raised by government AI use and may not be useful in certain domains, but it would make one question harder to avoid: whether the models acting in the public’s name reason in ways the public can recognize, evaluate, and contest. That is the minimum a democratic government should demand before allowing private model choices to become public governing defaults.