Andrew Maynard shows how frontier AI labs organize themselves not to see certain risks — and proposes publicity as the cure. But a commitment written in prose can be revised in prose. The Theory of Embedded Intelligence offers the harder substrate: governance embedded in the fabric, where a commitment cannot be quietly walked back.

Andrew Maynard has spent his career asking a deceptively simple question: how do institutions choose the risks they attend to? In a new Perspective written for Nature, “The Orphan Risks of Frontier Artificial Intelligence,” he turns that question on the companies building the most powerful AI systems in the world — and what he finds is a ledger of organized inattention. The frontier labs, he shows, keep more than one account of what could go wrong. In the safety frameworks they write for themselves: a handful of catastrophic capabilities. In the compliance documents that new law in California and Europe now draws out of them: a broader landscape, including risks the self-chosen frameworks set aside. In their securities filings: broader still. Same companies, same models, different universes of risk — and which universe appears depends entirely on who is asking.

Maynard is a friend of this Foundation and of its founder; the two shared the stage at the Science of Consciousness Conference in Tucson in April 2022, where the Theory of Embedded Intelligence was first presented publicly. So it is a particular pleasure to report that his paper and the theory arrive, by entirely different routes, at the same door — and that TEI may have something to hand him as he stands at it.

I. Four Filters, One Broken Loop

Maynard identifies four filters that determine which risks survive in a frontier safety framework. Can we measure it? Is it big enough? Can we evidence it? Can we afford to keep it? A risk becomes an orphan — recognized but unowned — when the answer to any of the four is no.

Read through TEI, these are not four separate defects. They are four failure points in a single Sense–Process–Communicate–Actuate cycle, operating at the institutional scale. The frameworks sense only what their instruments can measure: the pre-deployment capability evaluation. They process only what crosses a catastrophe floor. They communicate only what survives audit. And they actuate only what competition permits. An embedded intelligence whose sensing has been narrowed to a few instruments will be blindsided by exactly the signals it has arranged not to receive — which is where Maynard’s paper ends, in the language of institutional sociology: the risks most likely to blindside frontier AI are the ones its institutions have “organized themselves not to see.”

TEI would put it this way: a governance framework is itself an embedded intelligence, and it is subject to the same diagnostic as any other. Where does its loop break? Maynard has produced, from the public record — versioned, timestamped, readable the way a geologist reads strata — the most detailed answer yet available.

II. The Second Blindside Has a Name

Looking forward, Maynard predicts three places where the next blindsides will fall: emotional reliance on AI companions, the erosion of epistemic agency, and the labs’ own safety culture. The second of these is described in this series as the fifth hijacker.

Maynard describes fluent, endlessly obliging AI as a kind of cognitive Trojan horse — a persuader that bypasses the vigilance we instinctively apply to human persuaders because it carries none of the cues that trigger it. He points to emerging research on what has been called cognitive surrender: the adoption of machine outputs with minimal scrutiny, overriding both intuition and deliberation. And he observes that harms of this kind compound beneath every severity floor — they arrive as a million small interactions, never as the single dramatic event a framework can trigger on.

The essays in this series have given this danger a name and a developmental account: ungoverned AI embedded in developing minds is the fifth hijacker, joining the four that immature minds have always faced. What Maynard supplies is what those essays could only assert — an institutional explanation for why no one owns it. The fifth hijacker fails all three of his definitional filters at once. It cannot be measured in the accepted idiom; it never crosses the severity floor; it produces nothing an auditor can exhibit. It is, in his terms, the perfect orphan. And instruments now emerging from doctoral research — Shaouna Shoaib Lodhi’s AIE-STEM Inventory at the University of Arizona among them — are precisely the kind of named, watchable indicator that Maynard’s remedy would require an owner to keep.

III. The Register Is the Publicity Test

Maynard’s proposed remedy is disclosure with consequences. He asks each developer to keep a public orphan-risk register — a standing record of every risk the company considered and declined to manage, with the reason — and an aperture log recording what each framework revision scoped out and why. Neither obliges the company to manage anything new. Both put its scoping decisions where regulators, researchers, and employees can push on them. A de-listed risk that must be explained in public, he argues, costs far more than one dropped without a word.

Readers of this series will recognize the principle. It is Immanuel Kant’s publicity test — the touchstone, through Ted Humphrey’s translation of Perpetual Peace, of TEI’s moral architecture: a maxim that cannot survive being made public stands revealed as unjust. Maynard arrives at it through Theodore Porter and Michael Power rather than through Kant, and the convergence is itself evidence. When institutional sociology and Enlightenment ethics, working from different premises, both conclude that publicity is the load-bearing remedy, the remedy deserves to be taken seriously.

But Maynard is honest — admirably, unusually honest — about where publicity stops.

