Compare the full-AI hypothesis with bounded alternatives

In the full-AI hypothesis, citizens set a mandate; AI systems help turn it into proposed laws, carry out authorized rules, and assess evidence and legal claims in disputes. Separate oversight institutions would examine those functions and could intervene. This makes the central design issue visible: who may propose, approve, act, interpret, challenge, pause, and correct? If one system can translate a mandate and decide whether its own interpretation complies, separation of powers may be only formal. The hypothesis exposes dependencies, but the evidence does not demonstrate that AI can perform these governing functions. [1][6][2][12][13]

To make the hypothesis concrete, imagine citizens authorizing a constitutional mandate while AI institutions make final ordinary decisions within it. A legislative AI adopts binding rules after considering public submissions and publishing its interpretation of the mandate; an executive AI makes and carries out binding decisions under those rules; and a separately operated judicial AI determines disputed facts, interprets rules, and orders remedies. A person could challenge an executive decision before that judicial institution, with an appellate AI able to uphold, reverse, or return the decision. A separately authorized oversight institution could pause a compromised function; a pre-authorized recovery procedure would restore a checked version and review affected decisions. Citizens retain authority over constitutional amendment and withdrawal of delegation, while AI institutions hold final authority in ordinary cases. This is a stress-test scenario, not a demonstrated capability or selected constitutional model. Separate AI bodies could still share models, data, infrastructure, or operators and reproduce the same error; appeals among them might not provide an understandable external challenge. A human fallback would protect people during failure but qualify the claim that AI performs every function. The packet establishes neither that such a fallback is viable nor that it can safely be omitted. [1][6][2][12][14]

Trace a citizen mandate from public choice to executable rule

The mandate process begins before a vote: someone chooses which options and evidence citizens see. As a provisional arrangement, an authorized agenda-setting body would record who is affected, how priorities were elicited, which alternatives were excluded, and what conflicts remain. Affected groups, including minorities, could challenge that framing through a separate reviewer. Where interests conflict, the record would show who benefits, who bears costs, and what protected interests cannot simply be traded away under the mandate; contested priorities would return to the authorized public process rather than be settled by predicted preferences. The body’s composition, participation method, protections, and treatment of competing interests are open choices requiring public justification. [1][13][12][25][26]

There is no single established way to collect citizen preferences. Choosing among policy bundles can simplify participation, but a package need not fit every citizen’s preferences. A contrasting proposal is to let people express preferences on individual policies, potentially with software agents reducing the effort; that proposal has not been shown to produce binding decisions, to be easier for users, or to earn their trust. A further possibility is an individual, citizen-controlled AI helper that predicts or expresses its user’s preferences. This would shift the representation question from one institutional model to many personal agents: who controls each agent, how can its user correct or reject it, and what happens when prediction differs from expressed choice? The report favors none of these options on the current evidence. A bounded comparison should assess accessibility, effort, control, trust, and how each route handles citizens whose preferences do not fit the available choices. [27][28][29][30][31]

Carry the rule through executive and judicial work without merging their powers

Once a rule is approved, an executive system could act only within its recorded scope. Its inputs would be the current authorized rule and information relevant to a particular case; its output would be an action or recommendation with a record of the factors used. Affected people could challenge missing or incorrect information, an application outside the rule, or an unexplained result. A separate reviewer could pause a contested action under a defined mandate, while an authorized executive or adjudicative body would decide the dispute and direct correction or reversal. Technical operators could maintain systems but should not acquire authority to change policy through software. These are proposed boundaries, not demonstrated operating procedures. For allocation of public resources, the proposed record should link allocation criteria to the authorized rule or mandate, show the relevant inputs and distributional trade-offs, and explain how those criteria produced the allocation. Affected people could challenge an allocation outside the mandate or an incorrect application of its criteria through the same independent review route. Unresolved distributive priorities would return to the authorized public decision-maker; the system would not select those priorities merely because it can calculate an allocation. [2][49][50][51][12][1][6]

In adjudication, the framework should distinguish assistance with organizing evidence from authority to decide contested facts, interpret law, or determine remedies. A proposed case pathway would give the adjudicator the relevant rule, submissions, and evidence; produce a reasoned decision identifying disputed points and its basis; and provide parties a route to challenge evidence handling, interpretation, or procedure before an independent authority. That authority could affirm, revise, or return the matter for reconsideration, and direct a remedy within its mandate. A system that helps produce a statement or predict a preference has not thereby demonstrated the capacity to adjudicate. Who may delegate judgment, what evidence is sufficient, and where appeal lies remain open choices. [12]

