Leveraging AI and Technology to Investigate Power

Editor’s Note: This is the fifth excerpt taken from GIJN’s in-depth report on “The Investigative Agenda for Technology and AI Journalism,” based on a day-long pre-conference event held on November 20 at GIJC25, where 100 investigative journalists, editors, tech experts, and researchers from nearly 50 countries and territories convened to examine the most urgent technology-related challenges and opportunities facing investigative journalism today. Credits and acknowledgments for this project can be found here.

Across the GIJC25 pre-conference day and in the survey that followed, one message emerged clearly: investigative journalists are not only concerned about artificial intelligence, but they are actively asking for support on how to use it.

Requests for guidance on integrating AI and advanced technologies into investigative work appeared as frequently, and in some cases more frequently, than calls for AI accountability reporting itself. For many participants, AI is no longer a distant or abstract topic. It is becoming part of everyday newsroom workflows — sometimes by choice, sometimes by institutional pressure — and journalists are seeking ways to use it effectively, responsibly, and without undermining core investigative standards.

Responding to these changes requires examining how AI and related technologies are already being used to investigate power, where they genuinely add value, where they introduce new risks, and why collective infrastructure, shared standards, and capacity-building are now essential. It also outlines the specific role GIJN can play in helping reduce inequalities of access and knowledge across the global investigative journalism community.

How to Integrate AI and Advanced Technologies Into Investigative Work

Survey responses following the pre-conference show strong convergence around several priorities:

• Using AI in newsrooms to gain efficiency and expand investigative capacity
• Access to practical tools, workflows, and shared resources
• Responsible use of AI in investigative journalism
• Building and running local or independent models
• Low-code and no-code tools for journalists without technical backgrounds

Respondents asked for practical guidance: how to analyze large document troves, connect disparate datasets, detect patterns, verify claims, map networks, or triage massive leaks without exposing sensitive data to third-party platforms.

AI as Investigative Tool: Where It Works

Speakers from organizations such as ICIJ and OCCRP presented concrete examples of how machine learning (in which computer models are trained on a “teaching set” of data to identify patterns, insights, or predictions at a scale and speed beyond human capacity) and AI-assisted tools have already been used in large-scale investigations.

As early as 2019, GIJN documented both the possibilities and limits of these techniques, highlighting that their value lies in clearly defined tasks with strong human oversight.

The work of the International Consortium of Investigative Journalists illustrates the investigative potential of machine learning. In investigations such as the 2018 Implant Files, ICIJ used machine learning to analyze more than eight million health records on adverse events caused by medical devices and obtained through freedom of information requests. Because deaths were often described indirectly in the narrative reports (using phrases such as “the patient expired” rather than the word “death”) machine learning algorithms were first trained to detect cases where a patient had likely died based on patterns in the language. Once the cases were identified, the findings were compared against the original classification of the event provided by the manufacturer, detecting cases in which the deaths were originally classified as a device malfunction or injury. Because deaths were often described indirectly (for example as “the patient expired”) algorithms were trained to identify cases where a fatality had likely occurred but was misclassified as a device malfunction or injury. This approach led ICIJ to identify over 2,100 previously obscured patient deaths. The results obtained through the machine learning process, as well as all the underreported deaths identified through the analysis, were fact-checked by a team of journalists.

Recent advances in AI have significantly lowered the barrier to entry for investigative journalists. Tools that once required dedicated data teams or custom development are now more accessible, easier to deploy, and increasingly integrated into newsroom workflows. As a result, machine learning and AI are proving most relevant and effective for investigative journalism today in the following cases:

• Navigating massive leaks and document troves, including through retrieval-augmented generation (RAG) systems that allow journalists to query large, curated document collections and use that information to respond.
• Detecting hidden patterns and misclassification, especially in large administrative or corporate datasets where harm or wrongdoing is obscured by inconsistent language or categorization.
• Extracting data from unstructured documents, such as pulling tables, names, or fields from PDFs and scanned records, when tasks are narrowly defined and outputs are systematically verified (for example police reports or court documents).
• Organizing investigations at scale, for example by automatically building timelines, chronologies, or entity lists from court records, contracts.
• Identifying specific visual elements within large datasets, using models trained to locate items such as passports or identity documents in leak tranches, with journalists reviewing and discarding false positives.
• Satellite and environmental investigations, where models trained to detect deforestation, illegal mining, or clandestine airstrips help reporters filter vast imagery archives to a manageable set of high-risk areas.

These uses share a common feature: AI works best when narrowly scoped, problem-driven, and embedded in a human-led investigative process, rather than treated as a universal solution.

