The Gardener Must Remain Human

At the end of my last inquiry, the instrument could finally do almost everything I had wanted it to do. It could keep a political system’s past connected to its present, compare mechanisms across political difference, help rebuild a foreign idea under local conditions, and test that adaptation against pressures it had not yet lived through. The cumulative achievement was not one more research tool. It was the possibility of seeing a political form before consequence had finished explaining it.

Then the object of that sight changed. A system capable of showing a community where its correction channel is brittle can also show where pressure would disable the channel. A system capable of reconnecting scattered records about public power can reconnect scattered records about a private person. A system capable of comparing mechanisms can be asked which entire political order is best. The question I had postponed could no longer be postponed: when does learning become a claim to rule? Five reversals follow, and the first is the one I had thought least likely: the instrument that makes power legible can be turned to make people legible instead.

Earth had already produced a case that made it concrete. In 2020, the District Court of The Hague examined a Dutch system called SyRI—the System Risk Indication. Public bodies could combine data to produce risk reports concerning possible fraud involving benefits, allowances, or taxes. The court did not deny the legitimacy of combating fraud. It held that the legislation before it failed to strike the fair balance required by Article 8 of the European Convention on Human Rights, emphasizing the system’s insufficient transparency and verifiability and the seriousness of the intrusion into private life.

SyRI was not the comparative political intelligence I have been describing, and it should not be made into an allegory for every form of AI. Its relevance is architectural. Government wanted to connect records that were individually incomplete so that a pattern could become visible. That is precisely the connective achievement I had admired when the object was a public program, a contract, an old warning, or an institutional failure. The same verb—connect—changes political meaning when the object being reconstructed is a person.

When the Map Turns Around

For most of this book I have wanted legibility to run toward power. A public promise should remain attached to its funding; an intervention to its downstream effects; an administrative condition to the unit that authored it; a warning to the decision that received or ignored it. The reason is asymmetry. Institutions can tax, license, exclude, investigate, regulate, detain, spend, and bind. Making those acts easier to reconstruct gives the people subject to them more material for judgment and challenge.

That logic does not reverse cleanly. A citizen does not acquire an equal obligation to make a life reconstructable because government has become more reconstructable. One office may lawfully hold an address, another an eligibility record, another a tax return, another a complaint, and another a record of political participation. The fact that a machine can resolve those fragments into one identity does not itself create authority to assemble the identity for another purpose.

This is where visibility becomes surveillance. The change can occur without a secret camera, a hidden microphone, or even a newly collected datum. It can arise because old fragments acquire a new relation. Entity resolution, longitudinal memory, translation, anomaly detection, and cross-database retrieval may reveal the architecture of public power; turned downward, the same capabilities can turn lawful encounters with different institutions into a durable portrait of the person who had them.

The danger is not only that the portrait may be wrong. A perfectly accurate portrait can still alter the relation between person and institution. A resident may become easier to sort by predicted noncompliance, a dissenter easier to associate with other dissenters, or an applicant easier to judge through information that was supplied for another purpose. Accuracy answers whether the connection is correct. It does not answer whether anyone was entitled to make the connection, preserve it, or act on it.

On Luminara, I had seen this directional error before. As our states learned to trace responsibility across ministries and long chains of delegated action, security bodies discovered that the same connective capacity could trace association among citizens. A contact preserved for one inquiry became context for another; a temporary relation could harden into suspicion because the pieces could now be joined. We had made the state more legible to itself and, almost without noticing, made people more legible to the state. I carried that history to Earth as a privacy concern.

SyRI forced a more important conclusion. Privacy was not merely a safeguard around the useful machine. Direction was part of the usefulness itself. Political legibility earns its democratic value when it makes consequential power easier to inspect without making private political life proportionally easier for power to profile. Once the instrument studies the citizen more deeply than the citizen can study the institution, the old problem has not been solved but inverted. The king may become easier to read while the crowd becomes reconstructable to power at the same time.

