Virtual Leader

Status of this page. This page describes a research programme. It reports no experimental results and announces no product. Statements marked our analysis, our proposal or our hypothesis are ours and could be wrong; statements about published work carry a numbered source, and the Limitations section says how those sources were read.

What the project asks

The project started from a provocation: can a political leader exist without a human leader? Software has no bank account, no family and no re-election campaign, and it can show its working. That makes the idea attractive, which is a reason to test it.

The programme now works on a narrower question: which parts of political leadership can be handed to a machine without also handing it the authorship of public values or the authority to bind people?

The name and its working names

"Virtual Leader" comes from the original concept note: a disclosed synthetic public figure with a stable character, a programme and a memory of decisions. Inside the research the idea has had several working names:

Here "Virtual Leader" names the research direction. It is not a product, a candidate or a persona.

Relation to the book

The author's book The Agentic State argues that a state which acts through AI agents needs counterweights built into the same system: more than one model, independent audit, and reasoning that is published and can be contested. This direction takes three of those ideas and asks whether they can be run from outside government, with no state power: a second, independent account of a real decision; the strongest objection recorded beside every option; and forecasts scored against what happened.

It is a separate project. The book is about how governments use AI agents and what citizens need on their side. This direction asks something the book does not: whether any part of political leadership can be separated from a human office-holder. Its conclusions are not part of the book's argument, and nothing here is offered as evidence for the book's thesis.

What the analysis separates, and what it does not

Our analysis. We treat political leadership as a bundle of three functions: authorship of the values that guide collective choices, agency in applying those values to cases, and authority to bind others. The distinction is old: Hobbes separated the actor from the author, who owns the act, and called the right to act "authority" [1]. Recent authors apply related distinctions to AI [2–7]; we claim no priority.

Sorting leadership functions by the human involvement each needs gives four groups:

• Delegable, with humans auditing: generating options, retrieving facts, forecasting, keeping a consistent record, explaining with disclosure.
• Executable by a machine only under values that people authored: setting agenda criteria, framing, prioritising, applying a declared value profile to a case.
• Requiring a party that bears consequences and is subject to the same law: coercion over persons, liability, sanction, irreversible decisions.
• Dependent on a body and a life: acts whose meaning is the leader's own exposure, such as shared mourning. They cannot be honestly simulated.

The main result is negative. On two value commitments that we declare and do not prove, the strong hypothesis fails. Reciprocity: whoever holds coercive authority should be subject to the same law and answerable for error. Non-origination: a machine and its operators should not be the source of a community's values. Readers who hold that good outcomes alone justify authority will reject the first. Related arguments appear in other work [2, 4, 6, 7], not always with the same conclusion: Abiri, for example, argues for public authorship of an AI constitution as a way to give AI systems legitimate authority [4]. Ours is a consolidation.

Two corrections follow. A machine executor is not interest-free: the interests of those who train, host, fund and instruct it can reach its outputs. And it is not value-free: which options are on the table, or how the future is discounted, are value choices whether or not anyone states them.

The ladder, and why the work stops at "shadow mode"

The programme uses a ladder of eight claims a synthetic political actor could make. It is not a roadmap; nothing implies that moving up is desirable.

Our analysis. An independent actor can defensibly work up to L3. Shadow mode means publishing non-binding analysis of real public decisions (options, evidence, the strongest counter-argument, uncertainty) and being scored afterwards against what happened. L4 and L6 are possible only through a human institution that chooses to use the tool. A mandate (L5) does not pass to software: in the offices whose rules we reviewed, eligibility turns on conditions that, on our reading, only a natural person can meet, though no text we read says "human". At L7 an earlier claim of ours was corrected: software can already issue rule-bound administrative acts as a tool of an authority where a statute permits; German law allows this only where a legal provision permits it and there is neither discretion nor a margin of assessment [8]. What stays unsupported is software being the authority.

Limits of authority

Our proposal. A system of this kind:

• holds no office, signs nothing, spends nothing and commands no one;
• applies declared values by default only where every authorised value position leads to the same option and the decision can be undone; otherwise it maps the disagreement and returns the choice to people;
• takes no part in elections or campaigns, does no persuasion, targeting or recruitment, and has no human likeness;
• is not a legal person; liability stays with named humans and institutions.

Human control

The first design put a Steward Board in charge: named, accountable people who can pause the system and cannot quietly change what it says. Later analysis found that insufficient, because control of software sits in many places: the constitution text, value weights, model choice, tests, signing keys, funding, and the model provider's usage rules.

Our proposal is to split those controls among separate roles (authors, operators, key custodians, auditors, outside witnesses) so that no single group can change behaviour silently, with changes delayed and announced and pausing kept cheap.

Our hypothesis, untested: the more faithfully a system applies whatever specification it is given, the more its safety depends on who can change that specification. We found no working precedent for such custody over a political agent, and the design depends on it.

