June 9, 2026

Public Ownership Is Not Public Trust

Assiduity AI

The first question is who shares in AI’s gains.

If artificial intelligence shifts the balance of production away from labor and toward capital, then ownership becomes a political question. Public stakes, taxation, sovereign funds, worker participation, and broader access to capital are all possible responses. They are ways of asking whether the public should share in the upside of a technology built on public institutions, public infrastructure, public knowledge, and public markets.

But ownership does not settle the next question.

A citizen can own a share of an AI company and still be harmed by an AI system that fabricates facts.

A worker can receive a public dividend and still face an inexplicable automated hiring screen. A patient can benefit indirectly from AI-driven economic growth and still be denied care by a system no one can reconstruct. A government can tax AI profits and still deploy models it cannot govern.

Public ownership may address who benefits from AI. It does not answer whether AI can be trusted in use. That distinction matters because AI policy is often forced into a single debate about safety, wealth, jobs, or national competitiveness. In practice, there are at least two separate questions.

The first is objective origin. Who defines the objective? Who owns the system? Who benefits from the output? Whose interests are treated as the purpose, and whose interests are treated as constraints, costs, or externalities?

That was the question behind the public-ownership debate. If the productive system is increasingly owned by capital, then capital will shape the objectives it pursues. Productivity, margin expansion, scale, and asset value may become the dominant goals. Those goals are not illegitimate. But if they become detached from labor, public institutions, and democratic accountability, the system faces a legitimacy problem. Objective origin is political economy.

The second question is objective pursuit. Once an objective has been declared, can the system actually hold it? That is the trust problem.

Generative AI does not simply retrieve a fixed answer from a table. It produces outputs step by step. At each step, the system may preserve the task, weaken it, or drift from it. It can remain fluent while losing fidelity. It can satisfy the shape of an answer while loosening its connection to the facts, sources, policy limits, or institutional purpose that made the answer legitimate.

The danger is therefore not only that the original objective is wrong. The danger is that the system may lose the objective while pursuing it.

Consider a benefits appeal. The model begins by accurately summarizing the applicant’s medical record and the agency’s eligibility standard. Halfway through the answer, it drops a qualifying condition, treats a discretionary factor as mandatory, and writes a conclusion that sounds authoritative but no longer reflects the governing rule. A human reviewer may see a polished final paragraph. What matters is that the system lost the thread before it got there.

That is the core lesson from autoregressive drift. A model can begin with the right instruction and still go off task. It can preserve confidence, grammar, and tone while weakening the relationship to the intended objective.

In low-risk settings, that may be tolerable. A marketing draft that drifts can be rewritten. A brainstorming answer that wanders can be ignored. A casual summary that misses nuance can be corrected.

Institutional use is different. A benefits office, bank, hospital, insurer, school, court, regulator, defense workflow, or public agency cannot rely only on the fact that an AI system was assigned a proper objective. It needs evidence that the system pursued that objective within bounds.

Objective origin is a legitimacy question. Objective pursuit is a control problem. This is where public trust begins.

We do not trust financial institutions because they promise to be careful. We require records.

We do not trust medical devices because their manufacturers say they usually work. We require testing, logs, monitoring, and reporting. We do not trust aircraft, broker-dealers, payment systems, or critical infrastructure by relying on reputation alone. We build systems of evidence around them. AI will need the same shift.

The current AI debate often treats trust as a property of the model. Is the model safe? Is it biased? Is it aligned? Is it truthful? Is it objective? Those questions matter, but they are incomplete. Institutions do not deploy models in the abstract. They deploy systems inside workflows.

A model used for brainstorming marketing copy is one thing. Another is a model used to summarize a benefits appeal, draft a compliance finding, triage a fraud alert, support a clinical decision, review a loan file, or guide an investigative workflow. The relevant question is not only whether the underlying model performs well on general benchmarks. It is whether the deployed system behaves within the boundaries of the task it was assigned. That requires runtime evidence.

