AI in the Loop vs Human in the Loop

The difference between AI in the loop and human in the loop is who owns the decision. In a human-in-the-loop system the AI runs the process and a person is brought in to approve, correct or reject at set points. In an AI-in-the-loop system a person runs the process and makes every decision, and the AI is brought in to suggest, draft or check. A third term, human on the loop, describes a system that acts on its own while a person monitors it and can intervene. The same model can sit in any of the three; what changes is where the authority is.

Human in the loop, human on the loop, human in command

The clearest official definitions come from the European Commission’s High-Level Expert Group on AI, in its Ethics Guidelines for Trustworthy AI↗ (opens in a new tab) (April 2019). It names three oversight approaches:

• Human in the loop (HITL): “the capability for human intervention in every decision cycle of the system, which in many cases is neither possible nor desirable.”
• Human on the loop (HOTL): “the capability for human intervention during the design cycle of the system and monitoring the system’s operation.”
• Human in command (HIC): “the capability to oversee the overall activity of the AI system (including its broader economic, societal, legal and ethical impact) and the ability to decide when and how to use the system in any particular situation.”

Two details in those definitions are worth holding on to. First, the group itself says intervening in every decision is often neither possible nor desirable, so human in the loop is a tool for particular decisions rather than a default for everything. Second, it adds that “the less oversight a human can exercise over an AI system, the more extensive testing and stricter governance is required.” Less supervision in the moment has to be paid for somewhere else.

AI in the loop

AI in the loop turns the picture round. In a 2024 paper↗ (opens in a new tab), Natarajan, Mathur, Sidheekh, Stammer and Kersting argue that many systems described as human in the loop are really the opposite: “AI-in-the-loop (AI²L) systems, where the human is in control of the system, while the AI is there to support the human.” Their point is that such systems should be judged on how well the person and the AI do together, not on the model’s accuracy alone, because the person is an active participant who shapes the result.

In everyday terms: if you would describe the work as “I did it, with help”, the AI is in your loop. If you would describe it as “it did it, and I signed off”, you are in its loop.

Who decides, side by side

                  Who runs it    Who decides          When the person acts
AI in the loop    a person       the person, always   throughout
Human in the loop the AI         a person, at gates   before set actions
Human on the loop the AI         the AI, within       after the fact, or to
                                 its limits           intervene and stop

The labels matter less than the question underneath them: for this piece of work, whose decision is it? Teams get into trouble when the answer is unclear, when a person believes they are approving everything and the agent is in fact acting on its own for most of the day, or when an agent is treated as a decider for something a person should own.

Examples of each

AI in the loop

• A developer writes a function and accepts or rejects the editor’s completions line by line.
• An analyst asks an assistant to summarise twenty support tickets, then decides which problem the team fixes first.
• A project lead asks an assistant to draft next week’s priorities, rewrites half of them, and moves the cards into Next Up herself.

Human in the loop

• An agent writes a database migration and tests it on a copy; a person runs it on live.
• An agent drafts a refund decision with its reasoning; a person approves anything above a set amount.
• An agent finishes a bug fix and comments its evidence; a person reads it and moves the card to Completed.

Human on the loop

• An agent files what it notices into To Do all day; a person triages the lane once a week and deletes the noise.
• An agent fixes lint errors on branches without asking; a person reads its activity each morning and can revoke it.
• A nightly script labels new bug reports; a person samples a handful and corrects the labelling rules when they drift.

None of these is better in general. Which one a job should get depends on how reversible it is, how far a mistake spreads and how easily the result can be checked, which is the subject of human in the loop vs fully agentic AI. Where a person should step in once they are in the loop is covered in human in the loop for AI agents.

The failure each one invites

• AI in the loop fails when the suggestions are accepted so reliably that the person stops deciding. Nothing changed on paper, but the loop has quietly turned round.
• Human in the loop fails when the gates are so frequent that approval becomes a reflex. The person is present and no longer participating.
• Human on the loop fails when nobody reads the record. Monitoring that does not happen is simply full autonomy with a comforting name.

All three failures have the same cure: make it visible who did what and who decided, so that drift from one arrangement to another shows up rather than hiding in habit.

How a shared board makes either visible

The arrangement a team believes it has is only worth something if the record agrees. A task board where people and assistants are both members is a good place to see it, because every change is attributed and the lanes show where the decisions happen.

• On fenbs, History records every change with who made it, and an assistant’s changes read as “Claude via” the person it acts for. Read a week of it and you can tell which arrangement you actually have: whose name is on the moves into Next Up and Completed, and whose is on the rest.
• Decisions are kept on their own page with their own refs, DEC-001 onwards, saying what was decided, why, and who decided. A decision marked Decided is always made by people. An assistant can write one down or ask an open question with fenbs_add_decision, but it is never recorded as the decider.
• Moving between lanes is its own permission. Give an assistant a role that can add and edit tasks but not move them, and every move to Completed is a person’s, which is human in the loop enforced by the tool.
• Scopes set the outer limit. Connect an assistant with read and comment only and it is in your loop: it can report and suggest, and people make every change.
• Each task’s test status says who has checked the work. Needs owner check marks something only a person can confirm, so it stands out until one does.

fenbs_add_decision
  title:    "Drop support for the legacy export format?"
  status:   open
  context:  "Two customers still use it. Keep, deprecate, or remove?"
  tasks:    ["ENH-052"]

The open question then sits in the “Needs a decision” queue for a person to answer. The assistant has done the part it is good at, noticing and framing, and left the decision where it belongs.