Human oversight UX for high-risk AI: designing for Article 14 of the AI Act
Gideon Awolesi ·
Article 14 of the AI Act asks that high-risk AI systems are designed so people can oversee them effectively while they're in use, including through "appropriate human-machine interface tools". In other words, a lot of the oversight duty lands on the interface. The Digital Omnibus on AI moved the dates, so you have more time than you might have planned for. Here's what the article asks for, and what it looks like on the screen.
01When it applies
The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published on 24 July 2026 and moved the dates for high-risk systems. The requirements for high-risk systems listed in Annex III, like AI used in hiring, credit scoring, or education, now apply from 2 December 2027. For AI that's part of a product covered by the laws in Annex I, like a medical device, they apply from 2 August 2028.
02What Article 14(4) asks the interface to support
Article 14(4) says the people assigned to oversee the system have to be able, as appropriate and proportionate, to:
- Understand what the system can and can't do, and monitor how it's working, including spotting anomalies and unexpected performance.
- Stay aware of automation bias, the tendency to rely too much on the system's output.
- Correctly interpret the output.
- Decide not to use the system, or to disregard, override, or reverse its output.
- Intervene or interrupt it, with a stop button or a similar procedure that brings it to a halt in a safe state.
Each of these is something a person does through the interface, so each one is a design requirement.
03Showing what the system can and can't do
Put the limitations where people make decisions, not only in the instructions for use. If the system is less reliable for some inputs or some groups of people, say so next to the output it affects. Show when something looks unusual, like an input outside the range the system was tested on, so people know to look more closely.
04Designing against automation bias
Automation bias is why a lot of oversight fails in practice. When the system is right most of the time, people stop checking. A few patterns help:
- Show the reasons or the evidence behind an output, so people have something to check instead of a bare answer.
- Show uncertainty, like a confidence level, in a way people can act on.
- For decisions that matter most, ask people to look at the evidence before they see the recommendation.
- Make agreeing and disagreeing equally easy, so accepting the output isn't just the path of least effort.
05Override, reverse, and stop
Make it obvious how to disregard or override an output, and record why people did it, because those reasons tell you where the system is weak. Make outputs reversible for long enough that people can catch a mistake. And where the system acts on its own, give people a clear way to stop it that leaves it in a safe state, and test that it works.
06What deployers have to do
If you use someone else's high-risk system in your organisation, Article 26 asks you to use it according to its instructions, and to assign oversight to people with the competence, training, and authority to do it, along with the support they need. Article 13 asks the provider to describe the oversight measures in the instructions for use, so check they're there before you rely on them.
07Test the oversight, not just the AI
The only way to know people can oversee the system is to watch them try. Give people from your user group realistic cases, including ones where the AI is wrong, and see if they notice, question it, and override it. If they accept a wrong answer, the oversight design isn't working yet.
If you'd like a second pair of eyes on an AI flow, send us one flow and one question for our free AI audit, and within five business days you'll get a recorded walkthrough and a one-page fix list.