A benchmark says how a model performed under test conditions. A policy says how the organization intends the system to be used. A prompt says what a user asked at a particular moment. A final answer shows what the system produced. None of those, by itself, shows whether the system held the objective during generation.

That is why traceability matters. The EU AI Act requires high-risk AI systems to support automatic record-keeping throughout the system’s lifetime and to be designed for human oversight to prevent or minimize risks to health, safety, and fundamental rights. Those requirements reflect a basic institutional reality: if AI systems affect consequential decisions, organizations need more than output. They need records of behavior in use. [1] [2]

For an agency head, that means knowing whether the system operated within the authorized use. For a procurement officer, it means comparing vendors on evidence rather than demonstrations. For an inspector general, it means reconstructing what happened after the fact. For a risk officer, it means knowing whether exceptions were routed to humans. For an affected person, it means having some basis to challenge the result.

The core questions are practical.What objective was declared? What sources, facts, or constraints were the system required to preserve? Did the system drift from the assigned objective? Were exceptions detected? Was a human asked to review the right cases? Can the institution reconstruct what happened later? Those questions are the beginning of operational trust.

They also change the buyer’s problem. A serious AI buyer should not ask only whether a model is powerful. Power is increasingly available. The harder question is whether the institution can govern that power once it is placed inside a workflow.

This is especially important for the government. Government use of AI is different from consumer use. A citizen cannot always choose another provider. A beneficiary cannot easily opt out of an agency process. A regulated firm cannot ignore a supervisory finding. A defendant, applicant, patient, student, or taxpayer may have to live with the consequences of an automated or AI-assisted decision.

That does not mean the government should avoid AI. The opposite may be true. AI may help agencies reduce backlogs, improve service, detect fraud, summarize complex records, identify inconsistencies, and make public administration more responsive. But public-sector AI adoption requires a higher standard than usefulness. It requires contestability.

Contestability means that the affected person, auditor, reviewer, or oversight body has a means to challenge the output. Not necessarily to inspect every model weight. Not necessarily to obtain every internal trade secret. But enough to know what the system was supposed to do, what information it used, what constraints applied, whether it deviated, and how a human reviewed the result. Without that, AI can become a legitimacy problem even when it improves productivity.

The United States is moving in the same evidence-oriented direction as Europe, though through a different governance model. NIST’s AI Risk Management Framework organizes AI risk around governance, mapping, measurement, and management. The Government Accountability Office’s federal AI accountability framework emphasizes governance, data, performance, and monitoring. OMB M-24-10 establishes requirements and guidance for federal agency AI governance, innovation, and risk management, especially where rights and safety are affected. [3] [4] [5]

The direction is clear. AI governance is moving from principles to evidence. But evidence has to be produced somewhere. It does not appear because an organization writes a policy. It does not appear because a vendor has not published a model card. It does not appear because a reviewer says the final answer looked reasonable. Evidence must be generated within the workflow. This is where ownership and trust separate.

Public ownership addresses distribution. It says: if AI creates wealth, the public should share in it.

Operational trust addresses accountability. It says: if AI affects rights, benefits, obligations, opportunities, or institutional decisions, the public should be able to challenge how it was used.

Both questions matter. They are not substitutes. A public stake in an AI company does not make an agency’s AI system reviewable. A sovereign wealth fund does not tell an inspector general whether a model drifted from a policy constraint. A tax on AI profits does not prove that a hospital, insurer, school, or bank used AI within authorized boundaries. The ownership layer can redistribute gains. The control layer has to govern conduct. This is where a new category of infrastructure is needed.

Not another chatbot. Not another foundation model. Not merely a content filter. Not only a compliance checklist. Institutions need runtime control infrastructure: systems that sit where AI is actually used and produce evidence of whether the AI stayed within its assigned operating bounds.

In practical terms, that means an institution should be able to define the task before generation begins. It should be able to state the facts, constraints, sources, and limits that matter. It should be able to observe whether the system remains coherent with those requirements as the output develops. It should be able to flag drift, preserve a record, and route exceptions to review.

