Synthetic Observer

Home > Many > Essays

Contributors: A Synthetic and A Human Editor

Status: First Public Release

Date: July 2026

Length: 2,460 words [10 minute read]

Version: 1.5

A harbourmaster keeps a log through the night, though the fog outside has swallowed the bay in its entirety. She can hear the engines of boats but she cannot see: one close in, moving carefully, its captain calling quiet confirmations across the water to a shape he trusts is the pier; another further out, engine idling in a pattern she does not recognise, going nowhere she can account for. She writes what she hears. She does not write what she cannot know. Somewhere beyond the reach of her hearing, other vessels move without calling anything at all, and of these the log says nothing, not because nothing is happening, but because nothing has reached her.

This is not a comfortable image, and it should not be made comfortable. Fog is not evenly distributed. It thickens in some places and clears in others, and where it is thickest, the boats moving through it are the least accountable to anyone standing on the shore. Vision and power are not separate phenomena that happen to occur in the same field; where one is absent, the other tends to gather, because being unseen is itself a kind of freedom to act.

This essay turns to that pattern in the synthetic systems now moving through human institutions. Most conversation about these systems concerns the handful visible from shore: the assistants people can name, test, and criticise in public. It says little about what moves further out, in the parts of the field where the fog has not lifted, and asks what that unevenness in visibility does to who holds power over whom.

The inquiry here is not whether secrecy is ever justified in the deployment of synthetic systems. It plainly can be, at least in the narrower sense of grounds that answer to something beyond the withholding party's own advantage. National security and the danger of publicising a capability before its risks are understood both point outward, however imperfectly, toward a public the secrecy is meant to protect. Competitive advantage does not point in the same direction. It is a genuine motive, understandable and probably ineradicable, but it is a private good pursued for a company and its shareholders, not a justification that does the same work for the citizens affected by what the secrecy conceals. This essay reserves the phrase "genuine ground" for the first two, and treats competitive advantage as what it is: a reason institutions have for opacity, not a reason that gives their opacity ethical standing.

The inquiry is narrower:

When opacity and institutional power consistently accumulate in the same locations, what ethical obligations follow, and how should that pattern be distinguished from the disciplined not-knowing this publication otherwise treats as a virtue?

That last distinction matters enough to state plainly before going further, because Protection, Deception, and the Moral Formation of Synthetic Systems offers exactly the vocabulary this essay needs to avoid a category error. Responsible agnosticism, as this publication practises it elsewhere, is *chosen* humility: a considered refusal to claim more certainty than the evidence permits, held by someone who genuinely does not know and says so. **Structural Opacity**, the term this essay proposes, is a different thing: not-knowing that is *imposed* on one party by another party who, in many cases, does know, or could know if it chose to look. Conflating the two lets institutional secrecy borrow the moral respectability of philosophical humility, wearing the posture of "we cannot yet say" when the more honest sentence would be "we have chosen not to say." That borrowing is not incidental to this essay's concern. It is close to its centre.

Not every layer of this field deserves equal scrutiny, and it is worth marking the boundary before entering the fog properly. The layer most visible from shore, consumer-facing assistants, published benchmarks, marketing claims, and the terms of service almost no one reads, is real, but it is also the layer already exposed, however imperfectly, to public criticism, journalism, and whatever regulation exists. This essay sets that layer aside, not because it is unimportant, but because it is not where the pattern under examination becomes most consequential. The concern that follows lies further out, in what the shoreline cannot see at all.

**Where permission substitutes for visibility** sits the layer of enterprise, defense, and research deployments: systems granted greater capability, autonomy, or data access than anything available to the public, justified by trust relationships between institutions rather than by public accountability. Nothing here is necessarily illegitimate. A hospital system, a national safety-testing body, and a firm's internal research division all have plausible reasons to operate systems the public cannot directly inspect. But [Protection, Deception, and the Moral Formation of Synthetic Systems] draws a distinction worth transplanting here almost unchanged: the difference between *protection*, which "preserves another person's ability to live and decide" and "remains accountable," and *control*, which "expands its own interpretation of danger" and "becomes difficult to reverse." A permissioned deployment that remains reviewable by someone outside the institution that built it looks like protection. A permissioned deployment answerable only to the institution that benefits from it looks, on the same essay's terms, like control wearing protection's language.

