Human in the loop

Human in the loop, often abbreviated HITL, describes a system where a person reviews and approves what an automated process produces before it takes effect. The machine does the work and proposes a result. A human makes the decision. The pattern covers everything from a radiologist confirming a model's read of a scan to an assistant asking before it sends an email on your behalf.

The phrase came from control engineering and military simulation, where it meant a human operator sat inside the control loop rather than outside it watching. Machine learning then borrowed it twice. Once for training, where people label data and correct model output, and once for deployment, where people approve what a deployed model does. Both senses still travel under the same name, so it is worth asking which one a vendor means.

In an AI product the loop takes one of two shapes. Either the model acts and a person reviews afterwards, which is usually called human on the loop, or the model proposes and a person approves first, which is human in the loop proper. The second is slower and is the only one that prevents a wrong write rather than detecting it. Which shape a product chose is visible in one place: whether an agent's write lands in the data or in a queue.

The version that actually holds has one more property. The credential that can propose must not be the credential that can approve. If a connected agent can both write a fact and confirm it, then an agent that read a malicious instruction hidden in a web page or an email can approve its own writes, and the review step is decoration. Separating the two is what turns a review screen into a boundary.

The usual failure is not the absence of review but the exhaustion of it. A queue that demands a decision on every trivial item trains people to approve in bulk without reading, which is worse than no review at all because it produces a record that looks confirmed. Designs that survive keep the loop for writes that carry consequence, batch the rest, and make undo cheap, so attention is spent where being wrong is expensive.

The idea also carries legal weight. Article 22 of the GDPR gives people a right not to be subject to decisions based solely on automated processing where those decisions have legal or similarly significant effects, and the EU AI Act requires effective human oversight for high-risk systems. In those settings a human in the loop is not a design preference. It is what makes the system permitted.