By Burak Oktenli
As artificial intelligence is implemented into naval operations, the authority to act must have limits that survive communications failure.
Imagine an unmanned surface vessel conducting a maritime surveillance mission. Its onboard software has received valid instructions to distribute sensor information through a designated network. Hours later, a change in the operational picture leads a watch officer to withdraw that authorization. The cancellation is issued. But the vessel has temporarily lost communication. The original instruction remains digitally signed, intact, and recognizable to the receiving system. Nothing has been hacked. No artificial intelligence model has necessarily made a mistake. Yet an instruction that was legitimate when issued may no longer reflect the commander’s decision. The critical question is no longer whether the machine received permission. It is whether that permission still authorizes the next action.
This is a hypothetical scenario, not a reported Navy incident. Nevertheless, it exposes a problem that deserves explicit attention as the U.S. Navy brings artificial intelligence, unmanned platforms, and distributed command systems into operational service. Naval AI assurance must extend beyond approving a mission. It must establish how delegated authority expires, how revocation takes effect, and how commanders can determine what happened when communications fail.
The Problem Hidden Inside Delegated Autonomy
The Navy has compelling reasons to distribute decision-making authority. The 2024 Chief of Naval Operations Navigation Plan identifies the integration of robotic and autonomous systems as well as more effective fleet-level command and control as major priorities. The Department of the Navy’s 2026 Strategy to Weaponize Data and Artificial Intelligence further emphasizes the operational importance of data and AI. The Navy’s more recent Fighting Instructions make an equally important point that naval forces cannot depend on uninterrupted communications or constant permission from higher headquarters. These ambitions are compatible, but only if delegated autonomy is understood as bounded authority rather than unrestricted machine initiative.
A human commander may approve a mission, a software service may authenticate the command, and an AI-enabled component may select an action consistent with its instructions. None of those facts establishes that the action remains authorized indefinitely. A valid digital signature proves something important about an instruction’s origin and integrity. It does not, by itself, prove that the underlying permission has survived a later operational decision. Nor does a human approval recorded at mission launch settle every subsequent execution decision. Approval and execution may be separated by distance, time, multiple software systems, and unreliable communications. That separation creates a governance problem with direct operational consequences.
Authority Must Travel With its Limits
The solution is not to require every unmanned platform to ask headquarters for permission before every action. Such a requirement could undermine the very operational resilience that distributed maritime operations seek to achieve. Instead, the Navy should treat delegated machine authority as a clearly bounded operational permission. An authorization should specify the permitted class of actions, the conditions under which those actions remain valid, their duration, and the point beyond which new authorization is necessary. For example, permission to adjust sensor-reporting schedules within an approved mission may remain available during a temporary communications outage.
Permission to change the intended recipients of sensitive information may require a fresh authorization or a more restrictive preapproved rule. These are distinct authorities, even when the same software controls both functions. A platform operating without communications should therefore have a defined envelope of permissible behavior. Inside that envelope, it can continue its assigned mission. Outside it, the system must obtain renewed authority or enter an appropriate restricted operating state. The duration and scope of these permissions should reflect operational risk. A single universal expiration interval would be a poor substitute for mission judgment. The Office of Naval Research’s communications and networking program already identifies interference, contested radio-frequency environments, and dispersed forces as central challenges for naval information exchange. The consequence for AI governance is straightforward: a control system that functions only when the network is healthy is not sufficient for maritime operations. A disconnected system must know not only what it is permitted to do, but also when that permission runs out.
Cancellation Is Not the Same as Revocation
A second distinction concerns the meaning of a stop order. Imagine that a commander withdraws permission while a remote service has already prepared an action. The cancellation message reaches the coordinating system, but the downstream component continues using an earlier authorization. From the commander’s perspective, the instruction has been withdrawn. From the executing component’s perspective, the last valid authorization may still be in force. An acknowledgment from the coordinating system does not necessarily mean that the downstream action has been prevented. This distinction appears in ordinary distributed software systems, where authorization decisions and persistent operations can occur at different moments.
The National Institute of Standards and Technology’s zero-trust architecture guidance emphasizes the importance of enforcing access decisions close to the protected resource. That principle is particularly relevant when AI-enabled applications can initiate operations through other services. Naval systems need an equivalent distinction between permission to propose an action and authority to make its consequences effective. Therefore, a command-revocation mechanism should distinguish between a cancellation request, receipt of that request, prevention of future execution, and actions that have already occurred. These states cannot be treated as interchangeable.