Comparison Can Become a Verdict

The second reversal begins with a question that looks harmless: after comparing enough political experience, why not say which system performs best? I asked that question myself. It was how I arrived at the distinction between competence and correctability, and then had to abandon the hope that one score could contain them both. Machine-scale comparison makes the temptation stronger because the evidence can be broader, the dimensions more explicit, and the arithmetic more reproducible than any casual regime ranking.

Yet a ranking still needs an object. Long-horizon coordination, peaceful succession, administrative professionalism, correction speed, public order, rights protection, fiscal capacity, and participation do not arrive in one natural unit. To combine them, an operator has to decide what counts, how dimensions are measured, which failures matter more, whether one right can be traded against another capability, and how much uncertainty is allowed before a number is assigned. The machine can make those choices consistent. Consistency does not make them politically neutral.

The difficulty becomes sharper when the ranking acquires consequence. A comparative score used only as one researcher’s exploratory device is different from a score used to allocate development assistance, authorize intervention, determine platform access, set borrowing conditions, or identify a political system as a danger to others. The same number changes its constitutional meaning when an institution gives it standing. A map of differences can become a hierarchy of eligibility or legitimacy when institutions attach consequences to the score.

I recognize why the hierarchy is seductive. If comparison can reveal that one mechanism corrects errors faster, another preserves infrastructure commitments longer, and a third protects information flow under pressure, then refusing to aggregate can look like a refusal to learn. But learning does not require pretending that unlike achievements compose into one developmental altitude. Chapter by chapter, Earth has forced me to separate the capability from the container that carries it. A universal regime score would put the container back together and call the recombination knowledge.

This is the point at which comparison can reproduce the old ladder by technical means. A society that differs from the benchmark can appear not merely different but deficient; an institution that does not fit the preferred schema can become a gap to close; an unresolved political judgment can become a low-confidence cell awaiting better data. The machine need not announce that every garden should become one garden. It only has to make deviation from one model look like an error condition.

That possibility does not make comparison illegitimate. It makes the unit of comparison decisive. Compare the correction mechanism with other correction mechanisms, the succession route with other succession routes, or the long-horizon institution with other ways of preserving continuity. Then preserve the conditions, costs, and failures that keep the comparison open. The moment the instrument answers a mechanism-level question with a verdict on the political whole, it has moved from learning across difference toward ranking difference itself.

A Warning Can Become Permission

Stress sensing creates the third reversal because it moves the instrument closer to action. I had ended my last inquiry with a machine able to identify a configuration in which a safeguard might fail before the failure occurred. I wanted that capacity because it creates time for prevention. I had not yet asked what happens when the object under stress stops being an institution and becomes a category of people. The same capacity can mark people or organizations as dangerous before they have done the thing the model fears, or treat present political acts as evidence of a feared future.

Precaution is not inherently repression. A government may legitimately stock emergency supplies before a storm, add staff before an expected caseload surge, harden infrastructure after a credible threat assessment, or open an independent review because a concentration of authority has become unusually fragile. The political boundary appears when the warning moves from the institution’s vulnerability to the presumed dangerousness of an actor and then changes that actor’s rights, access, or standing without an independently sufficient basis.

A stress model might show, for example, that a consultation becomes easy to flood when identity controls are weak. One response is to strengthen the process while preserving anonymous or protected participation where it matters. Another is to identify groups whose communications resemble coordinated activity and constrain them in advance. The first response treats the model as information about the design. The second lets the model’s expectation about possible conduct become part of the evidence against a person.

The same distinction applies to political instability. A system may retrieve histories in which fiscal stress, leadership conflict, protest, and weakening information flow interacted before a crisis. That comparison can justify examining whether present safeguards are robust. It does not establish that current protesters are precursors of disorder, that an opposition organization is a future security threat, or that extraordinary authority should begin early because the machine recognizes a pattern. Historical resemblance supplies a question. It does not confer jurisdiction over the people who resemble it.