Make independent oversight capable of acting—and bounded in what it can do

A tentative proposal gives two independent oversight bodies authority to pause a harmful AI executive, with separate administrators and approval credentials. This addresses an identified shared-administrator risk, but effectiveness and practical independence are unverified. Shared systems, actors, or operational dependencies might still undermine both bodies; no specific additional dependency or demonstrated failure is identified. The working design should require each body to document the basis for a pause, state its permitted scope, and provide a route for review and resumption. Do not rely on the proposal until the bodies’ ability to detect and carry out intervention despite relevant shared dependencies has been examined. [52][53][14][54][55]

A stop rule should identify the interests and rights-holders it protects, the actions it prohibits, the trigger, the evidence that may support intervention, and who decides when the trigger is disputed. One hypothetical approach is an initial protective pause on a credible indication of specified harm, followed by prompt independent review; this could reduce exposure while avoiding an unreviewable veto. It is a design option, not an established standard: the evidence does not define harm, establish an evidentiary threshold, or select the decision-maker. Constitutional placement and a proposed two-thirds threshold for removing a trigger may resist simple-majority repeal but could also make necessary revision difficult. Neither the threshold nor the protections covered is established as appropriate. [51][56][57][7]

Pause power should not automatically include power to change rules or control the executive’s software. Conversely, technical operators should not gain policy-making authority by implementing changes. A proposed oversight record would state whether a body is pausing an action, referring a rule for revision, or asking an adjudicator to resolve a dispute; each action would have its own authority and challenge route. Review of a pause could lift it, narrow it, or return the issue for authorized decision rather than permit indefinite obstruction. Triggers, time limits, records, appeals, and the treatment of changes that are both technical and policy-related remain unspecified.

Require meaningful challenge, not merely a human in the loop

The philosophical source distinguishes hiddenness about data collection and use from opacity about whether people can understand reasoning behind decisions. Its concern is that algorithmic systems may limit participation and understanding, including when a person is formally present but unable to challenge a system they cannot understand. This is an argument about legitimacy, not a measured finding about every algorithm. A practical design response is to provide affected people with relevant factors, reasons, and a way to contest them, without assuming that they must understand every technical detail. What level of explanation makes challenge effective remains unsettled. [62][34][63][64][49]

Privacy and participation are related but distinct design questions. A person may receive an explanation and an appeal route while still having little control over what personal information was collected, how it was used, or who could access it. Any civic-data test should therefore assess collection, use, retention, sharing, privacy, surveillance, and applicable legal risks separately from the adequacy of explanations. These are necessary inquiries, not evidence that a particular privacy safeguard is sufficient. Collection or sharing should not be relied on for a test until the relevant protections and legal conditions have been examined. [34][35][62]

A further conditional risk is behavioral steering: a system could influence people toward or away from actions while leaving its reasons, data practices, or promised benefits unclear. This is a warning relayed through a source, not an established finding about a particular governance tool. Testing should ask whether the system steers behavior, what benefits it claims, what data practices accompany that steering, and whether affected people can inspect and contest them. The threat assessment should also examine provenance and possible manipulation of training data and development information. Open code or data rules may be options to evaluate, but do not by themselves establish protection from manipulation or capture.

Use civic evidence for what it tested, not for a broader claim of governing competence

The Habermas Machine provides a concrete but bounded example: it privately collected participants’ views, generated a group statement, and revised it using critiques. Its training goal was to produce statements with strong group endorsement; that objective should not be confused with a finding that endorsement training caused participants’ views to change. In the reported UK study, participants preferred AI statements in a comparison with human mediators, and group agreement increased after AI-mediated discussion; an unmediated exposure comparison did not show the same increase. The human comparison did not use the same critique phase. These findings support further inquiry into AI as a discussion aid, not a conclusion that it can select policy or exercise public authority. [84][85][86][8][87][88]