How evaluation is intended to work

No results are reported here. The programme fixed in advance what would count against it:

1. If the executing model, and not the published constitution, drives the recommendations across several independently developed models, the approach fails.
2. If outputs do not move as the published rule predicts when declared weights are changed, it fails.
3. If almost no real decisions turn out to be robust to value disagreement, the machine's role shrinks to mapping.
4. If trained human staff applying the same public rules do as well, the office should be staffed by humans. This is the most important test, and it has not been run.

In shadow mode only one kind of output yields evidence of skill: forecasts of observable outcomes, fixed in advance and scored against named baselines. A claim that an alternative policy "would have worked" cannot be scored, because the alternative is never observed. Results, including null results, will be published only with the full method.

Conclusions and practical implications

• The useful question is which functions are separated and who authors the values the machine applies.
• Anyone assessing an "AI minister" or "AI candidate" can ask three things: who authored its values, who can change it, and who answers for its acts.
• We treat transparency as a means of contesting decisions and do not claim that it creates legitimacy.
• The closest existing institutions are human ones under a published mandate, such as a central bank that is given its target by government and must explain publicly when it misses by more than a set margin [9]. Our pattern is assembled from such parts.

Limitations and open questions

• The two commitments behind the negative result are declared, not demonstrated.
• Whether a machine adds anything over human staff is unknown.
• Whether custody can be made to work is unknown.
• Who authors values for people who cannot take part (future generations, non-citizens) is unsettled.
• How reversibility should be measured, and by whom, is unsettled.
• Our survey of real cases is partial and the facts move, so this page describes no country cases.
• Parts of the adjacent literature are unread; our framing may be anticipated further than we know.
• How the sources were read: Hobbes, the German statute and the Bank of England page [1, 8, 9] were read in full on 5 October 2026. The academic works [2–7] are cited only for the general point named beside them; some were read in full and others at abstract or introduction level, and none of their authors has seen this page.

Actual stage

Research and design. No system has been built or deployed, and nothing has been tested with the public. A working paper exists as a restricted draft and is not released: it has had no external expert review, and several facts need re-checking on release day. One small technical test has been run. It used synthetic problems and stand-in value profiles, on models from a single developer (OpenAI), and its numbers are not reported here because one model family cannot show what the test is meant to show. Studies with human participants have not started and need an academic partner and ethics approval.

Disclosure

AI models (Anthropic's Claude) assisted with the research, the internal audits and the drafting of this page. The programme analyses documents published by AI developers, including Anthropic's constitution for its models and OpenAI's Model Spec, and treats them as company self-description. This is a conflict of interest. No reviewer independent of those developers has yet re-read the material.

Sources

1. Thomas Hobbes, Leviathan (1651), ch. XVI, "Of Persons, Authors, and Things Personated". Project Gutenberg text: https://www.gutenberg.org/cache/epub/3207/pg3207.txt
2. Seth Lazar, "Automatic Authorities: Power and AI", arXiv:2404.05990 (2024). https://arxiv.org/abs/2404.05990
3. Seth Lazar, "Lecture I: Governing the Algorithmic City", arXiv:2410.20720 (2024). https://arxiv.org/abs/2410.20720
4. Gilad Abiri, "Public Constitutional AI", arXiv:2406.16696 (2024, revised 2025). https://arxiv.org/abs/2406.16696
5. Simona-Nicoleta Vulpe, Cosima Rughiniș, Răzvan Rughiniș and Dinu Țurcanu, "Governing AI through character: moral agency and corporate authority in Anthropic's public texts", Frontiers in Sociology 11 (2026). https://www.frontiersin.org/journals/sociology/articles/10.3389/fsoc.2026.1912778/full
6. Mohammed Salah, Fadi Abdelfattah, Aisha Al Araimi and Alhamzah Al Sayed Noor, "Before agentic AI scales in government: the democratic authorization gap", Frontiers in Political Science 8 (2026). https://www.frontiersin.org/journals/political-science/articles/10.3389/fpos.2026.1956489/full
7. Aleixandre Brian Duche-Pérez, Marco Tulio Falconí Picardo, Emmanuel Neptalí Augusto Chávez Urquizo and Cintya Yadira Vera-Revilla, "From algorithmic efficiency to democratic legitimacy: rethinking AI governance in the Global South", Frontiers in Political Science 8 (2026). https://www.frontiersin.org/journals/political-science/articles/10.3389/fpos.2026.1875291/full
8. Verwaltungsverfahrensgesetz (Germany), § 35a, "Vollständig automatisierter Erlass eines Verwaltungsaktes". https://www.gesetze-im-internet.de/vwvfg/__35a.html
9. Bank of England, "Inflation and the 2% target". https://www.bankofengland.co.uk/monetary-policy/inflation