Imagine you are a future person. You are not a combatant by any definition. But to a machine you are a pattern — a configuration of data points within parameters set by a human being some time earlier, and from a room you never knew existed. You match a machine’s target profile. You match that target profile with sufficient probability to engage you.
No human looked at you. No human made a decision about you specifically. Humans set the rules. The machine found you. The machine engaged you.
This is not science fiction. It is the logical endpoint of permissive legislative language currently being written in Washington and debated in Geneva. The people writing it probably know exactly what they are permitting. If they don’t know, that might be more frightening.
2026
2026 arrived carrying a deadline. The United Nations Secretary-General called for a legally binding instrument on lethal autonomous weapons systems to be concluded by year’s end. One hundred and fifty-six nations voted to support this. NGO’s have spent over a decade angling on it. Congressional offices are now drafting legislation. The language of “human control” is here, per schedule.
And yet the future you remains exactly as exposed as you would have been without the language. The reason is in the language itself.
The Language of Human Control
Human control, meaningful human control, human involvement, human authorization, autonomous and semi-autonomous — these terms share a common structural feature. None of them specify what a human being must actually do before a machine engages. What do these terms permit?
These terms permit humans to set parameters of engagement. Parameters to define target profiles. They permit target profiles made up of attributes like behavioral patterns, location data, communication signatures, movement characteristics. Machines to be deployed. Machines finding persons matching the target profiles. These terms permit machines engaging persons.
At what point in that sequence did a human authorize the engagement of any specific person?
The answer, under the current language, is that no such authorization is required. What is required is human involvement at the parameter-setting stage. What is required is that humans remain accountable for the rules under which the machine operates. What is not required — what the current language carefully avoids requiring — is that a human look at a specific person and then decide that this specific person be engaged.
That distinction is the essential concept of human control, and yet it has been tossed away in today’s language of human control.
Consider the AI Guardrails Act, introduced by Senator Slotkin. Section 2(b)(3) requires “appropriate levels of human judgment and supervision” of autonomous systems. Yet the Act never defines whether that judgement and supervision must occur before a lethal action — or merely after, or alongside, or at the policy level. The Act constrains these systems within its apparent oversight framework, but does not require that a human authorize the engagement of any specific target. The word “appropriate” is doing all the structural work of the Act, and it is doing that work without a definition.
Section 2(c) includes a waiver provision. The Secretary of Defense can waive the autonomous weapons prohibition for up to one year, renewable, if “extraordinary circumstances affecting national security” require it.
Section 2(c)(1): The waiver standard is also that the machine’s error rate must not exceed the error rate of trained human operators doing the same job. That’s not a human control standard. It is a machine performance standard. Slotkin’s Act explicitly benchmarks the machine against the human and says: If the machine is as good as or better than the human, the human can be removed from the decision.
None of these things are drafting oversight. It is the pattern. Across current legislative and treaty language, “human control” consistently refers to control over systems, over parameters, over rules of engagement — not control over the specific decision to engage a specific person. The 2026 standards are being written to accord the fullest potential lethality of AI weapons systems.
Parameters Are Not Decisions
Setting the rules under which a machine will later make ten thousand engagement decisions is not accountability for those decisions. It is accountability for the rules. The machine holds the engagement mechanics. And the machine will never really be able to answer for those mechanics.
A parameter set once that engages ten thousand times is not ten thousand human decisions. It is one human decision — shaped by policy makers, engineers, and military or law enforcement commanders — followed by ten thousand machine engagements.
For decisions around the use of machine force, humans are being expected to retain accountability. Humans seem to be expected to be accountable for the use of autonomous weapons systems. But read carefully: These expectations are that humans are accountable for the system, for its deployment, for its parameters. None of these expectations require a human to be accountable for the specific decision to kill a specific person — because under the permissive language of 2026, no human makes the decision to engage. The machine just engages.
Human Control of Weapons Act
The 2014 Human Control of Weapons Act was a legislative proposal designed specifically to close this gap between human and machine. The Act required positive human action to authorize the machine engagement of each specific individual target. Not parameters. Not profiles. Not rules of engagement. Not what a machine must or must not do. But what a person must do.
That standard was not written in ignorance of how AI weapons systems operate. It was written in full awareness that AI-driven weapon systems by their very nature avoid exactly any requirement of individual-level human control. Of course, the policy language being developed alongside today’s new weapons is being shaped to accommodate that avoidance.
On the way to 2026, the weapons policy community tried hard to classify, and utilize, human-machine interactions as “in the loop,” “on the loop,” or “out of the loop.” But those distinctions, too, blurred the a moral gap between “humans setting the parameters” and “humans authorizing each specific engagement.” That gap is the difference between human control and human-initiated machine autonomy. It is the difference between accountability and the appearance of accountability.
In truth, the humans responsible for a machine’s parameters are, at the moment of engagement, doing something else entirely. Playing video games. Or trolling in X. Maybe idling their engine in a Starbucks drive-through. Some will be sleeping.
The humans matching the target profile will be the only humans held accountable in any meaningful way.
AI Arms Race
Who wins? Who loses? Who pays? Today’s permissive language will also surely not slow worldwide autonomous weapons proliferation. It is authorizing the proliferation.
If the standard will be that humans set the parameters within which autonomous systems operate, then the incentive structure of military competition drives weapons development toward systems that operate faster, at greater scale, across broader parameters. Nothing in today’s permissive language prohibits that. 2026 legislation is the permission structure for the arms race it seems to be trying to claim to govern.
The person who builds the fastest, most scalable system operating within the broadest parameters retains every competitive advantage any arms race offers. In the age of AI, accepting a standard requiring specific human authorization for each engagement would require accepting a significant operational constraint. The incentive to abandon the human control of weapons is undeniable. And the language being written in 2026 has already accommodated that abandonment.
Your Future Person
Return to the person at the beginning. You are a pattern. You matched a target profile. No human decided to engage you. A human decided to build a system. Other humans decided to deploy it. Still other humans set its parameters. But the decision about you — the specific, irreversible choice about you — was made by a machine, under rules written to ensure that no one would ever have to make it.
Human institutions cannot step outside “the machine” to stop it. The institutions writing this language are themselves a part of the machine — a part of something that determines who profits, who labors, and who gets targeted and/or denied.
Demanding better language from institutions will still only get us the world they want, written in terms they find acceptable.
In the language of autonomous machines, it will get us a world in which human life is laid before the speed and efficiency of the machines of those who control the machines.
What remains is individual choice. This is not a choice to demand better human institutions, although we can still choose to do that. It is a choice to refuse the premise entirely — that machines should ever engage human beings on our behalf.
That clarity does not require the language of institutions. It does not require any permission. It requires only that we each make that choice.
The language our society is using to govern lethal force tells us what our society believes about human life. The language of 2026 is being written now. Read it carefully.
Senator Slotkin’s Proposed AI Guardrails Act
Senator Schiff’s Proposed Act