US Federal AI Policy — Structural Alignment

Part of Structural Alignment in Practice. A structural reading, not legal or policy advice.

As of this reading, the United States has no comprehensive federal AI law. Federal direction currently comes from executive action, a national-security memorandum and a legislative proposal still in draft. This page reads those instruments through the six pillars of structural alignment, and notes where they have little or nothing to say.

Basis and status. Read in September 2026 from public summaries and reporting and, for NSPM-11, from its published pillars. This is not a clause-by-clause review of primary texts. The instruments are moving: the act is a discussion draft, and executive instruments can be replaced. Where a summary was silent, this page says "not found in the sources reviewed", not "absent".

The instruments read

A word on "alignment"

In US policy discussion, alignment often refers to harmonizing rules across levels of government, as in the proposed preemption of state law. That concerns which government sets the rules. Structural alignment concerns whether the rules, whoever sets them, preserve accountable reasoning inside systems. This page uses the word only in the second sense.

At a glance

Reading through the pillars

Structural Embedding

NSPM-11's assurance pillar reaches furthest. It calls for rigorous testing and verification throughout, and for assurance, by contract or other means, that a fielded system cannot be disabled, degraded or modified without federal knowledge and approval. That is a control over change, close to a promotion gate, but it applies to national-security systems. The draft act and the June order work through disclosures, audits and voluntary evaluations, which are events. None of the summaries reviewed asks that continuity be built into how a system operates.

Decision Lineage

Disclosure of model information and published safety frameworks describe the model. In the summaries reviewed, nothing requires preserving the reasoning behind individual decisions made with a deployed system.

Continuity of Agency

NSPM-11's accountability pillar keeps responsibility with commanders, directors and agency heads at every level of command. That establishes who answers. It does not address whether the responsible person's understanding of the system stays current between decisions, which is the temporal condition Continuity of Agency names.

Audit and Repair

Third-party audits and incident reporting in the draft act, and testing and verification in NSPM-11, correspond to the audit side. The structural questions are whether audit works from preserved state or from reconstruction, and whether repair keeps the record intact. The summaries describe audits as periodic and incident reports as events; they do not say.

Stewardship of Structure

Licensed independent verification organizations, a codified federal standards center and whistleblower protections in the draft act create custodians of verification and standards. Who is answerable for the continuity infrastructure of a deployed system is not addressed in the sources reviewed.

Visibility

Incident reporting and disclosure make some outcomes visible to authorities. Nothing in the summaries reviewed requires that determinations made in autonomous execution be traceable to where they were made. Research on agent infrastructure names attributing actions to specific agents as a core function (Chan et al., Infrastructure for AI Agents); the instruments reviewed do not appear to address it.

What the gap suggests

Read together, the instruments concentrate on developers of frontier models, on national-security use and on periodic events: disclosures, audits, reports and evaluations. The pillars concerned with how deployed systems operate over time (embedding, agency and visibility) have little or no counterpart in the sources reviewed. Structural alignment holds that these are where accountability is lost.

This is a finding about coverage, not a criticism of any instrument's purpose. A memorandum on national-security adoption, a developer-focused draft and a voluntary evaluation framework were not written to answer these questions. It is also an observation about the instruments read, not a claim that no such provisions exist elsewhere in US law: state laws, sector regulators and future drafts may address them. By contrast, the EU AI Act reaches providers and deployers across the system lifecycle, and each of the six pillars has a reading there.

This page takes no position on whether any instrument should change.

Related