Human oversight is designed in, not added on
High-risk systems must be designed so that humans can effectively oversee them — understand the capacity and limits, stay alert to automation bias, interpret the output, decide not to use it, and intervene or stop the system. They must also log automatically throughout their lifetime.
Why it is like this
Oversight that depends on the operator's goodwill is not oversight. Putting it in the design makes it something a conformity assessment can check.
Not to be confused with
A human clicking approve. Article 14 names automation bias explicitly: an interface that nudges the operator to accept the output has not provided effective oversight.
In practice
For certain biometric identification systems, Article 14(5) adds the four-eyes rule: no action on the basis of the identification unless separately verified by two people.
Source: Regulation (EU) 2024/1689 (AI Act) — Article 12 and Article 14. Read on .
Odd one out Check yourself
Three of these are capabilities Article 14 requires oversight to provide. Which one is not?
Show the answer
Retrain the model when accuracy drops
Retraining is the provider's lifecycle duty under Article 9, not an oversight capability the interface must give the operator.