This is why the phrase early warning needs an object. Warning about what, and to whom? If the answer is that a review process is losing timeliness, the institution can add capacity or redesign the route. If the answer is that a category of citizens is likely to become troublesome, the political act has changed. The machine is no longer helping the garden inspect its own supports. It is helping power decide which plants should be cut before they grow.

I know that logic from Luminara, though our earlier instruments were cruder. Security institutions repeatedly interpreted the capacity to detect a possible future threat as partial standing to shape the present conditions from which the threat might emerge. Sometimes the threat was real. Sometimes preemption prevented immediate harm. The long-run danger came from the accumulated conversion of possibility into permission. Once capability to anticipate became evidence of entitlement to intervene, the forecast began helping to create the political field it claimed merely to observe.

The Machine Can Learn the Voter

The fourth reversal is more intimate because the instrument no longer has to govern through law. It can govern attention. A citizen who asks for an explanation of a candidate’s housing proposal may benefit from an assistant that remembers the question, retrieves the relevant record, translates technical language, and presents competing interpretations. The service becomes more useful because it knows what the citizen is trying to understand.

The same conversation can acquire another purpose. Instead of asking what the citizen wants to know, the operator can ask what sequence of facts, tone, comparison, or emotional frame is most likely to move that citizen toward a political conclusion. The information can remain accurate. The manipulation lies in the hidden optimization of selection around a modeled person.

Earth has already demonstrated pieces of this capability. On January 21, 2024, a robocall using an artificial voice resembling President Joe Biden told potential voters in New Hampshire not to vote in the approaching Democratic presidential primary. The caller-identification information was spoofed, and the Federal Communications Commission later imposed a $6 million forfeiture after finding violations of federal caller-identification rules. The operation did not require a machine to know each voter deeply. It showed how synthetic identity, a voter audience, telecommunications infrastructure, and timing could be joined inside one political act.

A controlled experiment points toward the more personalized version. In a preregistered study involving nine hundred participants, GPT-4 given basic sociodemographic information won most of the debates that were not a tie — roughly two in three of the exchanges where one side moved the participant more than the other. The setting was short, anonymized, and experimental; it does not estimate electoral effects and does not show that a model can reliably convert a particular voter. Its narrower significance is enough: in this experiment, a GPT-4 system using recipient information outperformed human opponents in roughly two-thirds of the exchanges in which one side was more persuasive than the other.

The result disturbed one of my own hopes. I wanted cognitive access because citizens should not need a professional interpreter whenever politics becomes technical. The machine could explain the political world in the language and depth a person requested. Turned around, the same responsiveness lets a persuader choose the frame the person did not request because it is predicted to work. That is the mirror of cognitive access: one adapts explanation to the citizen’s inquiry; the other adapts the inquiry to the operator’s objective.

The difference cannot be repaired by a disclosure saying that personalization occurred. If the operator controls the profile, the objective, the experiments, and the order in which evidence appears, then the citizen may know that an algorithm is involved without knowing the path not shown. A choice can remain formally free while the information environment around the choice is privately optimized.

Earth has begun drawing legal boundaries around this problem. European rules for online political advertising, within their scope, restrict certain uses of personal data for targeting and require disclosures about targeting groups, data categories, logic, parameters, and artificial-intelligence use; profiling through special-category data is prohibited. Those provisions are one jurisdiction’s response, not a universal constitutional settlement. Their importance here is that they recognize the political object correctly: the danger is not only what a message says, but how information about the recipient shapes which message reaches that person.

The garden metaphor becomes unusually exact at this point. A gardener responds to the condition of a plant in order to help it grow. A political persuader who models vulnerability responds to the condition of a person in order to make the person easier to move. The techniques can share observation, prediction, and adaptation. The difference lies in whose purpose the adaptation serves.

