Companion to Human oversight of AI
Each relationship explains what connects two topics, which sources support it and where the evidence stops.
Read the full brief →Showing all 5 connections
Lines indicate a relationship discussed in the research. They do not establish causation, strength or probability. Records, our analysis and open questions are labelled separately. Edition 1 · 14 November 2026.
Documented relationship
The reviewed EU provisions connect human oversight with authority and intervention capabilities.
Evidence limit: The legal text does not establish who has working permissions in a particular system.
Sources: [2]
Documented relationship
The experiments linked explanations with acceptance, including wrong advice.
Evidence limit: Bounded by the tested tasks and designs.
Sources: [4]
Documented relationship
Agudo’s experiments varied when incorrect support was shown and found different judgment outcomes.
Evidence limit: An experimental effect is a reason to test workflow, not a universal design prescription.
Sources: [6]
Documented relationship
Cognitive-forcing designs reduced overreliance in Buçinca’s study while receiving less favorable ratings.
Evidence limit: More friction is not automatically better; effects were context-dependent.
Sources: [7]
Our analysis
Our inference from GAO’s assessment structure is that traceable records make oversight claims examinable.
Evidence limit: Records enable scrutiny; their existence does not prove that harm was prevented.
Sources: [8]