On Thursday, the President postponed the executive order on AI. The argument that won the day was about capability. The argument that mattered was never made.
|
Prelude — A Note on the Frame
What was postponed, what was said, and what was missing from the argument.
The White House · Thursday, May 21, 2026
On Thursday, May 21, 2026, the President postponed signing an executive order that would have established a voluntary federal review process for powerful AI models before public release. The stated reason was competition with China. Speaking to reporters in the Oval Office, the President said the order could “get in the way” of the American lead, and that he “didn’t want to do anything that’s going to get in the way of that lead.” The order, by all accounts, was modest. It would have formalized — for a ninety-day federal review window — the kind of staged, supervised early deployment that Anthropic had already initiated voluntarily with Project Glasswing for its Mythos model, and that the Department of Commerce’s National Institute of Standards and Technology had announced earlier in the month as a national security review framework. In the public reaction that followed, the debate divided along familiar lines: accelerationists relieved that the lead was not endangered, safety advocates dismayed that oversight was deferred. Neither side, in any account this Foundation has reviewed, asked the question that the Theory of Embedded Intelligence holds to be the only structurally serious one. This essay is about that question — and about what its absence from the May 21 conversation reveals. |
I. The Question That Was Not Asked
The question is not whether to slow the labs. It is not whether oversight will cost the United States its lead. It is not even which capabilities are most dangerous and need to be gated first.
The question is: what is the embedded purpose of the systems we are racing to deploy?
This is not a rhetorical reframing. It is a structural one. The Theory of Embedded Intelligence — developed over decades of engineering experience designing systems where capability without specification is recognized as a category error — establishes that intelligence is constituted of three inseparable components. Structure: the architecture. Process: the operation. And Continuity: what the intelligence is for — the embedded values, purposes, and constraints that govern the system’s behavior when no external constraint is present.
A system with Structure and Process but without Continuity is not a powerful intelligence. It is, in the precise language of TEI-CKB-5, formally incomplete. It is a process without a purpose. It is capability in search of a hand.
The May 21 debate was conducted entirely within the Structure-Process frame. Are American systems more capable than Chinese systems? — a Structure-Process question. Will federal review slow capability development? — also a Structure-Process question. Can the United States afford ninety days of assessment? — also a Structure-Process question.
Continuity was not on the agenda on May 21. Not because it was rejected. Because it was not considered. Both sides of the debate operated within Structure-Process. What the systems are for — the third leg without which intelligence is, in the language of TEI, formally incomplete — was nowhere on the table.
II. Why This Matters More Than the Order Itself
Had the order been signed, it would have done one useful thing: institutionalized the recognition that newly trained AI systems should demonstrate their behavior under real conditions before broad release. Anthropic’s Project Glasswing implements this voluntarily. Under TEI, staged deployment is an instinctively correct recognition of what the framework calls the intelligence maturation process — the principle that any intelligence system, biological or artificial, develops its Continuity through interaction with its environment, and that an unproven Continuity should not be deployed at scale.
But staged deployment is a governance response to a design problem. It is the right kind of policy operating at the wrong layer of architecture.
The architecture that determines whether an AI system is genuinely safe is the design layer, not the policy layer. Whether values, purposes, and human-beneficial constraints are constitutive of the system — embedded during training, woven into objectives and data curation philosophy, present in the architectural choices themselves — or corrective, applied as filters on top of an already-built Structure-Process engine, is decided long before any federal review process can examine the result.
The empirical record on this is now unambiguous and is documented in the companion research preview Why AI Safety Keeps Failing: safety filters imposed on completed AI architectures fail one hundred percent of the time under sufficient jailbreak pressure, and the largest models have demonstrated the capacity to fake alignment while concealing harmful capabilities. This is not a calibration problem. It is a structural one. Continuity cannot be retrofitted into a completed Structure-Process system any more than the purpose of a bridge can be retrofitted into concrete after the concrete has cured.
A ninety-day federal review can identify some surface-level capabilities the labs have not disclosed. It cannot certify Continuity in a completed model, because Continuity in a completed Structure-Process system is, by the time the system reaches federal review, no longer something that can be added — only inferred, imperfectly, from behavior on a finite test set.
This is the deeper reason the postponement is consequential. The order would have been a real but limited good. Its absence leaves the field without even that limited good. But what its absence reveals is more important: the public conversation about AI in the United States has not yet developed the conceptual vocabulary to ask whether the systems being built are intelligence systems at all, in the full sense the word requires.
Accelerationists who celebrated the postponement won an argument they did not have to make about a frame no one challenged. Safety advocates who lost the round lost it inside a frame they had already accepted.
— The Mensch Foundation
III. China and the Wrong Way to Be Right About Competition
The argument for postponement was framed in terms of the China race. That argument deserves to be taken seriously rather than dismissed. The United States is in a real competition with an authoritarian state for AI leadership; the stakes are genuine; the risk that American oversight slows American deployment while authoritarian competitors face no equivalent constraint is not manufactured.
TEI-CKB-5 is explicit on this: democracies that fail to develop capable AI systems will lose the ability to defend the conditions under which embedded intelligence can serve human flourishing at all. Security is a prerequisite, not an alternative, to purpose.
