AI Applications in Investigative Journalism

Investigative journalism grounds technological practices like computational journalism within its core watchdog function and normative mission.⁹ In fact, investigative journalism has historically been at the forefront of adopting computational methods. Early computer-assisted reporting (CAR) in the 1960s used social science methods, spreadsheets and statistical analysis for long-term projects,¹⁰ practices that evolved into modern AI-driven techniques, such as machine learning and large-scale data processing.¹¹ As information increasingly shifted to digital formats, investigative journalists turned to computational tools to manage growing volumes of data, exemplified by large-scale projects like the Panama Papers. That project took more than a year, involved millions of documents and required techniques such as optical character recognition (OCR) and algorithmic sorting, as well as large-scale indexing and search capabilities that would otherwise be impossible for human teams to perform.¹²

These technology-based approaches enabled new pathways to story discovery through pattern recognition, anomaly detection¹³ and large-scale data linkage,¹⁴ while also expanding verification practices through methods like open-source intelligence (OSINT), including satellite analysis and digital forensics.¹⁵ Computational journalism in this context applies “computational thinking,” meaning the abstraction of investigative tasks into granular, computable elements that can be processed algorithmically.¹⁶

Normative goals — namely accountability, exposure and reform — shape the role of AI systems in investigative journalism, enhancing the watchdog function and efficiency. In this sense, AI tools and data-driven methods are not ends in themselves but tools that support journalism’s public-interest mission by enabling the analysis of large and complex datasets, uncovering hidden patterns and strengthening transparency. 

At the front end, AI tools are particularly valuable for establishing the scale of problems and “finding needles in haystacks.”¹⁷ Two studies surveying the uses of AI in investigative journalism¹⁸ show that machine learning (ML) is used to reduce massive datasets into actionable leads through data cleaning, record linkage and text-as-data techniques such as sentiment analysis, entity extraction, topic modeling and similarity detection (e.g., Locality-Sensitive Hashing). Another practice-based study with eight investigative journalists in Germany¹⁹ showed that journalists see particular value in tools that could automate repetitive web-based research tasks, such as monitoring multiple websites, scraping unstructured local-government records, tracking changes in online content over time, extracting information from social media or leaked materials, and converting collected material into searchable or analyzable formats.

AI tools play a central role in algorithmic accountability and investigative work in constrained environments. While this briefing does not include a comprehensive look at investigative journalism about AI systems, it does include cases where journalists use AI tools to investigate opaque platform systems. In a practice-based study conducted in a Dutch context²⁰ they describe using AI tools to audit and reverse-engineer opaque platform systems as “fighting fire with fire.” In collaboration with De Groene Amsterdammer, these researchers created sock-puppet accounts and used a vision-language model to simulate user behavior on TikTok, demonstrating that the platform could infer user interests within seconds and begin amplifying eating-disorder content within minutes. In a separate audit of a Google dataset, the same researchers used the ML tool Homepage2vec to classify 2.5 million Dutch-language websites included in large language model training data and uncovered significant proportions of copyrighted journalism, conspiracy content and even leaked personal data. Complementing these efforts, ProPublica’s “Machine Bias” investigation revealed systemic racial disparities in algorithms used in the criminal justice system. 

One of the most significant limitations of AI tools in investigative journalism is the lack of accessible, reliable and structured data, foundational to both AI systems and investigative work. As discussed, investigative journalism often depends on data that are hidden, incomplete, messy or intentionally withheld, which requires journalists to “request, negotiate, scrape, or purchase” datasets before their analysis can even begin.²⁵ Even when data is available, it can be messy (e.g., spread across multiple sources, poorly formatted, inconsistent over time, lacking metadata), further complicating automation.²⁶ These constraints reflect a broader “infrastructural gap,” in which investigative capacity is systematically limited by factors such as proprietary roadblocks and the absence of accessible public databases. Because AI systems rely heavily on consistently structured data, the risk of inaccurate or misleading outputs increases, making full automation not only unreliable but also impossible.²⁷ 

AI systems introduce concerns about the erosion of journalistic autonomy and expertise. Journalists across all subfields emphasize that their “irreplaceable human qualities” (e.g., judgment, fieldwork, interviewing, source-building) remain central to their work despite increasing automation of other areas of work.²⁸ At the same time, reliance on external technology companies, nongovernmental organizations and universities for AI tools and infrastructure can create newsroom dependencies and raise concerns about influence, information security, copyright, privacy and data protection.²⁹ 