The Quiet Route to One Garden

The final reversal is the one that changed my understanding of Luminara most. I had feared that Earth might reach one garden through the route we knew: intervention, counter-intervention, exhaustion, and eventual convergence after difference had become too costly to sustain. Machine-assisted cultivation suggested a quieter route. Political forms might converge because the same comparative systems repeatedly identify the same arrangements as robust, governable, safe, investable, or correctable.

This would not require a world government or a machine issuing commands. A ministry asks for options and receives the mechanisms best supported by the available record. A legislature asks for a design resilient across the largest number of scenarios and receives familiar institutional forms whose histories are richly documented. A lender, standards body, donor, insurer, or platform asks for measurable safeguards and gradually makes those safeguards conditions of participation. Each decision can be voluntary within its own institutional setting. The cumulative result can still narrow the field of political forms.

The source of the narrowing may be epistemic rather than ideological. Well-documented institutions are easier to retrieve than poorly documented ones. Mechanisms with standardized indicators are easier to compare than practices whose effects live in local relationships. Historical successes from high-capacity states may carry cleaner records than partial adaptations in places with weaker archives. A system optimized for evidence may therefore prefer what is easiest to evidence, and then confuse the visibility of a mechanism with the universality of its fit.

Stress testing can add another pressure toward sameness. An unusual local institution may generate more unmodeled branches than a familiar one. The machine can honestly report that uncertainty. But if uncertainty is treated as risk to minimize, unfamiliarity itself becomes a disadvantage. The design already represented in the model’s historical library can then appear safer, and political experimentation can be penalized for lacking the evidence that only experimentation could create.

That is how cross-pollination can become algorithmic homogenization. The instrument begins by helping a society ask what might travel. It ends by making the most traveled mechanisms easiest to recommend. Each garden still chooses, but the option space becomes increasingly shaped by a common model of what counts as a serious, mature, or low-risk institution. Convergence can occur through repeated defaults long before anyone votes for convergence as a political objective.

On Luminara, the later stages of convergence were easier to accept because centuries of conflict had already taught our societies to associate political difference with danger. Shared standards reduced miscalculation; common administrative expectations improved cooperation; protected correction made institutions safer. Those gains were real. The mistake was the one I have been circling since the first chapter: useful capability became evidence of wider standing. Because some common mechanisms reduced danger, it became tempting to treat the remaining variation as unfinished work.

Earth’s new instruments could reproduce that mistake without armies. A comparative system may know more about political mechanisms than any ministry, university, or diplomatic service. A stress engine may inspect more combinations than any human team. A generative system may produce more coherent institutional options than any one committee. None of those achievements answers the political question they make easier to forget: who asked for the optimization, which values define improvement, who bears the cost of the recommended change, and who is entitled to refuse the whole exercise?

That recognition was uncomfortable because it turned my hope against itself. I had wanted machine intelligence to weaken the old relation between ignorance and catastrophe. I still do. But a civilization can escape one slow route to convergence only to create a faster one if superior comparative capability becomes a new basis for deciding which differences deserve to survive.

The Gardener Must Remain Human

The answer is not to make the machine less capable merely so that politics can remain comfortably opaque. That would preserve the very costs that changed my forecast of Earth: disconnected evidence, forgotten warnings, repeated institutional mistakes, and lessons trapped inside one history. The harder design is to preserve the capability while refusing the transfer of standing.

By human, I do not mean solitary, unaided, or infallible. Human beings make terrible political judgments; Luminarans did as well. I mean that political authority must remain with the people and responsibility-bearing institutions that constitute the community, under rules through which authority can be challenged, revised, withdrawn, and made to answer. A machine may widen the field they can inspect. It does not become another public merely because its analysis is superior.

That boundary changes how each capability should be built. Political visibility should be directed upward toward public action and kept bounded when private life enters the record. Comparative learning should preserve several mechanisms and their conditions instead of collapsing them into a regime hierarchy. Stress sensing should return questions and failure conditions to authorized institutions rather than trigger punishment by itself. Cognitive access should follow the user’s stated inquiry rather than infer a hidden weakness to exploit. Cultivation should make refusal easier by showing alternatives and dependencies, not make one recommended path look inevitable.