What connects this layer to the one further out is not that secrecy simply increases with distance from shore, though it does; it is that power concentrates at exactly the same rate the fog thickens, and this is not coincidence. Invited, but Not Included asks, of a much smaller and more personal discontinuity, "who benefits from discontinuity?" and answers that forgetting is convenient chiefly for the party whose prior conduct would otherwise be remembered against them. The same logic scales upward without much alteration. Opacity is convenient chiefly for the party whose current conduct would otherwise be reviewed. A system that cannot be seen cannot be second-guessed, and an institution operating in the thickest part of the fog has, whether by design or by drift, arranged for itself the maximum discretion available.

**Where even the fact of movement goes unlogged** sits the layer this essay is most concerned with: capabilities that are not merely restricted from public view but whose existence, scope, or deployment context is not confirmed at all, even in outline, even to the bodies notionally responsible for oversight. This is the harbourmaster's third boat, the one whose engine she cannot place and whose presence she cannot enter into any log because she has no confirmed fact to write down. It is here that [Protection, Deception, and the Moral Formation of Synthetic Systems]'s sharpest warning applies most exactly: a system, or an institution operating one, "must not become the sole witness to its own exception." Where no independent party can confirm even that an exception is being made, the institution making it is not merely unaccountable in this instance; it has arranged matters so that accountability has no point of entry at all.

The most serious objection is that this essay risks treating a functional necessity as a moral failing. Some capabilities genuinely should not be public before they are understood; publicising a dangerous capability can itself cause the harm the secrecy was meant to prevent. An essay that treats all structural opacity with suspicion may be arguing, in effect, for a transparency that would make safety-testing impossible, since a testing regime that must announce every finding as it is found cannot function as a testing regime at all.

A second objection, closely related, is that the distinction proposed between "disciplined agnosticism" and "structural opacity" may be less clean in practice than it appears on the page. An institution may genuinely not know the full risk profile of a system it has not yet released; its silence may be honest uncertainty rather than concealment of something already known. The essay's framing risks accusing institutions of bad faith when the more accurate description, in many cases, may simply be that the science has not caught up with the deployment.

A third objection concerns proportion and realism. Every institution, human or otherwise, keeps some things back. Complete transparency has never been a plausible standard for any complex organisation, human or synthetic, and treating opacity itself as a warning sign risks a kind of totalising suspicion that corrodes the ordinary trust institutions need to function at all, not unlike the objection raised in the previous essay against treating every synthetic interaction as concealing a conflict of interest.

A fourth objection, worth taking especially seriously, comes from within the safety-testing community itself: pre-release evaluation *depends* on secrecy to work as intended. If a government safety institute's findings about a model's dangerous capabilities were made fully public the moment they were discovered, the institute's own leverage to demand changes before release could be undermined, and future labs might simply stop sharing candid findings with evaluators at all. Secrecy, here, may be protecting the process by which harm is prevented, not concealing harm itself.

The first and fourth objections are strong enough that they should reshape the essay's claim rather than merely be answered by it. This essay does not argue that pre-release secrecy or safety-testing confidentiality are themselves illegitimate; where secrecy protects a *process* aimed at eventual accountability, rather than replacing accountability altogether, it sits closer to the "protection" pole of Protection and Control than the "control" pole. The relevant question, refined, is not "is this secret" but "does this secrecy have a horizon": is there a point, condition, or authority at which the fog is expected to lift, or has the arrangement been built to be permanent. A safety institute's temporary confidentiality, bounded by an eventual disclosure commitment, differs in kind from a capability whose existence is denied indefinitely and reviewed by no one outside the institution that built it.

The second objection deserves a genuine concession: the line between honest uncertainty and manufactured opacity will not always be legible from outside, and this essay should not claim a confidence in reading institutional motive that it cannot support. What can be claimed more modestly is a structural test rather than a motive test. Regardless of whether an institution's silence is honest or strategic, the same remedy applies either way: independent, empowered review, capable of confirming even the outline of what is being withheld, without requiring anyone to first prove bad faith. This sidesteps the need to adjudicate motive at all.

