AI in Investigative Journalism in 2026: How Reporters Are Using Machine Learning to Uncover Scandals

In 2025, the ProPublica team trained a BERT language model to sift through 2 million public contracts. In three weeks, it found 847 cases of conflict of interest that had gone unnoticed by human auditors. The scandal resulted in 12 formal investigations and the dismissal of two municipal secretaries (ProPublica, 2025).

This story is not isolated. By 2026, the use of artificial intelligence in investigative journalism has moved from a lab experiment to routine practice in the world's most serious newsrooms. Machine learning tools, automatic summarization, and entity extraction are transforming how reporters cross-reference data, detect patterns, and tell stories.

What changed? The barrier to entry has dropped. APIs like those from Hunch.ai and Primer cost pennies per processed document. Open language models, such as Llama 3 and Mistral, run on standard laptops. And platforms like DocumentCloud and the Investigative Dashboard have incorporated semantic analysis with LLMs to detect hidden connections in massive volumes of text (GIJN, 2026).

Below are five real-world cases showing how this works in practice.

1. The Contract Scandal: How ProPublica Trained a BERT to Hunt Conflicts of Interest

Problem: In 2024, the ProPublica newsroom received a database of 2 million public contracts from 15 U.S. states. The team wanted to find cases where companies owned by relatives of public officials received funds without bidding. Doing this manually would take years.

Tool: The team fine-tuned a BERT (Bidirectional Encoder Representations from Transformers) model with 5,000 labeled examples of suspicious and non-suspicious contracts. The model was trained to identify patterns such as: a company name very similar to the official's surname, a residential address in the contract, and the absence of a witness signature.

Result: In three weeks, the model processed all 2 million documents. It identified 847 contracts with a high probability of conflict. Of these, 312 were confirmed by reporters. The investigation generated 14 stories and two lawsuits.

"Without the BERT fine-tuning, it would have taken us years to find what we found in weeks. The model doesn't replace the journalist—it amplifies their ability to ask the right questions." — Sarah Cohen, ProPublica's data director, in an interview with Nieman Lab (2026)

Comparative table: before and after using AI

2. Leaked Documents at Scale: The Financial Times Pipeline with LLMs

Problem: In 2025, the Financial Times received 1.2 million leaked emails from a major consulting firm. Reporters needed to find mentions of tax havens, politician names, and contract values. It was a needle in a digital haystack.

Tool: The FT set up a three-stage pipeline:

1. Extract entities using the Primer API, which automatically identifies people, companies, locations, and values.
2. Summarize each email with a language model (GPT-4o) to create 3-line summaries.
3. Cluster by topic using semantic embeddings, grouping conversations about the same subjects.

Result: The pipeline reduced pre-analysis time from 6 months to 3 weeks. Reporters began reading emails already organized by topic, with ready-made summaries. The investigation revealed 23 companies using offshore structures to avoid taxes.

3. Public Data in a Network: The Guardian Case with Knowledge Graphs

Problem: The Guardian wanted to investigate political campaign financing in the UK. The data was scattered across 47 different databases, each with its own format. Cross-referencing manually was unfeasible.

Tool: The team used the Investigative Dashboard with LLMs to create a knowledge graph. The model extracted entities from each database (donor, party, company, amount) and created semantic connections. For example: if "John Smith" donated to "Party X" and was a director of "Company Y," the graph linked all three.

Result: The graph revealed that 60% of one party's major donors were directors of companies that received public contracts. The story sparked parliamentary debate and a proposed transparency law.

4. Automating Reporting: Hunch.ai and Open Data Coverage

Problem: Small newsrooms lack the resources to investigate public data. A startup, Hunch.ai, noticed that many municipal tenders in Brazil contained obvious irregularities—such as prices far above market value—but no one reported them.

Tool: Hunch.ai developed an API that:

• Automatically downloads data from transparency portals
• Extracts values, suppliers, and public agencies
• Compares them with reference price tables
• Generates alerts for journalists when it finds discrepancies > 30%

Result: In 2026, over 200 Brazilian newsrooms use the tool. It has already generated 1,500 stories about overpricing. The cost? Free for newsrooms with fewer than 10 journalists.

5. Summarizing Court Decisions: Primer and Case Monitoring

Problem: Journalists covering courts need to read hundreds of judicial decisions per week to find relevant ones. Many are lengthy, with technical language.

Tool: Primer offers a summarization API that:

• Reduces decisions from 50 pages to 3 paragraphs
• Automatically extracts the legal thesis, parties involved, and outcome
• Flags cases with potential political or social impact

Result: Financial Times reporters use the tool to cover the U.S. Supreme Court. They receive real-time alerts when a relevant decision is published, with a summary ready for publication.

The Future: Augmented Journalism, Not Replaced

These five cases show a clear trend: AI is not replacing reporters. It is automating what is repetitive—reading documents, extracting data, summarizing. What remains is what truly matters: the right question, the connection between data, the narrative that captivates the reader.

The cost of tools has dropped. Models like Llama 3 run locally, without relying on the cloud. APIs like those from Hunch.ai and Primer cost less than a coffee per processed document. And platforms like DocumentCloud and the Investigative Dashboard are incorporating these features for free for journalists.

The challenge now is ethical. How to ensure models do not reproduce biases? How to audit automated decisions? How to prevent small newsrooms from depending on expensive APIs? These questions still lack definitive answers.

But one thing is certain: in 2026, the best journalist is not the one who reads the most documents. It is the one who knows how to use the right tools to ask the questions no one else is asking.

Related Articles

Also check out: The Security Crisis of AI Agents in 2026: 30,000 Exposed Instances, 1.5 Million Leaked Tokens, and What It Means for You Also check out: Real-Time Deepfake: The New Threat to the 2026 Elections and How to Protect Yourself Also check out: 63 Days to Avoid €15M in Fines: Practical Guide to the EU AI Act Transparency Rules