// DefinedTermSet

Human-in-the-loop glossary

Plain, citable definitions for the vocabulary of human oversight of AI — the terms the rest of the writing leans on, from the AI Act to the everyday craft of keeping people in the loop.

01Human-in-the-loop (HITL): A design in which a person reviews, approves or corrects an AI system’s output before it takes effect. The human sits inside the decision path, so no consequential action is taken without explicit human input.
02Human-on-the-loop (HOTL): A design in which the AI system acts on its own but a person monitors it and can intervene, pause or override at any time. Oversight is supervisory rather than step-by-step, which fits high-volume or time-critical tasks.
03Human-in-command: The principle that people retain overall authority and responsibility for an AI system — deciding whether, when and how it is used — even when day-to-day operation is automated. It frames AI as a tool serving human goals, not the other way round.
04Meaningful human oversight: Oversight that genuinely lets a person understand, question and stop an AI system — not a rubber-stamp click. The EU AI Act (Art. 14) requires high-risk systems to be built so an overseer can grasp their limits, watch for automation bias and intervene or halt operation.
05Automation bias: The tendency to over-trust a machine’s recommendation and stop checking it — accepting wrong answers and missing errors a human would otherwise catch. It is the central failure that meaningful human oversight is meant to counter.
06Augmentation vs. automation: Two opposite goals for deploying AI: augmentation makes a person better and faster at their work while they keep deciding; automation removes the person from the task entirely. The choice shapes job design, risk and where human judgement still applies.
07Escalation / hand-back: The moment an AI system passes a case to a human because it is uncertain, out of scope or high-stakes. Well-designed hand-back includes the context and reasoning the person needs to take over without starting from scratch.
08Calibrated trust: A user’s reliance on an AI system that matches how reliable it actually is — trusting it where it is strong and doubting it where it is weak. Good interfaces and clear performance signals help build it; opaque ones push people toward over- or under-trust.
09Automated decision-making: A decision produced solely by automated processing, without a human assessing it. Under the GDPR (Art. 22) people have the right not to be subject to such a decision when it has legal or similarly significant effects, save for narrow exceptions with safeguards.
10Transparency obligations (AI Act Art. 50): Duties under Article 50 of the EU AI Act to tell people when they are dealing with AI: chatbots must reveal they are machines, and synthetic or manipulated media (deepfakes) must be marked as artificially generated. These rules apply from 2 August 2026.
11High-risk AI system: A category in the EU AI Act for AI used in sensitive settings — hiring, credit, education, healthcare, justice and the like — where errors can seriously affect people’s rights. Such systems carry the strictest duties, including risk management, data quality and human oversight.
12Deployer: In the EU AI Act, the organisation that uses an AI system in its own operations (as opposed to the provider that builds it). Deployers of high-risk systems must assign competent human oversight, follow the instructions for use and monitor how the system behaves in practice.
13Reskilling and upskilling: Reskilling trains people for a different role as their old tasks are automated; upskilling deepens their current skills to work alongside AI. Both are how organisations adopt AI while keeping their workforce, rather than replacing it.
14Change management: The deliberate work of guiding people through a new way of working — communication, training, redesigned roles and feedback — so an AI rollout actually sticks. Most failed adoptions are change-management failures, not technical ones.
15Oversight UX: The interface design that makes human oversight workable: showing the AI’s confidence and reasoning, flagging cases that need review, and making intervention quick and low-friction. Poor oversight UX is a common reason oversight exists on paper but not in practice.
16Accountability chain: The documented line of who is answerable for an AI-supported decision — from the provider and deployer to the person who signed it off. A clear chain ensures that when something goes wrong, responsibility rests with people, not the model.
17Fail-safe / stop button: A reliable way to halt an AI system and return to a safe state — the “stop button” that meaningful human oversight depends on. To be real it must remain reachable, fast and effective even when the system is confident it is right.
18Model drift: The gradual decline in an AI model’s accuracy as the real world moves away from the data it was trained on. Because drift is silent, it is one of the strongest reasons to keep humans monitoring AI long after launch.