Artificial intelligence tools have entered this environment at a moment of genuine need. The risk is that they arrive dressed as a solution to the wrong problem — that organisations adopt them expecting automation of the core editorial task of verification, and find them wanting. This article argues for a different relationship: AI as infrastructure for analysis, not as a replacement for judgment. Through the innovative work of Spanish fact-checker Maldita.es, we will show that rather than telling us what is true, these tools help us see what is happening at a scale and speed that would otherwise be invisible.

A Problem That Outgrew Human Capacity

Information integrity has a precise meaning: It’s the degree to which information is accurate, traceable, and presented in its appropriate context. All properties have come under systematic pressure from the digital environment, and the mechanisms are by now familiar enough to deserve naming directly.

The first mechanism is platform ranking, which rewards emotional resonance and shareability This can even overturn more objective metrics, such as temporal recency or belonging to one’s own network. The result is a structural bias in what gets views: not the most reliable content, but the most engaging.

The second mechanism is the ease with which content is detached from its original source, date, or context, making source attribution increasingly difficult. A video filmed in one country is labelled as coming from another; a statement made years ago is recirculated as if it were today's news. The problem is not human error or bad faith in individual cases, but a format that systematically severs content from the metadata needed to evaluate it.

The third pressure mechanism is scale. Generative tools and coordinated distribution infrastructures have radically reduced the cost of producing and spreading large volumes of content. What once required significant financial and organisational resources (such as a coordinated information operation, a network of synthetic accounts seeding a narrative) can now be attempted at low cost and with a trial-and-error approach that makes detection harder. Just as social media reduced the cost for sharing and distributing content ; generative AI is now enabling the cheap creation of high-quality, personalised content production across all media formats. Informational flooding is easier than ever, and malicious actors are betting on volume rather than precision.

The consequence for monitoring organisations is that the pace of content generation has outrun the capacity of human teams to track it in real time. A team of analysts, however skilled, cannot process thousands of posts per hour, detect emerging patterns across platforms, or maintain continuous surveillance of a rapidly evolving information landscape.

This is not a question of resources that more hiring could solve. It is a structural mismatch that requires structural tools.

The challenge dimension was obvious during the 2024 floods that happened in Spain and killed 238. The level of misleading content during the aftermath is almost unmatched, stemming from involuntary misunderstandings, partisan interest and coordinated campaigns. During this episode, Maldita.es received 4 times more fact-check requests than usual: 13.000 in 7 days. Thanks to our tool for AI-powered monitoring, we could enabled us to group those messages into underlying narratives (e.g. the floodings were the result of a plan from the elites, authorities were hiding the real number of casualties, etc). This allowed us to understand the informational environment and populate it with empirical data, a crucial step before drafting policy reports on platform regulation like this one.

What AI Tools Can and Cannot Do

The most persistent and damaging misconception about AI tools in this space is that they can determine whether something is true. They cannot. Not reliably, not at scale, and not in a way that could substitute for editorial judgment. At least not yet. Getting this boundary wrong leads to poor tool design, misplaced trust, and ultimately to outputs that erode rather than support credibility.

There is a fundamental distinction that tends to get collapsed in public discussion: the difference between detecting synthetic or inauthentic content and detecting inaccurate content. These are different technical problems that require different solutions.

• Detecting a synthetically generated image or a network of coordinated inauthentic accounts is, at its core, a pattern-recognition problem that can be approached with machine learning tools designed for that purpose, such as watermarking, fingerprinting, behavioural analysis.

• Detecting a factual claim is an epistemic problem: it requires access to evidence, the ability to evaluate sources, and the exercise of judgment about what counts as reliable. Large language models (LLMs) are not built for this, and their tendency toward fluent confabulation makes them actively dangerous when deployed as fact-checkers.

Where LLMs and related AI tools do reliably deliver value is in tasks that do not require external knowledge — tasks that are fundamentally about transformation rather than verification, such as classification into a set of topics, labelling of the wording, clustering, or summarisation.: these are operations that organise and reframe existing information rather than adjudicating its truth. For example, Maldita.es participates in the European project Prebunking At Scale by using Large Language Models (LLMs) to process thousands of claims in short-form videos sourced from TikTok, YouTtube and Instagram. They are categorised into five categories: health, conflict, European Union, migration and climate. An AI tool that takes a stream of ten thousand social media posts and labels them is doing something genuinely useful: it is turning noise into a structured dataset that analysts can work with. It has not verified anything. It has made verification easier.

The practical implication is clear: AI tools in this domain show their best potential as analytical assistants (not leaders) that transform and organise data. Anything beyond that should be treated with caution, whether it appears in a product pitch or a newsroom workflow.

Building a Monitoring Pipeline: Data, Memory, and Noise

The practical challenge of AI-supported monitoring is architectural before it is analytical. Before a tool can surface patterns or group narratives, it needs a reliable flow of data, a structure for storing and linking what it finds, and a strategy for managing the noise that is inherent in any automated system at scale.