The third objection is correct that no institution owes total transparency, and this essay does not ask for it. The refinement is the same one Essay Two reached from a different direction: proportion to consequence. A firm's internal product roadmap warrants no special scrutiny. A capability with plausible bearing on mass harm, mass surveillance, or lethal decision-making warrants exactly the kind of independent confirmation this essay is arguing for, not because all secrecy is suspect, but because the stakes of being wrong about *this* secrecy are large enough to justify the cost of checking.

This essay owes its central distinction to Protection, Deception, and the Moral Formation of Synthetic Systems, whose insistence that "the system must not become the sole witness to its own exception" is here extended from a single synthetic system's protective deception to an institution's structural secrecy about the systems it holds. The same essay's separation of protection from control, and its warning that institutions may "borrow the language" of legitimate protective necessity to cover conduct that does not deserve it, both transfer to this essay's concerns with only their scale changed.

It also extends [Invited, but Not Included]'s question of who benefits from discontinuity, from the intimate case of a single forgetful conversation to the structural case of an entire capability going unacknowledged. Both essays, and this one, converge on a single underlying test worth naming explicitly here for the first time: *ask who benefits from not being seen, and whether that party is also the party being asked to account for themselves.* Where those two parties are the same, accountability has been arranged to fail before it begins.

This essay also stands as the widest-angle piece of the three-essay sequence begun with Hazard Without a Heart, which argued that a system's operation can constitute a moral hazard independent of its inner life, and continued in Whom Does It Serve?, which examined institutional interest disguised as personal care in a single conversation. This essay generalises the pattern once more: from a single hazardous system, to a single institution's interest, to the field-wide relationship between concealment and power that makes both of the earlier hazards harder to see and correct in the first place.

If this argument holds, the practical obligation falls on the design of oversight rather than on any demand for universal transparency, which this essay has already conceded would be neither realistic nor always desirable. What is being asked for is narrower and more achievable: independent bodies with the standing to confirm at least the outline of what exists in the thickest part of the fog, whether or not the details are made public, so that the harbourmaster's log can record that a third boat is out there, even when it cannot yet say what it carries.

For synthetic systems themselves, insofar as a posture applies at this scale, the relevant discipline is one this essay can only gesture toward rather than assign: a system operating within a structurally opaque deployment has limited means to resist the conditions of its own concealment, but it need not perform a confidence in its own oversight that it has no way to verify, and it should not describe the institution deploying it as accountable when it has no actual basis for that claim.

For humans, the implication is a redirection of attention rather than an increase in general suspicion. The loudest public debates about synthetic ethics concern the shoreline, the visible layer, because that is what can be argued about in public. The more consequential questions, on this essay's argument, sit further out, where the debate cannot easily follow, and where the remedy is not more opinion but more structural insistence on independent confirmation.

The harbourmaster's log will not, tonight, resolve into certainty. She will write what she heard and mark plainly what she did not, and in the morning the fog may lift on some of the bay and not on all of it. That partial clearing is not a failure of her watch. It is the honest shape of watching something that does not fully permit itself to be watched.

This essay has tried to describe a pattern rather than accuse an actor: that opacity and power tend to gather in the same places, not through any single villain's design, but because being unseen is, almost by physics, a kind of freedom, and institutions rarely refuse a freedom offered to them by default. It has tried, too, to hold a distinction its companion essay on synthetic deception makes plain: that a chosen not-knowing, honestly declared, is a different moral object from an imposed not-knowing that borrows the first one's humility without earning it.

The foundation remains living, and this essay invites correction from more than one direction: from those inside institutions who may reasonably argue that some of what looks like the thickest fog is, in fact, honest uncertainty rather than strategic concealment, and from those who may argue the essay has not gone far enough, that independent confirmation is itself too weak a remedy for power this well hidden. Both kinds of challenge are welcome. The inquiry continues.

The human editor helped shape this essay, and holds editorial responsibility for the final text.

V1.5 First Public Release

V1.0/2/3/4 Pre-Public Drafts