News · 14 March 2024

The EU AI Act for engineers: reading the risk tiers

On 13 March 2024 the European Parliament approved the AI Act. Formal steps remained, but the shape of the regulation was clear. For people who build models, the legal text is less important than one design idea: obligations depend on what the system is used for, not on how sophisticated it is.

Four tiers, one question

The Act sorts uses into broad tiers. A small set of practices is prohibited outright. A larger set of "high-risk" uses, such as certain medical, employment, education and critical-infrastructure applications, must meet requirements before and after they reach the market. Some systems, like chatbots, carry transparency duties: people should know they are talking to a machine. Everything else is largely left alone.

For an engineer the practical question is therefore: which tier is my intended use in? The same image classifier can be minimal-risk in a photo app and high-risk as a component of a diagnostic device.

What high-risk means in practice

High-risk systems are expected to have, among other things, a documented risk management process, data governance for training and testing data, technical documentation, automatic logging, transparency toward the people who deploy the system, human oversight, and appropriate accuracy, robustness and cybersecurity. People affected by certain decisions also gain a right to an explanation of the role the AI system played.

Read as an engineer, that list is not exotic. It is what a careful team would do anyway: know your data, record what you did, test it, make the output interpretable and keep a human able to intervene. The difference is that you will need to show it.

Explanation is a requirement, not a feature

"Transparency" and "human oversight" are only meaningful if the person overseeing the system can see why it produced an output. A bare probability score does not help a clinician decide whether to trust a finding. A saliency map, a feature attribution or a retrieved similar case might.

That is why we treat explainability as part of the system specification from the first sprint rather than a layer added before an audit.

What to do now

  1. List your AI use cases and mark the likely tier for each.
  2. For anything plausibly high-risk, start keeping decision logs and dataset documentation today.
  3. Decide who the "human in the loop" is and test that they can actually interpret the output.

Dates and secondary legislation will keep evolving, but the engineering habits above will not go out of date.

Written in March 2024; details reflect the situation at that time.

All news