Semantic Patterns of Prohibited AI Systems in the EU AI Act

Delaram Golpayegani*, Harshvardhan J. Pandit**, Dave Lewis*

* ADAPT Centre, Trinity College Dublin

** AI Accountability Lab, Trinity College Dublin

​

​

​

Presentation for NeXt-generation Data Governance workshop, 3 September 2025

​

Introduction

• EU AI Act
• Regulation for AI (entered into force Aug 2024)

​

• Risk-based approach

​

​

​

Transparency Risk

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Prohibited AI Practices

• Strictly banned
• Non-compliance → heavy fines
• Challenge: vague legal wording, hard to determine prohibited systems

​

Research Objective:

• Facilitate identification of prohibited AI systems by using Semantic Web technologies

​

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Our Prior Work on using Semantic Web technologies for Compliance with the AI Act

AIRO

AI Risk Ontology

VAIR

Vocabulary of AI Risks

has specialisation

Annex III high-risk determinator

(SHACL shapes)

Queries for generating documentation

(SPARQL)

AICat

(DCAT extension)

AIUP

(ODRL extension)

describes constraints using

queries

reuses

is aligned with

AI Cards

​

​

​

​

​

​

​

is used to generate

Integrated into DPV

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Methodology

1) Annotate prohibited AI conditions to extract 5 key concepts (from our prior work): domain, purpose, capability, deployer, subject

2) Assess sufficiency of the 5 concepts to uniquely describe prohibited conditions

3) Identify additional concepts

4) Extend AIRO & VAIR with new concepts (to be proposed to DPV)

5) Express rules in machine-readable formats

AI Act analysis

Developing Semantic Web-based approaches

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Annotation of the Prohibited Conditions

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Key Concepts for Determining Prohibited AI

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Patterns of Prohibited AI

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Codified Rules – SHACL

• SHACL: W3C standard for constraint validation
• Used to encode prohibited AI rules as shape graphs
• Need to use negation (sh:not) to produce validation reports
• Challenge:
• Complex nested structures for risk representation
• Use of negation of rules

​

Example: AI chatbot impersonation case → prohibited art 5(1)

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Codified Rules – N3

• N3 rules
• Simplified if-then representation
• Easier to express complex prohibited conditions
• Offers more flexiblity

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Advantages of Semantic Web-based approach

• Standards-based, interoperable, and open
• Supports automation of compliance tasks
• Transparent rule-checking
• Extensible with future Commission guidelines and amendments

Limitations

​

• Framework supports but does not replace legal expertise

​

• Scalability challenge for annotating AI use cases

​

​

Semantic Patterns of Prohibited AI Systems| Delaram Golpayegani, Harshvardhan Pandit, Dave Lewis| golpayes@tcd.ie | NXDG | September 2025

Semantic Patterns of Prohibited AI Systems in the EU AI Act

Delaram Golpayegani, Harshvardhan J. Pandit, Dave Lewis

golpayes@tcd.ie, Pandithj@tcd.ie, delewis@tcd.ie

​

11

https://regtech.adaptcentre.ie

This work has received funding from the European Commission's Horizon Europe Research and Innovation Programme under grant agreement No. 101177579 (FORSEE), the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 813497 (PROTECT ITN), and from the ADAPT Centre for Digital Media Technology, which is funded by Research Ireland and is co-funded under the European Regional Development Fund (ERDF) through Grant#13/RC/2106_P2. Harshvardhan J. Pandit is a member of AI Accountability Lab, which is funded under John D. and Catherine T. MacArthur Foundation grant with project #216001 and award #19034.