Login Get a Demo
01 Glossary · Governance and risk

What Is Human Oversight?

What EU AI Act Article 14 requires, who has to deliver it, and why human-in-the-loop is one way to meet the requirement and not a synonym for it.

The short answer

What is human oversight of AI?

Last updated
Read time
4 min read
From
Cranium AI
Also called
Human oversight of AI
Share on LinkedIn

Human oversight is the set of design features and organizational measures that let people understand, monitor, intervene in and, when needed, override or stop an AI system while it is in use.

Its purpose is to prevent or minimize risks to health, safety and fundamental rights that other controls do not remove.

Key takeaways
  • EU AI Act Article 14 requires high-risk AI systems to be designed so that natural persons can effectively oversee them.
  • Deployers must assign oversight to people with the competence, training, authority and support to do it (Article 26(2)).
  • Human-in-the-loop is one way to deliver oversight. It is not a synonym for it.
  • Article 14(4)(b) names automation bias: the overseer has to stay aware of the pull to over-rely on output.
By the numbers Each figure links to its sourceAll sources 
02 Article 14

What must an overseer be able to do?

Article 14(4) of the EU AI Act lists five abilities. A high-risk system has to be provided in a way that makes each one possible.

Article 14(4)The person overseeing the system must be able to
(a)Understand the system’s capacities and limitations and monitor its operation, so anomalies and unexpected performance are caught
(b)Stay aware of automation bias, the tendency to rely too much on the output
(c)Correctly interpret the output
(d)Decide not to use the system, or to disregard, override or reverse its output
(e)Intervene in its operation or halt it, for example with a stop button
03 Compare

Human oversight vs. human-in-the-loop

Human oversight is the requirement. These three are ways to meet it, as defined in the European Commission’s 2019 communication on trustworthy AI.

Human-in-the-loopHuman-on-the-loopHuman-in-command
What the human doesIntervenes in every decision cycle of the systemCan intervene during design and monitors the system in operationOversees the system’s overall activity and decides when and how to use it
FitsLow-volume, high-stakes decisions about one personHigh-volume or real-time systems with thresholds and alertsDecisions to deploy, restrict or withdraw a system
LimitThe Commission notes it is in many cases neither possible nor desirableDepends on alerts being seen and acted on in timeToo far from single outputs to catch an individual error
04 Who does what

Who is responsible for human oversight?

01

The provider designs it in

Under Article 14(3), oversight measures are built into the system by the provider, or identified by the provider for the deployer to implement. They must match the system’s risks, its level of autonomy and its context of use.

Article 14
02

The deployer staffs it

Article 26(2) requires deployers to assign oversight to natural persons with the necessary competence, training and authority, and the support to act. Article 26(5) adds a duty to monitor operation.

Article 26
03

Outside the EU

The voluntary NIST AI RMF asks that oversight processes be defined, assessed and documented (MAP 3.5). For US federal agencies, OMB M‑25‑21 requires human oversight, intervention and accountability for high-impact AI.

NIST AI RMF
05 Keep going

Related reading.

06 Questions

Frequently asked questions.

Straight answers, kept current. Last updated .

Ask us anything →
Is human oversight the same as human‑in‑the‑loop?

No. Human oversight is the outcome: people who can understand, intervene in and stop a system. Human-in-the-loop is one design for getting there, in which a person intervenes in every decision cycle. The European Commission’s 2019 communication lists it beside human-on-the-loop and human-in-command, and notes that it is often neither possible nor desirable.

Does human oversight mean a person must approve every AI decision?

No. Article 14(3) ties oversight measures to the system’s risks, autonomy and context of use. There is one stricter rule: for remote biometric identification, Article 14(5) allows no action on an identification unless at least two natural persons have separately verified it, with exceptions for some law enforcement, migration, border control and asylum uses.

When do the EU human oversight rules apply?

Article 14 is a requirement for high-risk AI systems. On the European Commission’s timeline, which reflects the Digital Omnibus on AI, the rules for Annex III high-risk systems apply from December 2, 2027, and the rules for high-risk AI embedded in products covered by Annex I apply from August 2, 2028.

How does human oversight apply to AI agents?

The abilities are the same, but the design question shifts from reviewing outputs to controlling actions. Decide which actions an agent may take alone, which need approval first and who can halt it, then keep a record of what it did. NIST MANAGE 2.4 asks for mechanisms, with assigned responsibility, to supersede, disengage or deactivate a system that acts outside its intended use.

07 See it for yourself

From the definition
to the evidence.

See how Cranium AI discovers, observes, governs, secures and proves every AI model, agent and vendor in your enterprise.