Skip to content
←Back to glossaryAI

Human oversight

An obligation, set by the AI Act, to maintain effective human control over a high-risk AI system in order to prevent or correct its failures.

Definition

Tempe, Arizona, March 2018: an Uber Volvo XC90 in self-driving testing killed Elaine Herzberg as she crossed at night pushing her bicycle. According to Hulu's records obtained by Tempe police, the safety operator was streaming The Voice on her phone minutes before the impact. Article 14 of the AI Act requires that a high-risk AI system can be overseen by natural persons while it is in use. The operator must be able to understand the system's outputs, spot anomalies, override a decision or stop the system when needed. The rule targets automation bias, the tendency to over-rely on a machine's recommendation, and works alongside other safeguards such as the human in the loop. A token presence is not enough: oversight takes skills, suitable interfaces and a real ability to act. Oversight that is missing or only for show is a breach, and it can ground a finding of fault or defect under the liability regime.

Its 6 cross-references, woven across the glossary's 18 categories
Human oversight: AI glossary illustration, AlgoPolis
Example

Textbook case: an underwriting-support tool must let an underwriter review and override an automated refusal. Otherwise the human oversight required by the AI Act exists only on paper.

Read next
AIHuman in the loopThe principle of keeping human oversight or validation inside an automated decision. Its real value depends on the quality of the control.Read the entry →AIAutomation biasThe human tendency to place excessive trust in the outputs of an automated system, to the point of ceasing to check them, which amplifies the AI's errors.Read the entry →Law & regulationEU AI ActThe European regulation governing AI through a risk-tiered approach, imposing enhanced obligations on high-risk systems.Read the entry →
Sources
  • Tempe Police Department, investigation report on the March 18, 2018 collision, released June 2018 (reported by Bloomberg, June 22, 2018)
Related terms
Related articles
#11The EU AI Act and Liability for High-Risk Systems
On the Wire, where this term came up
September 18, 2026Mosaic ships its first AI product eight days after launching its lab
Stay informed

Follow the research

One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.

No spam, one-click unsubscribe, no data ever sold.