GPAI-DSL: A Domain Specific Assessment Language Ecosystem for General Purpose AI

Title: GPAI-DSL: A Domain Specific Assessment Language Ecosystem for General Purpose AI

General Purpose AI (GPAI) systems present vast opportunities alongside substantial risks. To ensure their responsible development and deployment, there is a pressing need for scalable, reproducible, and transparent methodologies to assess their capabilities and potential harms. 

Building on recent research into DSLs (Domain-Specific Languages) for fairness and bias auditing, this proposal introduces a broader vision: a modular ecosystem of DSLs designed to structure, standardize, and operationalize GPAI assessment across multiple dimensions. This ecosystem will provide a formal, extensible framework for defining evaluation criteria, simulating risk, auditing compliance, and benchmarking performance in line with EU values and the upcoming AI Act. The GPAI-DSL ecosystem offers a transformative approach to GPAI assessment: formal, flexible, and aligned with European principles. By enabling auditable, reusable, and standardised evaluation frameworks, it equips stakeholders with the tools needed to ensure trustworthy and responsible AI at scale.

Work Packages

WP1: Architecture & core infrastructure
This work package defines the overall system design, develops shared semantic components and ontologies, and sets up the architecture for interoperability between modules. It establishes the core domains to be supported and ensures future extensibility.

WP2: DSLs development
Each key domain—bias, robustness, alignment, interpretability, compliance, capabilities, and emergent behaviour—will be addressed by a dedicated DSL. These DSLs will define domain-specific semantics and syntax, with prototype tooling to support validation and use.

WP3: Integration & testbeds
This package connects the DSL ecosystem with existing AI models and platforms. It includes the development of benchmark scenarios and integration into testbeds representing diverse application domains.

WP4: Regulatory & ethical alignment
WP4 focuses on ensuring that DSL outputs align with AI Act requirements and fundamental rights. It includes legal analysis and collaboration with standardisation initiatives to ground the DSLs in current and emerging policy.

WP5: Dissemination & ecosystem building
This package handles the release of open-source tools, documentation, and community engagement activities. It includes events, training materials, and strategic outreach to stakeholders.

WP6: Pilots and validation
To test the ecosystem in real-world conditions, WP6 conducts pilots in sectors such as healthcare, finance, and education. It collects feedback, validates usability and generalisability, and refines the DSLs accordingly.