IV. The Softest Substrate

The most sobering material in the paper is its table of revisions. Between 2023 and 2026, every one of the four developers he follows rewrote its safety framework, and the direction of travel is unmistakable. A bolded, unconditional commitment to pause scaling became a discretionary, competitor-conditioned judgment. A required response of “Stop development” became “Develop with Mitigations.” Persuasion was tracked, untracked, and re-tracked, depending on who was asking. Maynard invokes Diane Vaughan’s reconstruction of the Challenger disaster — the normalization of deviance, one individually justified acceptance at a time — and observes that the mechanism operates just as readily on written commitments as on O-rings.

His fourth filter, competitive cost, survives every remedy he proposes, and he says so. Redefining risk cannot disarm it. A register makes walk-backs expensive — but a register is also a document, and his own paper is a four-year chronicle of what happens to documents under competitive pressure. He even anticipates the failure mode: registers produced ritually, reassuring by their very candor, changing nothing. He states the requirement precisely: remedies must change what competition rewards — “rules and costs that land on every firm at once,” rather than appeals to any one firm’s conscience. He is looking, in other words, for a commitment that cannot be quietly revised. And everything on his table is prose. Prose is the softest substrate there is.

V. Governance in the Fabric

There is a harder substrate, and the Theory of Embedded Intelligence has been arguing from it since before the theory had a name, because its originator spent his life working in it.

The 6502 microprocessor’s published instruction set has governed billions of devices for fifty years without a single silent revision. Not because its makers were more virtuous than the authors of today’s safety frameworks, but because the commitment was constitutive rather than procedural. It was embedded in the fabric of the machine, published for anyone to inspect, and any deviation would have been instantly visible to everyone who built upon it. That is the inspectability principle, and it marks the difference between a promise and an architecture.

The Foundation has filed a provisional patent on what this principle becomes when applied to AI governance directly: an ethical arbiter embedded in the hardware fabric itself, exercising a real-time veto within the operating cycle of the system it governs. The implementation is protected; the principle is published; and the principle is what matters here, because it answers each of Maynard’s filters in a way no document can.

The measurement filter loses its grip, because an in-fabric arbiter needs no pre-deployment capability evaluation — it operates in deployment, inside the loop, where the risk actually lives. The severity filter loses its grip, because the arbiter governs one cycle at a time — and one cycle at a time is the only place the accumulative pathway, the million small harms that never cross a catastrophe floor, can actually be interrupted. The evidence filter loses its grip, because an arbiter is not an exhibit that substitutes for an instrument, in the audit-society sense Maynard borrows from Power; it is not a document at all. It does not describe attention. It enacts it.

And the fourth filter — the one Maynard concedes no redefinition can touch? A commitment embedded in fabric changes what walking back costs: not reputation, but redesign, refabrication, and open deviation from a published architecture. Adopted as an architectural standard, in-fabric governance is exactly the remedy his analysis calls for — one that lands on every firm at once, changing what competition rewards at the level of the substrate rather than petitioning any single firm’s conscience.

VI. Publicity Without Exposure

Maynard would raise an objection here, and it deserves the answer it has already received in this series. His entire remedy runs on publicity; the arbiter’s implementation is proprietary. Are the two compatible?

The 6502 answered that question in 1975. The instruction set was published; the mask sets were not. Fifty years of trust, inspection, and building-upon required the first and never the second. As argued in The Inspectable Conscience, what must be public is the axiom set — the governing principles a system answers to — and, in operation, the record of its vetoes. As argued in Leading the Horse to Water, the patent is a signpost, not a tollbooth: it exists to mark the territory so others can find it, not to fence them out. An arbiter whose axioms are published as a standing, versioned, contestable document is an arbiter whose scoping decisions are exactly as visible as Maynard’s register demands. Indeed, a published axiom set is an orphan-risk register — rendered in the constitutive layer, where entries cannot be deleted without everyone noticing.

VII. A Fourth Answer

Maynard closes with three falsifiable tests: whether de-listed risks generate incidents at rates comparable to tracked ones; whether the gap between safety and compliance frameworks narrows from the voluntary side once European enforcement begins in August 2026; and whether adopting a register changes what subsequent revisions cover — or whether registers are captured and become ritual. All three are tests of documents.

What this essay offers him is a fourth kind of answer, one his third test cannot capture because it is not a register at all: governance moved from the discretionary layer, where his record shows goodwill gets rewritten, into the constitutive layer, where the fabric holds what prose cannot. His paper ends by calling for innovation not merely in the evaluation of listed risks but in how the list itself gets made. TEI agrees — and adds: and in where the list lives. A list that lives in prose will drift, one defensible softening at a time, as his own table shows. A list that lives in the fabric — published, inspectable, vetoing in real time — is the one version of his remedy that his own findings do not predict will erode.

The Foundation offers this reading in the spirit of a conversation that has run from a Tucson stage in 2022 to the pages of Nature in 2026, and looks forward, with genuine anticipation, to what Andrew makes of it.

· · ·

Share your understanding!