Ariadne AI Lab
Unraveling the thread of complex government procedures
Ariadne AI Lab is a joint lab of the Archimedes Unit of the Athena Research Center with the Ministry of Digital Governance and partner public agencies. It develops and evaluates AI tools for the public sphere and civic services, so that public agencies handle their work faster and more consistently and citizens get better services.
The lab works where AI meets administrative practice, on legal and administrative text, on the workflows of public agencies, and on the requirements that AI systems in the public sector must meet. Tools are built on public or partner data, evaluated by domain experts, and operated under human oversight. Every tool is developed together with the agency that will use it and is tested on real work before it goes into use. Results feed publications and benchmarks.
Research areas
Legal NLP
Legal NLP applies natural language processing to legislation, case law, and administrative acts. The lab develops methods for structure extraction from long Greek legal documents, resolution of amendments into consolidated text, and extraction and classification of provisions and competences, with the Government Gazette as the main corpus. Consolidated, machine-readable legislation removes most of the manual work of codification and gives agencies and citizens a reliable picture of the law in force.
Retrieval-augmented generation
Retrieval-augmented generation couples a language model with a retrieval component, so that answers draw on a defined corpus instead of model memory alone. The lab builds curated knowledge bases of hierarchical legal sources, retrieval methods over legislation, decisions, and guidelines, and generation that attaches a citation to every claim, and compares open and closed models on the same tasks. Grounded answers with verifiable sources are a precondition for the use of language models in legal and administrative work, where an unsupported claim has consequences.
Responsible AI
Responsible AI covers the properties an AI system must have beyond accuracy, in particular explainability, fairness, and robustness. The lab develops methods that explain why a model produced an output, detect bias in model decisions, and assess robustness. AI in public administration affects rights and obligations, and systems whose behavior can be explained and audited are the ones that agencies and citizens can trust.
AI evaluation and benchmarking
Evaluation and benchmarking measure whether an AI system performs a task to the standard its use requires. The lab builds expert-labeled test sets and benchmarks for Greek legal and administrative tasks, maintains leaderboards of open and closed models, and produces reproducible evaluation reports across quality dimensions. The same infrastructure measures every tool the lab builds, before and during deployment. Public bodies need evidence before they adopt an AI tool and while it runs, and shared benchmarks make that evidence comparable across models, providers, and versions.
Human-in-the-loop AI for public administration
Human-in-the-loop AI keeps a person responsible for every decision while an AI system takes over reading, retrieval, drafting, and checking. The lab designs assistants for the officers who process cases, permits, and legislation, embedded in the workflow of the agency, with review and approval of every output, and studies how to measure usefulness through expert feedback, how to design review so that oversight stays meaningful, and how to move a tool from public data to live cases. Administrative decisions carry legal responsibility, and assistants that support the officer without taking the decision can be deployed in public agencies and improve throughput and consistency without transferring accountability to a system.
Projects
See the Research page for details and for datasets and code.
