Caught On Camera: Moscow On Fire

THE HILL REPORT

Terminal Guidance Is Not Autonomy

Connor Hill · InsightfulWord · September 2, 2026

A great deal of the language used about machine decision-making in weapons is imprecise in a way that makes claims impossible to assess. The imprecision is not accidental; the terms carry weight in policy debates, and stretching them is useful to whoever is stretching them.

The underlying distinction is between two very different capabilities that are frequently described with the same words. One is the ability of a device to recognize a shape and steer toward it. The other is the ability of a device to decide what constitutes a target.

The first is old. Machine vision guidance in the terminal phase of a flight has existed since the 1970s, when systems began comparing what a sensor saw against a stored reference and correcting course from the comparison. Nothing about it required the machine to choose anything; the choice had been made before launch and encoded as the reference.

The second is what policy documents mean by autonomy in the selection and engagement of targets, and it is defined in writing, subject to a review process, and constrained by a directive that has been public for over a decade.

That directive is the reason precision matters here. It creates a category, defines it, and attaches procedural requirements to anything falling inside it. A system that merely recognizes a shape does not fall inside it. A system that selects targets without further human input does.

Reporting rarely makes the distinction, partly because the observable evidence is the same. A device that flies to a target and strikes it looks identical from outside whether the target was designated by a person minutes earlier or chosen by the device itself.

Distinguishing them requires information about the control architecture rather than the outcome, and that information is either disclosed by the operator or it is not available.

What follows is what the governing directive actually requires, the difference between recognizing and deciding, the review that occurs before such a system can be fielded, where the human sits in each architecture, and what would constitute genuine evidence that a system was operating autonomously.

What the Directive Actually Requires

The relevant American policy instrument is a Department of Defense directive on autonomy in weapon systems, first issued in 2012 and reissued in updated form in January 2023.

Its central requirement is a phrase that has generated a substantial interpretive literature: systems must be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force.

The phrase is deliberately not a fixed rule. It does not mandate a human approving each engagement, and it does not permit unsupervised operation. What it requires is that the level of human judgment be appropriate to the system, the mission and the circumstances, and that the appropriateness be established through a defined process rather than asserted.

The directive also creates definitions that carry legal and procedural consequence. An autonomous weapon system is one that, once activated, can select and engage targets without further intervention by an operator. A semi-autonomous system is one that engages only individual targets or groups that have been selected by a human operator.

That definitional line is where nearly all the confusion in public discussion sits. A weapon that a person points at a designated object, and that then completes the engagement using onboard sensing, is semi-autonomous under this definition regardless of how sophisticated the onboard sensing is.

The directive further establishes a senior review requirement, an obligation that such systems be subject to rigorous verification and validation, and an instruction that they be designed to minimize unintended engagements and to allow human operators to terminate an engagement.

The Difference Between Recognizing and Deciding

The technical distinction beneath the policy one is worth stating plainly, because it is what makes the definitions tractable.

Automatic target recognition is a classification problem. A sensor produces data, an algorithm compares it against stored characteristics, and the system determines whether what it is seeing matches a category it was built to match. The output is an identification with some confidence attached.

Selection is a different operation. It requires determining that an identified object should be engaged, which involves rules about what is a legitimate object, whether the circumstances permit engagement, and whether the expected consequences are acceptable. Those are judgments with legal content.

A system can be extremely good at the first and have no role whatsoever in the second. Indeed most fielded precision weapons are exactly that: highly capable at finding and hitting a designated thing, with no authority to decide what the designated thing should be.

Conflating them produces two symmetrical errors. Sophisticated terminal guidance gets described as autonomy, which inflates what exists. And genuinely autonomous selection, where it occurs, gets described as an incremental improvement in guidance, which understates it.

Both errors serve someone. The first sells urgency; the second deflects scrutiny.

There is a third capability that sits between them and is often the real subject of a claim: target prioritization. A system may be given a set of candidate objects by human operators and then order them, allocate effectors among them, or choose timing. That is machine decision-making with real consequence, and it is neither classification nor selection in the definitional sense. It is also where most current investment actually sits.

The Review That Happens Before Fielding

The procedural machinery attached to the autonomous category is the part most often omitted from discussion, and it is substantial.

Systems falling within the definition require senior review before formal development and again before fielding. The review involves the under secretaries responsible for policy and for research and engineering together with the vice chairman of the joint chiefs.

The review examines whether the system has been subjected to rigorous hardware and software verification, whether realistic operational testing has been conducted, whether the human-machine interface is understandable to trained operators, and whether the system will function as anticipated in realistic conditions against adaptive adversaries.

There is also a separate and older requirement: a legal review of any new weapon for consistency with the law of armed conflict, conducted before acquisition and again before fielding. That obligation applies regardless of autonomy and predates the directive by decades.

The 2023 reissue added a working group to oversee implementation and clarified that the directive applies to systems incorporating artificial intelligence capabilities, alongside the department's separate responsible artificial intelligence guidance.

None of this establishes that the process is adequate, and reasonable people disagree about whether it is. It does establish that the category is administratively real, which is the relevant point when a claim is made that something has crossed into it.

Where the Human Actually Sits

Control architectures are conventionally described in three arrangements, and the vocabulary is useful for reading any claim about a system.

In the first, a human authorizes each engagement. The machine may identify, track and calculate, but it does not act without a specific instruction for that specific engagement.

In the second, the machine proceeds unless a human intervenes. Authority has been granted in advance within defined boundaries, and the human role is supervision with a veto. Defensive systems against saturating threats have operated this way for decades, because the timelines involved are shorter than human reaction.

In the third, there is no human in the loop for the duration of the engagement, either because communication is impossible or because the design does not provide for it.

Most real systems sit in the second arrangement with boundaries that vary by mode, and the boundaries are what determine the answer. A system may require authorization in one mode and operate on supervision in another, with the mode selected by a person before the engagement begins.

This is why the question that resolves most claims is not whether a system uses machine learning. It is which of the three arrangements applies, under what conditions, and who set them.

What Would Constitute Evidence of Full Autonomy

For a reader trying to assess a claim rather than accept or reject it, there are specific things that would count as evidence and specific things that would not.

Video of a device striking a target is not evidence, because the footage is identical under every architecture. Nor is the presence of onboard image processing, which is standard in terminal guidance and has been for many years.

What would count is documentation of the engagement authorization: a statement of what a human did and when, relative to the strike. Operators who describe their systems as human-authorized generally say so specifically, because it is a claim they want on the record.

Procurement and requirements documents count, because they state what the system was specified to do. So do the results of the review process where a system falls inside the definition, and the existence of a review is itself informative.

Investigative reporting that identifies the control architecture from interviews with operators counts. Statements by manufacturers about product capability are weaker evidence, since capability marketing and fielded configuration routinely differ.

The absence of all of these, combined with dramatic footage and confident description, is the ordinary condition of coverage in this area and should be treated as an absence of information rather than as an implication either way.

The general observation is that the category exists, is defined, and has a process attached; and that whether any particular system falls inside it is a question of fact that specific documents would answer and dramatic imagery never will.