When space systems increasingly rely on automation and AI, there needs to be ways to audit the decisions those systems make. (credit: Boeing)

Decisions without deciders: authority at machine speed in the space enterprise

by Bharath Gopalaswamy and Daniel “Sphinx” Dant
Monday, August 10, 2026

In February, Operation Epic Fury opened with space and cyber forces rather than with aircraft. Less remarked was a second shift underneath it. A growing share of the decisions that determine who sees, who talks, and who moves in a contested environment are now executed by software in milliseconds, against rules written months in advance. The governance problem this creates is simple to state: machine-speed allocation now makes sovereign decisions before any accountable human can reconstruct them. At machine speed, governance is whatever was specified before the crisis.

This had to happen, and it is the right development. At the tempo of modern conflict and the scale of proliferated constellations, no human-in-the-loop process could keep pace with the volume of allocation and routing choices a crisis generates. The programs that delivered this capability moved quickly and were right to move quickly. What has not advanced at the same rate is the record that lets anyone reconstruct, afterward, why a system decided as it did. That gap is narrower and more fixable than it looks.

Let’s start with how these decisions actually present. They rarely look like decisions. In a crisis touching orbital infrastructure they appear as configuration: bandwidth prioritization, service tier enforcement, geofencing, imagery release thresholds, and the throttling rules that determine whose traffic degrades first. Each is defensible as engineering. In aggregate, under pressure, they allocate national capability, and the commercial incentives bearing on those choices are not the same as the operational ones.[1] A commercial operator exercising ordinary contractual discretion can determine which military formation stays connected, which hospital network holds, and which exchange clears. It is acting within its rights. The open question is not whether it may, but whether anyone can afterward establish why.

When allocation is performed by a learned model rather than a rule table, that answer is frequently unavailable. That’s not because anyone is concealing it, but instead because the system was never asked to produce it. Logs record what happened, but they rarely record what was weighed, which alternatives were scored and set aside, or under what authority the action was taken. Without that record there is no way to verify that contractual priority was honored, no way to compensate a party that went dark, and no way to improve the system from the event. This is a specification gap rather than a limit of the technology.

The operational reality is already here. Maven shows how far the capability has come. Established in 2017, it became a program of record at the National Geospatial-Intelligence Agency in 2023 and now supports the Pentagon’s combined joint all-domain command and control effort. What began as a system for identifying objects of interest can now integrate data from multiple sources, track targets, recommend which weapons are available against them, and compress sensor-to-shooter timelines from hours to minutes.[2] In March 2026, the Deputy Secretary of Defense consolidated oversight under the Chief Digital and Artificial Intelligence Office and named AI-enabled decision-making the cornerstone of that architecture.[3] This is genuine progress, delivered at pace. The natural next step is to specify what such systems should be able to show about their own reasoning.

Golden Dome makes the timing concrete. An architecture of proliferated sensors, space-based interceptors, and battle management at machine speed, carrying a 2028 demonstration, will encode allocation rules whether or not they are debated.[4] The engineering choices being made this year are governance choices in a different vocabulary, and they are far cheaper to set now than to revisit once fielded.

There is a strategic dimension as well. Dual-use sensing and interceptors blur the boundaries between defense, surveillance, targeting, and preemption, an ambiguity that long predates autonomy and that has shaped space security debates since the first anti-satellite tests.[5] When sensing and interceptors are cued automatically on commercially operated systems, an adversary assesses American intent against a process that is hard to reconstruct even from the inside. Deliberate ambiguity is a legitimate instrument of strategy, but ambiguity the originating party cannot resolve is not.

Other sectors settled this in calm conditions, and the precedents are binding, not advisory. Telecommunications operates priority frameworks written into obligation, among them Telecommunications Service Priority and the Government Emergency Telecommunications Service.[6] Electric reliability standards, aviation safety duties, and financial stress-testing follow the same logic: obligations defined before the event, binding on the operator, surviving a change of ownership. Department of Defense Directive 3000.09 governs autonomy in weapon systems and requires appropriate levels of human judgment over the use of force.[7] It does not reach the allocation layer, where a great many crisis decisions will actually be made.

Civil practice is moving first. On August 2, European obligations for high-risk artificial intelligence took full effect, requiring that such systems allow the automatic recording of events across their lifetime.[8] Those provisions reach hiring tools and credit scoring. No standing equivalent applies to a system that recommends which weapon answers which target. That is not because anyone judged military systems to need less traceability. It is because the question has not yet been put.

