Human-in-the-loop vs on-the-loop vs out-of-the-loop.
Updated 16 October 2026 · BlackByrne · ~5 min read
Human-in-the-loop means a person must approve each action before it happens. Human-on-the-loop means the system can act on its own, but a person supervises and can intervene. Human-out-of-the-loop means the system acts entirely on its own, with no one able to intervene in the moment. The distinction is not academic: it decides who is accountable when an effect is committed — and whether anyone could have stopped it.
The three levels, plainly
- In the loop. The machine proposes; the human disposes. Nothing happens without an explicit human decision for that action. Accountability is unambiguous.
- On the loop. The machine acts; the human watches and can veto. Control depends on the human noticing in time — which, under saturation or a lost link, is exactly when supervision fails.
- Out of the loop. The machine acts alone. Fast, and unaccountable in the moment. This is the category the public debate about "autonomous weapons" is really about.
What DoDD 3000.09 actually requires
DoD Directive 3000.09, reissued January 25, 2023, requires that autonomous and semi-autonomous weapon systems be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force. It does not ban autonomy; it sets review, testing, and design requirements so that human judgment is preserved where it matters. Separately, the DoD adopted five AI Ethical Principles in February 2020 — responsible, equitable, traceable, reliable, governable — that an AI advisory system is expected to meet.
Why "by architecture" beats "by policy"
Many systems claim to be human-in-the-loop as a configuration or a policy — a setting an administrator could change, or a mode an operator could switch off under pressure. That is a promise, not a guarantee. A stronger claim is human-in-command by architecture: the system is built so that no effect can be committed without a sustained human authorization, with no mode, setting, or failure path that routes around it. The difference is auditability. A policy can be quietly disabled; an architectural constraint can be tested and proven.
That is the line DaggerOS holds. Its reasoning layer proposes and ranks; it never arms, fires, or emits on its own; and every effect passes a single Human Gate enforced in code, aligned to DoDD 3000.09. You can even hold the gate yourself and watch an effect fail to commit when you let go early.
FAQ
What is the difference between human-in-the-loop, on-the-loop, and out-of-the-loop?
In-the-loop: a person must approve each action. On-the-loop: the system can act but a person supervises and can veto. Out-of-the-loop: the system acts alone with no human able to intervene in the moment.
What does DoDD 3000.09 require?
Reissued January 25, 2023, it requires autonomous and semi-autonomous weapon systems to be designed so commanders and operators can exercise appropriate levels of human judgment over the use of force. It sets review and testing requirements rather than banning autonomy.
Sources: DoD Directive 3000.09 (reissued 25 Jan 2023); DoD AI Ethical Principles (24 Feb 2020). Primary documents at esd.whs.mil and ai.mil.
Related: Security & compliance · Air-gapped C2 (DDIL) · Verify it yourself