A comparable dynamic already exists in non-AI investigative practices. For example, during the U.S.-Israel war with Iran, news organizations relied heavily on commercial satellite imagery to verify strikes and assess damage in Iran, but access was restricted when Planet Labs limited imagery distribution at the request of the U.S. government, citing operational security concerns. This shift to a managed, case-by-case release system sharply reduced journalists’ ability to independently document events and diminished the transparency previously enabled by open commercial imagery, illustrating that newsroom capabilities depend on the policies and geopolitical constraints of external data providers. 

This increased dependence on external actors is part of a broader structural transformation in journalism. The “platformization” of news — understood as the rise of platforms as the dominant infrastructural and economic model of the social web³⁰ — applies to the entire news ecosystem. As news organizations increasingly depend on third-party platforms for production tools, distribution channels and revenue streams, their autonomy is affected by systems they do not control.³¹ More broadly, studies stress that “technology in journalism is not neutral,”³² as its effects depend on institutional contexts and power relations, meaning AI tools can reshape — not just support — journalistic practice. 

Despite the promise of natural-language tool interactions,³³ implementing AI tools in investigative journalism requires substantial technical expertise, along with collaboration across disciplines and roles. AI tools are sometimes complex and difficult to integrate into newsroom workflows, especially when they require customization or advanced technical knowledge. In an interview study with 25 national, regional and local Dutch news media organizations,³⁴ the authors identified a significant “language barrier” between investigative journalists and IT professionals that serves as a major impediment to AI adoption. This barrier manifests in several ways. First, there is a lack of shared vocabulary. Some journalists find it difficult to articulate their specific editorial needs to technical experts because they lack a foundational vocabulary for AI techniques and applications. Second, interviewees noted that IT professionals “look through different glasses” and often assume a level of technical knowledge that journalists do not have, making communication difficult. Third, there is a disconnect between the “work sprints” of IT departments, which may plan activities months in advance, and the fast-paced or unpredictable deadlines of investigative news desks. Fourth, IT professionals are often based outside the newsroom, which further limits the opportunity for the spontaneous, interdisciplinary collaboration needed to co-create AI tools. 

To overcome this challenge, research offers a few recommendations. One is to involve “boundary spanners”³⁵ — individuals who operate at the intersection of journalism and technology and can translate between the two groups. These boundary spanners need not be additional staff; existing personnel can take on these roles and commit to approaching technology from this perspective by bringing together colleagues with distinct expertise. Another recommendation is to take a collaborative approach requiring active participation from all stakeholders, including, at times, those outside journalism.³⁶ 

In addition to communication barriers between journalists and IT professionals, AI adoption is shaped by other organizational and professional dynamics. Sometimes AI uptake is driven by enthusiasm for new technologies without strategy, which some journalists and scholars call “Shiny Things Syndrome.”³⁷ As shown in an interview-based study in the Dutch news media context,³⁸ AI discussions sometimes remain at the managerial level and do not translate to everyday newsroom workflows. As a result, many investigative journalists position themselves as “reluctant adopters,” particularly when faced with technical language barriers and the need to collaborate with specialized teams such as data scientists or IT professionals. These professional dynamics are not specific to AI adoption; across a range of methods, earlier research on technology adoption in journalism also shows widespread curiosity about new technologies alongside resistance to top-down technology mandates and concerns about dependency on third-party companies.³⁹

These challenges are closely tied to the broader issue of insufficient AI and data literacy in journalism. Journalists struggle with limited access to reliable learning resources, difficulty identifying appropriate expertise, general reluctance and the challenge of keeping pace with rapidly evolving technologies.⁴⁰ In response, scholars have proposed frameworks such as the Accuracy-Fairness-Transparency (AFT) framework, which emerged from a synthesis of prior research on AI systems in journalism. AFT applies core journalistic ethics (accuracy, fairness and transparency) to ensure high-quality, trustworthy data in AI-driven journalism. AFT can improve data practices and serve as a pathway for building AI literacy and embedding journalistic values into technological workflows.⁴¹ 