The precedent for setting rules ahead of the event also exists inside this domain. The first National Security Space Strategy, issued in 2011, described a space environment growing congested, contested, and competitive, and argued for shaping that environment deliberately rather than responding to it after the fact.[9] Much of that agenda was carried out. The allocation layer is the part of it that was never finished, largely because the technology that would make it urgent did not yet exist.

Nor does the fix require research. Auditability is a design property, not an overlay, and it means something specific. It is neither a dashboard nor an explanation generated after the fact by a second model, which yields a plausible account rather than a true one. At a minimum, it means three records: the state observed and the confidence assigned, so a reviewer can distinguish a poor decision from one made on poor information; the options available and the scores they received, since an allocation is only assessable against the alternatives; and the model version in force at the moment of action, because thresholds are tuned and models are retrained. All of this is routine in fields where decisions must be defended. Published work has already demonstrated response systems that reason under explicit legal constraints, using deontic logic to return a set of permissible options rather than a single opaque output.[10] The methods exist and are documented. What is absent is the requirement to use them.

Three acquisition requirements would close most of the gap, and the Department of Defense can impose all three now, through instruments it already holds. The first is decision provenance as a contract requirement: any system holding a sovereign-critical allocation function should produce an auditable record of the inputs received, the rule or model invoked, the alternatives considered, and the authority claimed, written into the request for proposals, scored in source selection, and demonstrated at acceptance. The second is a priority framework for orbital services, negotiated in peacetime and written into contract, with defined tiers, override conditions, and compensation for commercial parties who lose revenue when a priority order is exercised. The third is a named authority able to order, override, and answer for an allocation decision under declared emergency conditions, with the legal basis established in advance rather than assembled during the event.

The owners of the problem are identifiable today: the Chief Digital and Artificial Intelligence Office, which now holds oversight of AI-enabled decision-making; Space Systems Command, which is writing Golden Dome’s requirements this year; and the combatant-command requirements writers whose operational plans assume the allocation layer will hold.

None of this runs against the commercial interest. Vendors build to the requirement they are given, and where a specification is silent no source selection will score a feature nobody asked for. A clear and uniform requirement applied before award is easier for industry to price and to meet than a standard assembled afterward through inquiry. That has been the experience of every sector that has been through this.

Capability has moved faster than the paperwork, which is what happens when a country is serious about delivering. The paperwork is the cheaper half, and it is the half still available to us. The specifications are being written now, and the only question is whether anyone accountable is holding the pen.

Notes

1. Bharath Gopalaswamy and Daniel Dant, “Golden domes, fragile firms: the business risks of AI-enabled space infrastructure,” The Space Review, March 16, 2026.
2. Palantir Technologies, “Palantir Expands Maven Smart System AI/ML Capabilities to Military Services,” September 20, 2024; Center for Strategic and International Studies, “What Is Maven Smart System, and What Does It Do?” June 2026.
3. DefenseScoop, “Feinberg’s new Maven directive sets AI-enabled decision-making as ‘the cornerstone’ for CJADC2,” April 3, 2026.
4. Space Systems Command, public release on the Space-Based Interceptor program, describing a demonstration capability integrated into the Golden Dome architecture by 2028.
5. Bharath Gopalaswamy, Final Frontier: India and Space Security (Westland/Tranquebar, 2019).
6. Cybersecurity and Infrastructure Security Agency, Telecommunications Service Priority (TSP) and Government Emergency Telecommunications Service (GETS) program documentation, cisa.gov.
7. Department of Defense Directive 3000.09, Autonomy in Weapon Systems, updated January 25, 2023.
8. Regulation (EU) 2024/1689, Article 12 (Record-keeping); obligations for high-risk systems applicable from 2 August 2026.
9. Department of Defense and Office of the Director of National Intelligence, National Security Space Strategy: Unclassified Summary, January 2011.
10. T. Deb, M. Jeong, C. Molinaro, A. Pugliese, A. Quattrini Li, E. Santos, V.S. Subrahmanian, and Y. Zhang, IEEE Transactions on Cybernetics 54, no. 12 (2024): 7147–7162, presenting a framework for multi-objective decision-making under legal constraints using deontic logic and Pareto-optimal status sets.

Bharath Gopalaswamy, PhD, is an aerospace, defense, and emerging-technology executive. Col. (Ret.) Dan Dant, Vice President at KBR, is a defense industrial-base strategist. Gopalaswamy and Dant are Senior Fellows of the National Spacepower Center. The views expressed are the authors’ own.

Note: we are now moderating comments. There will be a delay in posting comments and no guarantee that all submitted comments will be posted.

Home