ICIJ used machine learning to help analyze large amounts of data on medical device issues, as part of its 2018 Implant Files investigation. Image: Screenshot, ICIJ

Limits, Pitfalls, and the Cost of Getting AI Assistance Wrong

Despite these opportunities, speakers consistently warned against AI over-reliance and hype. Several recurring limitations were highlighted during the pre-conference day and in survey responses. These included:

• Hallucinations and false confidence: As Helena Bengtsson noted, the more complex the task, the more likely models are to “guess.” AI tools may produce fluent but incorrect outputs, invent links or entities, or silently distort repeated data. This makes manual verification non-negotiable: a human always needs to remain in the loop.
• A prompt is not a methodology: Multiple speakers emphasized that AI outputs are not inherently reproducible. Two journalists using the same prompt may receive different results.
• Legal, security, and privacy risks: Uploading sensitive investigative data to commercial platforms can expose journalists and sources to legal and security risks. Concerns around personal data protection, copyright, and data retention were repeatedly raised, alongside fears of surveillance metadata.
• Cost and capacity divides: Running secure local infrastructure, training models, or maintaining specialized tools requires financial and technical resources that many newsrooms do not have. As Jelena Cosic from ICIJ stressed, even well-resourced organizations struggle to sustain tool development in competition with Big Tech.

Paradox in Practice: Investigating Tech While Using It

A central ethical tension runs through contemporary investigative work: journalists are increasingly investigating powerful technology companies and state systems while utilizing tools produced by those same actors.

This creates multiple contradictions:

• Depending on the platforms whose business models and power structures journalists are investigating.
• Relying on closed or opaque tools trained on unknown, biased, or extractive datasets.
• Indirectly reinforcing harmful human and environmental impacts, from data-labeling labor in the Global South to the excessive water footprint of AI infrastructure.
• Risking source exposure and investigative leaks, particularly when sensitive material is processed through commercial or cloud-based systems.

Speakers did not frame this paradox as a reason to disengage from technology, but rather as a call for radical transparency and editorial discipline. Publishing methodologies, explaining tool choices, documenting limitations, and distinguishing AI-assisted analysis from verified findings were cited as essential to maintaining credibility.

One approach discussed draws on practices developed by the Pulitzer Center, which encourages newsrooms to explicitly assess whether AI tools are used internally (to support reporting workflows) or externally (to generate outputs visible to audiences), and to evaluate risks accordingly. This internal/external distinction offers a practical framework for aligning tool use with journalistic ethics, accountability, and source protection.

Collaboration and Capacity Building as a Necessity

One of the strongest conclusions from both the pre-conference day discussions and the follow-up survey is that no single newsroom can solve these challenges alone.

Participants emphasized the need for:

• Cross-border collaboration on tools, workflows, and standards.
• Ready-to-use tools and practical guides for investigations.
• Partnerships with universities, technologists, and civil society.
• Open-source and independent alternatives to reduce Big Tech dependency.
• Training pathways for non-technical journalists.

Without collective solutions, AI risks deepening existing inequalities between well-funded investigative centers and smaller or exiled newsrooms.

GIJN’s Role: Reducing Inequalities of Access

In this context, the Global Investigative Journalism Network is uniquely positioned to act as a connector, curator, and capacity-builder.

Survey respondents explicitly called on GIJN to:

• Provide practical guides to AI-driven investigative tools.
• Develop step-by-step resources on building local or task-specific models.
• Curate trusted toolkits and low-code solutions accessible to non-technical reporters.
• Facilitate knowledge-sharing across regions and languages.
• Support ethical frameworks grounded in investigative practice.

Rather than promoting a single technological path, GIJN’s value lies in helping journalists make informed choices: when AI adds value, when it does not, and how to integrate it without compromising security, legality, or editorial standards.

Priorities identified: 

• Practical guidance on using AI as an investigative tool.
• Clear frameworks for responsible and transparent use.
• Shared, open, and secure tools, scripts, and workflows.
• Support for local and independent models where possible.
• Training that makes AI accessible beyond technical specialists.
• Collective strategies to reduce dependency on Big Tech.
• GIJN as a hub for capacity-building, coordination, and equity.

Sandrine Rigaud is the program director of GIJN. She is an investigative journalist, director, and Emmy-winning producer who served as editor-in-chief of Forbidden Stories from 2019 to 2024. In that position, she led international collaborations to continue the work of assassinated or under threat reporters, coordinating investigations involving up to 100 journalists and 30 media outlets, including Le Monde, The Washington Post, The Guardian, Der Spiegel, Haaretz, and El País. She teaches investigative journalism at the School of Journalism of Sciences Po Paris and is co-author of “Pegasus: How a Spy in Your Pocket Threatens the End of Privacy, Dignity, and Democracy.” A Nieman Fellow at Harvard in 2024/2025, she worked on global investigative collaborations, leaked data management, and Artificial Intelligence.