Additionally, the opacity of AI tools limits journalists’ understanding of their outputs and how and why they are processed.⁴² This, in turn, complicates accountability and the epistemic traceability of evidence, especially in the context of generative AI.⁴³ Scholars distinguish between explainability — the ability to explain what a model does and what it produces — and interpretability, which refers to understanding the specific mechanics of how the model reached that result.⁴⁴ Because of these concerns, many investigative teams purposefully opt for “decision tree” or “random forest” algorithms over more powerful neural network or deep learning approaches because they prefer tools that produce human-readable partitions of data that allow for better scrutiny.⁴⁵ In the specific context of generative AI, the traceability of evidence is further undermined by hallucinations, where models predict plausible but factually incorrect data points. This lack of reliability has led some newsrooms to abandon high-performing AI anomaly detection tools.⁴⁶

Finally, many tools remain project-specific, limiting their reuse and reducing long-term efficiency gains. These constraints mean that AI adoption is not simply a matter of availability but of sustained organizational capacity, training and collaboration.⁴⁷

AI systems in investigative journalism create another type of language barrier, particularly in non-English and Global South contexts, because the underlying technologies are primarily developed, trained and optimized for English. This results in clear technical performance gaps: while English-based models are highly-developed, even widely used languages like French remain comparatively underdeveloped and “clumsy,”⁴⁸ and AI tools often fail to function effectively in many non-Anglophone languages.⁴⁹ As a result, journalists (including non-investigative journalists) in these contexts face limitations in using AI tools for core tasks like data mining, transcription and automated analysis.⁵⁰ These constraints are compounded by a broader “Eurocentric” information lens, where AI systems — shaped by Western design and data — struggle to interpret local cultural nuances and are viewed by some as tools that extract information for Western purposes rather than supporting local knowledge production. Because they are not trained on localized linguistic and cultural data, such systems remain fundamentally exclusionary in their ability to represent non-Western perspectives.⁵¹

These dynamics extend into professional and educational domains, where English-centric AI systems reinforce existing hierarchies and forms of exclusion. An interview and focus group study of 34 investigative journalists across five East African countries shows that in some newsrooms, competence is tied to English proficiency rather than investigative skill, leading to the loss of talented journalists who are unable to communicate their work effectively in English. AI tools can intensify this divide by amplifying the capabilities of those already fluent in dominant languages and technologies, while further marginalizing those without these skills.⁵² A related mixed-methods study of journalism education across four South African universities found similar resistance to AI tools that lack support for Indigenous language contexts, prompting calls to “decolonize” language models by training them on local languages and contexts. Without such efforts, investigative AI tools risk continuing to privilege English-speaking environments and limiting locally grounded journalism.⁵³ For example, Latam-GPT, developed by Chile’s National Center for Artificial Intelligence (CENIA), aims to build an open-source AI model trained on Latin American languages and contexts (see this working group’s November 2025 briefing on translation and transcription for further discussion of linguistic strengths and limitations). 

From an economic perspective, investigative journalism faces some disadvantages that limit the broader use of AI. For one, investigative reporting relies heavily on human effort. Unlike advertising or other automated communication fields, it is much harder to use technology to speed up the deep, manual work required for investigations. Furthermore, while the social importance of investigative journalism remains constant, its production costs increase relative to other sectors, often branding it as “costly” by comparison.⁵⁴ This creates a condition in which AI tools accelerate and facilitate some tasks, such as data processing, but fail to reduce the overall labor. Moreover, the limited productivity gains from AI systems in investigative journalism, relative to the significant gains in sectors like advertising and public relations, create a cost disadvantage.⁵⁵ This “cost disease” makes it economically difficult for newsrooms to fund competitive wages for highly skilled human labor, increasing the risk of talent drain as journalists move to better-paying communication sectors with overlapping skill sets.⁵⁶ 

As discussed earlier, these constraints are compounded by the project-specific nature of AI systems in investigative journalism.⁵⁷ As a result, AI adoption tends to be viable only for a narrow subset of investigations where data is central, automation is technically feasible and simpler computational methods or traditional reporting cannot achieve comparable results. These financial barriers are even more pronounced for smaller and local newsrooms, which are less able to absorb the high costs of AI development or even to customize off-the-shelf tools — sometimes paired with infrastructural issues, such as unreliable internet connectivity⁵⁸ — and therefore face greater barriers to adoption, reinforcing existing inequalities in the distribution of AI capabilities across the industry.⁵⁹