GAIJ undersøkte korleis moderne språkteknologi kunne støtte meir systematisk analyse av skattemateriale og tilhøyrande forvaltningsdokument. Som eit pilotprosjekt utvikla prosjektet ein fungerande prototype som vart publisert på nett. Prosjektet hadde som hovudfokus å hente ut strukturert informasjon frå i stor grad ustrukturert tekst, til dømes opplysningar om inntekt, frådrag, eigarskap og omtale av motpartar, og å representere dette i ein konsistent datamodell. Ved å kombinere informasjonsekstrahering med enkel kopling av identitetar og aktørar, bygde arbeidet sporbare samanhengar for “kven–kva–når” som gjorde det lettare for analytikarar å orientere seg i store mengder skattemateriale og peike ut saker som burde undersøkast nærare. Vektlegginga var metodisk og operativ: å styrkje openheit og etterprøvbar tolking av dokument, heller enn å lova automatisk avdekking av lovbrot.
The project has delivered a functioning prototype that demonstrates, in a bounded way, that Norwegian scanned tax records can be transformed into a structured knowledge graph and explored through an interactive web interface. In practical terms, we have processed on the order of 33 000 documents, constructed a Neo4j graph with companies, people, auditors and addresses, and exposed this via a browser-based tool that a small group of investigative journalists has begun to test. At this stage, however, the actual outcomes are still modest: the system is closer to an advanced demonstrator than to an operational tool that could be relied upon for high-stakes investigations.
The potential impact is considerably larger, but contingent on addressing several non-trivial obstacles. First, the reliance on optical character recognition (OCR) and large-language models (LLMs) means that data quality remains uneven, especially for numerical and tabular content, and systematic validation is currently hampered by the lack of temporally aligned registry data. Second, the computational cost of the extended extraction pipeline severely restricts scalability: extrapolating from current performance, full coverage of the available corpus would require resources well beyond what was available in this project. Third, the graph database and visualisation stack begin to show performance and usability limits as the data size grows, casting doubt on the feasibility of naïvely scaling the current architecture to a truly national scope with substantial IT support.
If these issues can be overcome – through more efficient models, better historical data access, and more scalable graph technologies – the approach has the potential to support more systematic scrutiny of corporate networks and tax practices, and to serve as a template for similar tools in other jurisdictions and domains. At present, however, the impact is best described as laying a technically credible foundation and clarifying the challenges ahead, rather than delivering a ready-to-use instrument for transforming journalistic or auditorial practice.
Illicit financial transactions constitute a serious ethical challenge to the proper functioning of society. Various regulatory frameworks enforce banks, companies, and governments to keep checks on illicit financial flows, yet in gross amount, these still comprise a significant portion of the world's trade value. At the forefront of exposing malicious entities involved in non-sustainable activities are investigative journalists. Yet assessing financial data in a digital world is a strenuous task. GAIJ – Graph-bound Artificial Intelligence Journalism – is a project centred on utilising modern open-source Large Language Models (LLMs) to classify illicit transactions in financial and tax records. The target of this pilot project is to develop a prototype open-source Artificial Intelligence (AI) model that examines and classifies transaction data, creating a graph of suspicious interactions between companies.
This proposal is at the forefront of current research on the application of AI tools in Investigative Journalism. One core and distinctly unconventional aspect of GAIJ is that it utilised LLMs to classify data, not generate text. GAIJ uses not one, but a collection of LLMs to generate text regarding interactions between companies and classify actual interactions as regular or suspicious. This is the most striking, cutting-edge method employed in GAIJ, and thus naturally also the most challenging. Various questions and research avenues are open: Are LLMs capable of serving as classification tools, particularly of unlikely events? Are LLMs sufficiently powerful to unravel relevant information from interaction data like financial transactions? Are they able to discern between regular and illicit activities? On a large scale, GAIJ falls into the present pertinent scope of understanding what role AI models have in serving as tools to improve society and how can they aid, for example, journalists in untangling unsustainable financial practices from companies.