The favorable findings do not settle whether the process represented all views fairly or why participants changed their positions. Overall movement favored the majority, even though minority positions sometimes gained support and revised statements gave minority views greater average weight. The researchers found no clear relationship between the tested majority-leaning or bias measures and movement toward the favored position; their reported analyses found no sign of political slant in the statements, while calling for continued checking. These results challenge particular explanations, not every possible bias or influence. External raters found no clear difference on whether statements included minority views or fairly summarized opinions. Any proposed use should therefore preserve visible dissent and provide a way to contest summaries; agreement or preference alone is not an adequate representation test. [9][42][43][44][89]

Use observable gates to decide whether to continue, revise, pause, or stop

A proportionate sequence begins by specifying the mandate, decision rights, challenge routes, and oversight dependencies for one bounded function. A nonbinding test could examine whether participants can identify consequential assumptions in a translated mandate, whether affected people can find and challenge relevant reasons, whether dissent remains visible, and whether independent reviewers can act. For civic-data uses, test privacy and access separately; for preference predictions, add group-specific and order-stability checks. These are proposed evaluation questions, not validated measures or thresholds. Broader delegation should not follow merely because a system performs well on a different task or because a test produces agreement. [13][25][40][3][52][53][34][116][107]

Before any test involving real citizens, public data, or consequential decisions, conduct a jurisdiction-specific legal and institutional feasibility review. It should establish—not assume—whether the relevant institutions have authority to run the test or delegate the function, what notice, evidence-access, appeal, privacy, and remedy requirements apply, and which independent body can review compliance. The packet includes a historical account of a legal right concerning certain solely automated decisions, with exceptions; it does not establish current law in any jurisdiction. If authority and applicable protections are not established, limit work to simulation or other non-consequential research. This review is a prerequisite to moving from the proposed design toward real-world testing, not a legal conclusion or a commitment to a jurisdiction. [50][118][34]

Where this leaves us

The ground-up exercise shows how a fully AI-operated model might be specified, but not that it is ready or legitimate. The provisional framework traces a citizen mandate through clarification, drafting, approval, bounded implementation, challenge, adjudication, remedy, and revision, with distinct oversight powers. Its value is to expose decisions that must be made publicly rather than hidden in technical implementation. AI-assisted deliberation warrants only authorized, nonbinding study as a supplement within a broader representative process, with fact-finding, moderation, and challenge routes outside the mediator and comparison with direct exchange before considering substitution. The evidence does not justify extending that capability to binding law, executive authority, or judicial judgment. [1][6][2][3][8][9][95][96][94][93]

The next choices for scrutiny are whether to clarify and approve citizen mandates through the proposed record-and-return process; which powers, if any, may be delegated in each function; who decides disputed stop triggers and appeals; and what evidence is required before a test continues, changes, pauses, or ends. Those choices must account for minority representation, meaningful explanation, privacy and data use, public capacity to participate, shared oversight dependencies, institutional capture, and recovery after compromise. A real-world test also depends on establishing authority and applicable protections in the selected jurisdiction; until that review is complete, simulation or other non-consequential research is the appropriate boundary. Human-led and mixed designs remain open alternatives, and the full-AI model remains a stress test rather than a deployment recommendation. [13][12][51]

Next steps

1. Review the proposed mandate-to-rule pathway: clarification record, draft and issues record, approval, challenge, and return of unresolved policy choices to an authorized decision-maker.

   This allocation of authority determines whether broad citizen choices can be translated without silently shifting policy decisions to a model. [13][12][25][1]

   Treat the pathway as a hypothetical working option; the authorization process and the body empowered to resolve ambiguity remain open.

2. Compare policy-specific input, bundled choices, deliberation, and individual AI helpers as alternative ways to gather or represent citizen preferences.

   These approaches distribute effort, control, and representation differently; the available proposals do not establish which works best or how any would affect binding decisions. [27][28][29][30][31][33][32]

   Assess usability, trust, data access, privacy, and citizens’ ability to correct an agent; do not treat prediction as authorization or assume bundles fit all citizens.

3. Map how rule-setting, implementation oversight, and decision review would connect across relevant local, citywide, national, or cross-border levels.

   The allocation affects authority, coordination, challenge routes, and accountability in a city- or country-scale design. [48][6][12]

   Keep the allocation open until a setting is chosen; do not assume direct participation in every function or that one arrangement transfers across jurisdictions.