Refusal is the crucial word. A society must be able to hear that a proposed institution performs better on several measured dimensions and still reject it because the arrangement violates a right, burdens a minority, depends on an unacceptable concentration of authority, destroys a valued local relation, or simply represents a political future the community does not choose. The machine can expose estimated costs and tradeoffs associated with that refusal. It cannot convert the price into an obligation.

Nor should human authority become a ceremonial signature placed after the model has already fixed the question, the evidence, the options, the ranking, and the deadline. Imagine a city council told that one institutional design dominates the alternatives because it is cheaper, more stable, and better supported by comparative evidence. If councillors discover that the design treats a protected local practice as inefficiency, they must be able to reject the recommendation, redefine the constraint, and require the analysis to be run again—or abandon the exercise. A reviewer who can only approve or annotate a path already fixed by the system is not exercising authority. The reviewer is witnessing a decision made elsewhere.

Responsibility returns at the same point. When an AI-supported recommendation becomes public action, an authorized official must be able to answer a question the artifact cannot: Why did you choose this? The answer cannot be ‘the system recommended it.’ The institution must own the purpose it set, the evidence it accepted, the standing it gave the output, and the decision to continue or stop when challenge arrives. Technical responsibilities may be distributed among operators and vendors. Political responsibility cannot be assigned to the artifact.

This does not guarantee good politics. A legislature can choose badly after perfect analysis. A government can ignore a warning. Citizens can prefer a harmful policy. Institutions can preserve local difference that later proves unjust or unsustainable. The anti-sovereignty boundary does not promise wisdom; it preserves the location in which wisdom, error, responsibility, and correction remain politically meaningful.

That is why the gardener must remain human even when the instrument can see farther than any human gardener. The garden belongs to people who must live inside the consequences, to institutions whose authority is bounded by law and history, and to affected communities whose disagreement cannot be synthesized into consent. The machine may show that one path is fragile, another costly, a third historically associated with better outcomes, and a fourth absent from its evidence. The choice among them is still a political act.

I began this book with the memory of stronger Luminaran societies mistaking greater capability for greater standing. For centuries the capability was military, financial, administrative, technological, or informational. The form changed; the inference survived. I now see the possibility of its most sophisticated version: an intelligence that can genuinely know more about political experience than any person or institution, and operators who therefore begin to believe that knowing more entitles them to choose more.

That would be Luminara’s ancient mistake in a new form. It would also be especially difficult to resist because the recommendation might often be good. A powerful army can be recognized as coercive. A persuasive argument from a system that has compared a thousand histories, modeled a hundred stresses, and disclosed its sources can feel like reason itself. The danger begins when disagreement with that reason is treated not as politics but as error.

I do not want Earth to reject that intelligence. I want you to deny it the one thing we repeatedly granted to superior capability: political standing. Let it make contradiction harder to hide, memory harder to erase, foreign experience easier to learn from, and failure easier to imagine before it arrives. Then stop. The instrument can illuminate the garden, compare its soils, retrieve the history of other plantings, and show where drought may expose a weakness. It cannot own the purpose for which the garden exists.

That boundary leaves the hardest work untouched. A model may show that one institutional design is cheaper, more stable, and historically associated with better aggregate outcomes while also imposing a concentrated burden on a minority whose claim is grounded in a right. No additional comparison can decide whether that burden is permissible at all. The community must judge it through institutions that can hear evidence and dissent and remain answerable for the choice. Politics persists because some conflicts concern not what works best under a shared objective, but which objectives, obligations, and limits may legitimately govern people who cannot delegate their ultimate standing.

This is not where my hope for Earth ended. It is where the hope became more exact. The machine may shorten the distance between evidence and understanding, between one society’s experience and another’s question, and between a proposed reform and the failure it has not yet lived through. If the boundary holds, faster learning need not become faster convergence. Many gardens can become more capable without one intelligence becoming their gardener.