A system cannot guarantee immediate compliance with a cancellation it has not received, whereas it can be designed to limit what remains permissible during that loss of contact. Similarly, a completed physical action cannot necessarily be undone because a later message withdraws its authorization. The realistic objective is not a magical remote stop button. It is an enforceable boundary on future actions, paired with an accurate account of those that have already taken place.
The Fleet Needs Evidence, Not Just an Approval Log
A third requirement is the ability to establish what happened. Suppose a command system records that an AI-enabled task was approved at one time and canceled later. Those records establish a history of command decisions. But they do not necessarily establish whether a remote component completed its task before or after the relevant cancellation became enforceable. This is why an execution record matters. For a properly designed system, evidence should connect the originating authority, the specific delegated permission, the execution decision, and the resulting operation. Such records would help commanders distinguish a valid action completed before revocation from an unauthorized action attempted afterward.
In addition, they would also support incident reconstruction, engineering correction, and more precise assignment of responsibility. Audit logs alone cannot prevent a dangerous operation. Cryptographic signatures alone cannot prove that an action remained authorized. Together with an enforceable execution boundary, however, they can provide a much stronger basis for operational assurance. This distinction matters even when existing authorization technology is functioning correctly. A policy engine can return the right answer while the resource that performs the action remains exposed to an outdated decision. The engineering challenge is to ensure that the final operation respects the relevant authority, not merely that an earlier approval was correctly recorded.
Put Revocation into Fleet Experiments
The Navy should turn these principles into testable requirements rather than leave them as broad assurances about human oversight. The Fourth Fleet’s hybrid-fleet experiment provides a relevant model for examining autonomous and networked capabilities in realistic operational settings. A dedicated command-authority exercise could begin with non-kinetic software operations, such as synthetic sensor-data distribution or controlled network configuration changes. The experiment should include a correctly authorized action, an action withdrawn before execution, competing approval and cancellation events, and an authorization that expires during a communications outage. It should then measure whether prohibited actions were prevented, when revocation became enforceable, what mission functions remained available during disconnection, and whether the recorded evidence accurately reflected the resulting operations. Most importantly, success should not be declared simply because a cancellation message was transmitted or a software service acknowledged it.
The meaningful question is whether the system prevented actions that were no longer authorized under the tested conditions. This approach would also expose the trade-offs. An excessively restrictive system could sacrifice operational continuity whenever communications deteriorate. An overly permissive system could allow outdated mission authority to persist long after its intended limits. Both failures matter. Testing should therefore measure mission continuity and authority enforcement together, rather than treating them as competing objectives examined in isolation.
Command Responsibility Cannot Be Delegated Away
There is a larger command issue behind this technical problem. The Navy’s Enhanced Mission Command framework recognizes that forces must continue operating when higher headquarters cannot provide continuous direction. AI-enabled systems can support that approach, but they also create additional layers between a commander’s intent and the operation ultimately performed. Those layers must not turn a limited delegation into permanent authority.
A commander should be able to determine what a system is authorized to do, what it could continue doing without communications, and which actions require renewed approval. When a mission ends or authority changes, the system’s permitted behavior should change according to rules established before the disruption. The Navy does not need to eliminate initiative from its autonomous systems, but it needs to define the conditions under which that initiative remains legitimate. The most valuable future naval AI system may not be the one that acts most independently. Instead, it may be the one that can continue acting appropriately when disconnected, recognize the boundaries of its delegated authority, and provide trustworthy evidence afterward.
The next naval AI assurance question should be simple: When an order is withdrawn, what prevents the machine from acting on yesterday’s permission? Until that question has a tested answer, approval at the beginning of a mission should not be mistaken for control over everything that follows.
Burak Oktenli, MBA, is a Master of Professional Studies in Applied Intelligence student (in progress) at Georgetown University in Washington, D.C. He received his bachelor’s degree in computer science and engineering from the University of South Florida. His work examines trustworthy artificial intelligence, autonomous systems, and the limits placed on automated decision and execution authority. His essay “Audit the Machine” received First Prize in the 2026 U.S. Naval Institute/Naval Intelligence Professionals Naval Intelligence Essay Contest.