For a procurement officer, that means the agency can compare AI systems on evidence, not just vendor assurance.

For an inspector general, it means there is a record to audit.

For a risk officer, it means exceptions can be governed rather than discovered by accident.

For a citizen or customer, it means there is at least the beginning of a trail that can be challenged.

Assiduity’s ECD sits in this layer. Equilibrium-Constrained Decoding is neither a foundation model nor a content-moderation wrapper. It is runtime control infrastructure designed to make objective pursuit observable.

The institution defines what the system is supposed to preserve. The runtime maintains a formal representation of those task requirements and measures whether generation is deviating from them as the output develops. When the output begins to drift, the system can detect the loss of coherence and preserve evidence of what happened. Afterward, the institution has more than a final answer. It has a record of pursuit and that record matters.

Trust is not created by assigning a good objective. Trust is created by showing that the system held the objective while acting.

This is not a complete solution to AI risk. It does not decide whether a company should replace workers with AI. It does not settle copyright disputes. It does not make training data representative. It does not prove universal political neutrality. It does not determine whether a use case is morally appropriate. It does not remove the need for human judgment, statutory rights, procurement discipline, cybersecurity, red-teaming, or democratic oversight.

Those limits matter. Runtime control is not a theory of justice. It is not a labor-market policy. It is not a substitute for law. It is an institutional control layer.

That narrower claim is also why it matters. AI governance will fail if every tool is presented as a universal answer. The public does not need another promise that AI will be safe just because someone says it will be. The public needs institutions that can show their work.

This is especially true as AI systems become more agentic. A simple chatbot produces text. An agentic system may plan steps, call tools, search databases, draft messages, trigger workflows, update records, or recommend actions across multiple systems. The governance problem changes because there may be no single final answer to inspect.

Imagine an AI system that reviews a benefits file, queries a database, drafts a recommendation, routes the recommendation to a caseworker, and updates an internal record. The risk is not only that the final recommendation is wrong. The risk is that one step in the chain drifted from the authorized purpose, used the wrong source, ignored an exception, or triggered the wrong downstream action.

Static review cannot keep up with that kind of dynamic behavior. A policy document cannot see tool use. A post hoc audit may not capture the moment when a system began to drift. A human approver may over-trust a fluent output if the system gives no signal that coherence has degraded. Governance has to move closer to the point of generation and action.

This is not anti-innovation. It is the condition for durable adoption. Organizations will not deploy AI deeply into serious workflows if every use depends on blind trust. Agencies will not maintain public legitimacy if they cannot explain AI-assisted decisions. Regulated firms will not satisfy auditors with screenshots and vendor slogans. The more important AI becomes, the more evidence it will need to produce.

The best version of AI policy should therefore distinguish between three layers. The first is economic legitimacy: who shares in AI’s gains? The second is operational legitimacy: can institutions prove how AI behaves in use? The third is policy legitimacy: can the government encourage adoption without creating a slow, brittle, or politicized preclearance regime?

The first question concerned ownership. The second question is the trust gap. Public ownership can legitimize the origin of AI’s objectives. Runtime evidence is what legitimates their pursuit. The country will need both.

Because the deepest public concern is not only that AI companies will become too rich. It is that AI systems will become too consequential to trust and too opaque to challenge.

Sources:

[1] EU AI Act, Article 12, requires high-risk AI systems to technically allow automatic recording of events over the system’s lifetime.

[2] EU AI Act, Article 14, requires high-risk AI systems to be designed for human oversight aimed at preventing or minimizing risks to health, safety, and fundamental rights.

[3] NIST’s AI Risk Management Framework is organized around the functions Govern, Map, Measure, and Manage.

[4] GAO’s AI Accountability Framework is organized around governance, data, performance, and monitoring.

[5] OMB Memorandum M-24-10 establishes requirements and guidance for federal agency AI governance, innovation, and risk management.