Human in, on and out of the loop

Every autonomous system does the same three things: it senses, it decides, it acts. The only question that really matters is where the human sits in that chain.

There are three answers, and the choice between them is a trade between speed and control.

01 · Human in the loop

Sense (detects) → Decide (recommends) → Human approves (go or no-go) → Act (executes)

The machine detects and recommends, but nothing happens until a person approves it. An AI flags a hostile drone on the operator's screen, and the operator decides the response.

This is the most accepted model on ethical, legal and policy grounds. The cost is speed. The human is the bottleneck, and the link back to that human has to stay up.

02 · Human on the loop

Sense (detects) → Decide (chooses) → Act (executes), with a human monitoring throughout and able to veto or stop it.

The machine acts on its own while a person supervises. This suits cases where there simply is not time to ask: point defence against incoming missiles, or against a drone swarm.

The risks are over-trust and automation bias, and the very short window the supervisor actually has to step in.

03 · Human out of the loop

Human sets the mission first, then Sense → Decide → Act runs alone.

Once launched, the machine senses, decides and acts entirely by itself. Fire-and-forget munitions work this way. It is the fastest model and it keeps working under jamming.

It also raises the hardest questions about accountability and compliance with the laws of armed conflict. Many states restrict it, and policies such as DoD Directive 3000.09 require careful review of any system that selects and engages targets autonomously.

Automatic, automated, autonomous

The three loop models describe the role of the human. These three words describe the intelligence of the machine, and they get used interchangeably when they should not be. The cleanest way to tell them apart is to ask how each one handles a surprise.

• +Automatic. Performs one fixed action when a trigger fires. No decision at all, just an if-then rule, in a fully predictable environment. A tripwire, a proximity fuse, an automatic door. Needs almost nothing from a human.
• +Automated. Follows a programmed sequence using sensor inputs, with pre-defined rules and branches, in a structured and well understood environment. A drone flying pre-set waypoints, or an assembly-line robot. Needs setup and monitoring.
• +Autonomous. Perceives, decides and adapts in order to reach a goal. Chooses among options and handles what it did not expect, in an environment that is uncertain, changing and sometimes contested. A drone that finds and tracks a target through GPS denial and jamming. Needs goal-setting, oversight, and trust.

The gap between the second and the third is the whole engineering problem. An automated system fails the moment reality stops matching the script. An autonomous one is judged on what it does when the script runs out.

That is the layer we build, and it is why the position of the human is a design decision rather than a policy footnote.