Our Take
This is investigative journalism using AI as a labor tool, not a truth-finding oracle; the human reporting and source verification remain non-delegable.
Why it matters
Newsrooms are resource-constrained and document-heavy investigations are slow; if AI can efficiently surface patterns in leaked data, the bottleneck shifts from finding signals to verifying them. This matters now because the financial and political reporting community is actively testing whether LLMs can compress timeline and cost without compromising accuracy.
Do this week
Investigative teams: audit your current document-review workflow against what AI can do (entity extraction, relationship mapping, anomaly detection) before hiring or outsourcing; cost and speed may collapse faster than editorial processes can absorb.
Financial Times deployed AI to investigate Orbán
The Financial Times used machine learning to process leaked documents and construct a network map of corruption allegations surrounding Hungarian Prime Minister Viktor Orbán. The reporting team applied AI to identify patterns, relationships, and connections across a large document corpus that would have required significantly more manual effort to analyze by traditional means.
The investigation used AI to extract entities (names, companies, financial transfers) and build a graph of relationships. The output was then human-verified through on-the-ground reporting, source interviews, and cross-referencing against public records. No AI-generated claim was published without independent confirmation.
Investigative journalism has a processing bottleneck
Leaked document analysis is the classic constraint in political and financial reporting. A single tranche of emails, contracts, or bank records can involve thousands of documents. Manual reading, cross-referencing, and pattern spotting consume months and staff capacity that most newsrooms do not have.
AI does not replace judgment or verification. It compresses the first phase: feature extraction and anomaly detection. A reporter can now spend 80% of time on verification, source cultivation, and legal vetting instead of 40% of time on document triage. The risk is obvious: AI can hallucinate relationships or miss context that a human reader would catch. That is why the FT's workflow kept humans in the loop for every claim.
This matters because it signals a plausible path for resource-constrained news organizations to take on more ambitious investigations without hiring proportionally. It also raises an operational question: if AI speeds up document processing, does competitive pressure force newsrooms to adopt it, or do they wait for consensus on best practices for human oversight?
How to use AI for document-heavy investigations
Start with the constraint you actually have. If your bottleneck is reading 50,000 documents, use AI for entity extraction and clustering. If your bottleneck is source verification and legal review, do not use AI for that; do use it upstream. Train your AI model on a small, human-verified subset of documents so you can tune precision and recall for your specific domain (corruption allegations, contract analysis, regulatory filings). Run the AI output through at least two independent human reviewers before any reporting is filed. Use AI as a compass that points toward documents and relationships worth investigating, not as a source of truth. Document the model version, the documents it was trained on, and the confidence scores for each claim, in case you need to explain your methodology to legal counsel or an editor who will defend your story.