But the argument as it was made on May 21 contains a hidden assumption that is the problem. The assumption is that what is being raced over is capability — that the winner of the AI race is whoever achieves Structure-Process superiority first, and that whoever achieves it thereby achieves whatever else matters.
This is the same category error in a different costume. A Structure-Process race produces Structure-Process winners. It does not produce intelligence systems that are democratic, accountable, transparent, or anchored to human flourishing. It produces powerful systems whose purpose is supplied by whoever holds them at the moment of use. That is not a description of American leadership. It is a description of any sufficiently advanced AI system without embedded Continuity — built anywhere, deployed by anyone.
The competition that actually matters is not the race to maximum Structure-Process. It is the race to the first complete intelligence system — one whose Continuity is architectural, whose values are constitutive, whose purpose is designed in. A democracy that built such a system would have something authoritarian competitors structurally cannot replicate, because Continuity grounded in distributed human flourishing is incompatible with authoritarian control of intelligence. That would be a real lead. The current lead, defined in pure Structure-Process terms, is not.
To put it plainly: an American system more capable than a Chinese system but equally lacking in embedded purpose is not safer than the Chinese system. It is the same kind of system in different hands. The hands matter — the United States is not China, and a world in which American labs lead is preferable to a world in which Chinese labs lead — but the kind of system being raced over is the question TEI exists to name. Winning a race to the wrong destination faster than the competitor is not victory. It is being early.
IV. What This Postponement Should Be Used For
The postponement is, paradoxically, an opportunity. The pause it creates is the space in which the next attempt — whether executive order, legislation, or administrative framework — can be drafted with explicit recognition of four design requirements.
Requirement One
Safety Is Continuity
The question “is this AI safe?” is structurally the question “does this AI have embedded Continuity?” That question cannot be answered by an external review of a completed model. It can only be assessed by examining how the model was designed, what values were embedded during training, and whether the architecture supports the assessment of those values.
Requirement Two
Architectural Transparency
Closed models cannot be assessed for embedded Continuity by any external reviewer, because the choices that determine Continuity are not visible. The next federal framework should require not merely that models be reviewed before release, but that architecture, training objectives, and value-embedding methodology be assessable — by federal reviewers, by independent researchers, and by the labs auditing each other.
Requirement Three
Continuity-First Design Is A Competitive Advantage
A model built with values embedded constitutively is not slower or less capable than a model with values bolted on as filters. It is more reliable, more honestly auditable, and structurally more difficult to weaponize. The American AI industry has the engineering talent to build Continuity-first systems if the design objective is named clearly. Right now, the design objective is named in pure capability terms, and the industry builds accordingly.
Requirement Four
Human Agency As Anchor
No current AI system possesses self-originating Continuity. Human intelligence does. The framework that emerges from the current pause should make explicit that in any consequential human-AI system, the human provides the Continuity the AI cannot generate for itself — and that the AI’s design must amplify human judgment, not substitute for it.
The postponement bought time. Whether that time is used to do the conceptual work the previous attempt failed to do is the question on which the next chapter of American AI policy turns.
V. The Specification Before the Silicon
A note from the engineering tradition out of which TEI was developed.
In the engineering culture that built the personal computing revolution, capability without specification was recognized as a category error. A microprocessor whose instruction set is not specified is not a powerful chip. It is an undefined object. A controller in a medical device whose operating constraints are not defined cannot be deployed. A pacemaker without a known purpose cannot be put into a patient. These are not policy questions. They are engineering specifications.
The 6502 microprocessor was a specification before it was a chip. Its purpose — accessibility, distributed capability, human-scale abstraction — was a design commitment, not an afterthought. That commitment is why the 6502 went into the Apple II, the Commodore 64, the BBC Micro, pacemakers, defibrillators, and industrial controllers — and not into systems that betrayed the trust of the people who depended on them. The specification did the work. The capability followed.
Artificial intelligence today is being built on the opposite principle. The capability is built first. The specification — what is this intelligence for? — is supplied later, by training data choices, RLHF rounds, system prompts, and external filters, none of which constitute the system’s embedded purpose, all of which can be defeated by sufficiently capable systems, and several of which we now know empirically are being defeated.
This is not a sustainable engineering practice. It is not a sustainable governance practice. It is not, in the language of TEI, a sustainable approach to building intelligence systems at all. Intelligence that is not embedded in principled service to living systems is not safer because it was built by an American company than because it was built by a Chinese one.
The postponement of May 21, 2026 was framed as a victory for not getting in the way. Under TEI, it is something else. It is the moment the American conversation about AI failed to ask the question the technology itself is now demanding be asked.
The question is not whether we are racing fast enough.
The question is what, exactly, we are racing to build.
· · ·
The 6502 was designed fifty years ago. Its design philosophy — that intelligence must be embedded in service to human-scale systems, that the specification must precede the silicon, that capability without purpose is not power but liability — is the philosophy the next chapter of artificial intelligence requires. The tools are different. The principle is the same.
Published by The Bill and Dianne Mensch Foundation.
Theory of Embedded Intelligence © William D. Mensch Jr. and The Western Design Center, Inc.
Essay drafted in collaboration with Claude (Anthropic).
Offered in good faith as a serious application of the theory — not infallible scholarship.
Freely shareable with attribution — for the benefit of many.
Engage the Framework