There are no common rules for risky AI — verifiable building blocks are needed
Sh0ny
18 августа 20262 min read
In short
For high-risk systems the problem is now not only how well the model works but how to prove it complies with different laws. The review proposes machine-checkable compliance artefacts, yet shows itself that without interoperability between regulatory regimes they will not remove the bureaucracy.
The main risk for developers of high-risk AI today is not the absence of ethical principles but the mismatch between mandatory rules. The EU, the US and China have different triggers for risk classification, different requirements, liability mechanisms and means of oversight. So one and the same system can comply with one regime and remain problematic under another.
The review's authors propose looking at regulation not as a set of general declarations but as a checkable structure. Their idea is Knowledge Blocks: machine-checkable compliance blocks described through RDF/OWL, SHACL and PROV-O. They can record requirements, constraints, data provenance and evidence that a particular product meets those requirements.
The practical sense here is fairly down to earth: instead of gathering documents by hand for every inspection, an organisation could accumulate an audit trail as the system is developed. The approach matters especially where AI directly affects people or allocates scarce resources.
As a test the authors consider three scenarios: rehabilitation robotics with EEG, AI-assisted debt collection in CBDC ecosystems, and the allocation of scarce GPUs in AI Factory infrastructure. These are not three finished products but a stress test for the question: can AI regulatory requirements be linked to sector norms and data protection rules?
But "machine-checkable" does not mean "automatically compatible". The authors highlight weak interoperability requirements between jurisdictions, the difficulty of applying AI regulation, sector norms and data protection simultaneously, and insufficiently developed governance of critical digital infrastructure. Knowledge Blocks can make evidence of compliance tidier but do not resolve the conflict of laws itself.
That is precisely why the value of this work is not a promise of a single global standard. It is rather an attempt to turn compliance from a scattered folder of documents into structured data that can be checked and reused. The only question is whether regulators will agree on even the format of such evidence.
If you had to launch a high-risk AI system tomorrow, which would be the bigger problem: proving compliance in one country, or reconciling the requirements of several regimes? Source: cs.AI updates on arXiv.org