Control · C-002 · Human Oversight & Contestability

Human-in-the-loop review

Consequential AI outputs and decisions are reviewed by a competent person, with the time, information and authority to change them, before they take effect.

Control type: Preventive
Family: 09 Human Oversight & Contestability
Risks addressed: 11
Audience: Heads-of · Practitioner
Frameworks: Article 14 · Clause 6 · AML.M0029 · Articles 22A to 22D
Responsible AI domain: 09 Human Oversight & Contestability
Lifecycle stage: 3 Design & build
5 Approve
7 Operate & monitor
Reviewed: 2026-11-11

What it does

It places human judgement between the system and the consequence. For defined classes of output, such as decisions about people, advice, published content and significant actions, a person examines the case and decides. The control is the whole arrangement: who reviews, what they see, how long they have, what they may do, and what happens when they disagree. The business owner of the process is accountable; operations provide the reviewers. It is designed with the system and operates on every case in scope.

How to implement it

Define which outputs need review, by consequence and reversibility.
Show the reviewer the inputs, the reasons and the system's uncertainty, not only its conclusion.
Where practicable, have the reviewer form a view before seeing the system's.
Set workloads from how long a careful review takes.
Give reviewers authority to override, and make disagreeing no harder than agreeing.
Record each review and its outcome, and feed overrides back to the model owner.

Evidence it operates

Design: a review design stating scope, information, time and authority. Operation: review records; override rates and reasons; workload figures against the design assumption; examples of outputs changed on review.

Where it fits

Between generation and effect. It is a primary control for inadequate human oversight design, ineffective human review, unfair algorithmic management and excessive agency. For agents acting on their own, approval gates for high-impact agent actions applies the same idea, and reviewer effectiveness testing checks that the review works.

Limits

People defer to automated recommendations, so a review can exist and catch nothing. It does not scale: as volume grows, time per case falls. Reviewers cannot check what they are not shown. Review adds cost and delay, which creates steady pressure to thin it out.

Risks this mitigates

Primary mappings

Excessive agency in autonomous agents High: A person reviews consequential outputs before they take effect. An agent is given more functions, permissions or independence than its task requires, so an error, an ambiguous instruction or a manipulated input turns into a damaging action that no one approved.
Inadequate human oversight design High: Designs a review step with the information and authority to change the outcome. An AI system is built with no human oversight, or with an oversight step that lacks the time, information, competence or authority to change an outcome.
Ineffective human review High: Gives reviewers the time, information and authority to disagree. A human review step exists and operates, but reviewers do not in practice detect or correct the system's errors.
Unfair algorithmic management High: Keeps a person responsible for decisions about individuals. AI is used in recruitment, task allocation, performance assessment or dismissal without fairness testing, explanation, human review or a route to challenge.

Supporting mappings

Automated decision-making without safeguards Critical: Meaningful human involvement takes the decision outside the solely automated category. AI is used to take significant decisions about individuals with no meaningful human involvement and without the safeguards UK data protection law requires.
Automation bias and over-reliance High: Review design can require an independent judgement before the output is shown. People accept AI outputs without the scrutiny they would give another source, and the system's errors pass into decisions unchallenged.
Defamation by AI output High: Human review of content that will be published. A system publishes a false statement about an identifiable person or organisation that damages their reputation.
Miscalibrated confidence Medium: Routes low-confidence cases to a person. The system presents uncertain outputs with unwarranted confidence, or gives no usable signal of uncertainty, so users cannot tell when to check.
Misinformation and confabulation High: A person reviews consequential outputs. A generative model states false or fabricated information with the same fluency and confidence as accurate information, and the output is relied on.
Prompt injection High: A person approves the consequential actions an injection would try to trigger. Text supplied by a user, or contained in content the model retrieves or receives from a tool, is interpreted as an instruction and overrides the instructions the system was given.
Unsuitable AI-generated advice or guidance High: A qualified person approves personalised recommendations. A system gives personalised recommendations in a regulated domain without the suitability process, permissions or accuracy that advice requires.

Framework and clause references

Framework	Clause	Title
EU AI Act	Article 14	Human oversight The human oversight requirement.
ISO/IEC 42001:2023	Clause 6	Planning & risk
MITRE ATLAS	AML.M0029	Human In-the-Loop for AI Agent Actions Human approval of agent actions.
UK General Data Protection Regulation	Articles 22A to 22D	Automated processing and significant decisions Meaningful human involvement.

How to cite

Axiom Verity (2026). Human-in-the-loop review. Reference Library, C-002. https://axiom-verity.com/controls/human-in-the-loop-review. Reviewed 2026-11-11.

Found an error or a better source? Report a correction.