EU AI Risk Categorizations
The law creates a lot of new terminology...and requirements.
I’ve read every word of the European Union (EU) Artificial Intelligence (AI) Act.
It’s a beast.
Frankly, this thing could have been a lot shorter if it didn’t have so much repetition and extremely convoluted language.
With that said, StackAware reviewed the legislation to deliver actionable guidance. In this post, we’ll take a look at the risk categorizations the Act establishes:
General purpose
These are systems built on (a) general purpose AI model(s), characterized by:
Ability to perform wide range of distinct tasks
Large amounts of training data
Self-supervision at scale
Additional characteristics are specified in Annex XIII.
Systemic risk
This categorization is specific to general purpose AI systems.
Systemic risk systems are those with a “significant impact on the Union market due to their reach, or due to actual or reasonably foreseeable negative effects on”:
Fundamental rights
Society as a whole
Public security
Public safety
Public health
Deep fake
This is image, audio or video content that appreciably resembles existing:
persons
entities
objects
events
places
and would falsely appear to a person to be authentic or truthful.
High risk
Any AI system that is or is involved with:
Intended to be use as safety component; and
Underlying product is already EU-regulated
Criminal behavior risk assessment
Education admissions/decisions
Job recruitment/advertisement
Exam cheating identification
Public benefit decisions
Emergency call routing
Migration and asylum
Election management
Critical infrastructure
Health/life insurance
Law enforcement
Credit scoring
Types of high risk AI systems are specifically laid out in Annex III of the Act.
Not high risk
Self-explanatory.
Prohibited
With the exception of anything related to national security (a huge carveout from the EU AI Act) systems doing the following are banned:
Real-time public biometrics for law enforcement
Inference of non-obvious traits from biometrics
Criminal profiling not on criminal behavior
Purposefully manipulative or deceptive
Inferring emotions in school/workplace
Exploits vulnerabilities of a group
Blanket facial image collection
Social scoring
Need to keep up with the EU AI Act’s requirements?
StackAware put together a detailed 5-day email course laying them out.