1) Data sources come with meaningful trade-offs that shape everything downstream. Platform APIs provide high volumes of content but at a cost: the signal-to-noise ratio is low because not all content is relevant, plus the most sensitive content is often filtered out by platforms before it reaches researchers, and access terms change unpredictably. Additionally, official APIs tend to be expensive for user crowdsourcing (tiplines, reporting tools, WhatsApp channels) produces higher-quality, more contextually grounded inputs, but coverage is limited and dependent on community engagement. Expert curation provides the most reliable and contextually rich data, but it is resource-intensive and cannot scale to cover broad information environments in real time. An example of an expert-sourced claim database developed by Maldita.es is Iberifier. With an Iberoamerican lens, this digital infrastructure allows fact-checking organizations from different countries to share and connect their findings. Most monitoring systems use some combination of all three, with the choice of mix shaping the tool's strengths and blind spots.

2) Storage and memory are underappreciated components of effective monitoring. A database that preserves previously analysed content does more than archive it: it makes the analysis cumulative. When a new claim arrives, it can be matched against existing records — not just to avoid duplicating work, but to measure recurrence, track virality over time, and build a longitudinal view of how a particular narrative has evolved. This idea is central to the Whatsapp tipline that structures Maldita’s informational monitoring. The metadata attached to each item (for example, platform of origin, date, linked accounts, prior appearances)_ becomes the raw material for investigation. Without this institutional memory, monitoring is episodic rather than structural: you can see what is trending today, but not how it connects to what was trending six months ago.

3) Importantly, no monitoring pipeline eliminates noise. This is a structural feature, not a fixable bug, particularly when the data source is platform APIs or when some labelling is automated. The mitigations available are partial but effective: human-in-the-loop workflows that route flagged content through editorial review before it influences outputs or external knowledge bases, (such as Wikidata) that help the system distinguish between homonyms, resolve ambiguous references, and avoid conflating distinct actors or events. The goal is not a noise-free system — it is a system whose noise is managed well enough that the signal remains usable.

From Isolated Claims To Underlying Narratives

Individual false or misleading claims are symptoms. The underlying disease is the narrative — the broader ideological frame within which individual claims acquire meaning, and which gives disinformation its durability and its persuasive force.

At first, Maldita.es spotted isolated claims echoing an article about ice size growth in the Antarctic ice shelf from 2009 to 2019. But threats to information integrity rarely operate through single discrete pieces. More commonly, they rely on a chain of apparently factual statements that together construct a worldview: a particular understanding of who is responsible for a problem, who benefits from a given policy, what historical events really meant. Later that year, our newsroom saw that posts started to emerge, claiming that Al Gore failed at predicting Ice Cap’s melting by 2014. Each individual claim in such a chain might be contested, exaggerated, or simply out of context, but the cumulative effect on a reader's beliefs can be substantial. A monitoring system that only processes individual claims in isolation (catching false statements as they appear) misses this deeper structure. It is treating the symptoms without seeing the disease.

Combining content databases with large language models offers a way to work at the level of narratives rather than individual claims. The core capability is grouping: given a large set of claims about a topic, an LLM can identify which ones invoke the same underlying frame, even when the specific wording, actors, or contexts differ. This overcomes a fundamental limitation of human analysis: no team of analysts can hold thousands of claims simultaneously in working memory, track the connections between them, and identify the pattern that ties them together across time and geography. Computational clustering and AI-assisted labelling extend what is cognitively possible. This technology allowed us to spot and dismantle a narrative that questioned the ongoing climate crisis through ice-cap-size related claims.

Identifying recurring narratives also enables a form of pattern detection that has investigative value beyond any single article: when the same framing appears in multiple geographies in quick succession, it is often a signal worth investigating. Narratives do not spread randomly — they are often seeded by actors who have reasons to promote a particular frame. Cross-national recurrence, when detected systematically rather than anecdotally, becomes a thread to pull. To follow this type of threads, Maldita.es coordinates a claim database funded by National Endowment for Democracy where organizations from Latin America and Eastern Europe anticipate and track Russian information operations.

Turning Monitoring Into Action

Analysis that stays inside a database changes nothing. The final and arguably most important design challenge in AI-supported monitoring is connecting the analytical pipeline to outputs that organisations can actually use, and that means thinking carefully about who those organisations are and what they need.

Different end users have different requirements. Researchers often want raw, exportable data, usually CSV files with full metadata, filterable by date, platform, and topic, that they can bring into their own analytical workflows. Newsrooms typically want something more visual and navigable, for example dashboards that surface emerging trends, configurable alerts that flag spikes in particular topics, interfaces that let journalists explore the data without needing to write queries. Civil society organisations working on public communication may need something different again. A monitoring tool designed without its users in mind will not be used, regardless of the quality of its underlying analysis.

Tools built around live, growing datasets also require ongoing maintenance that one-time analysis projects do not. Unlike an investigation of a document leak, which has a fixed corpus and a defined endpoint, a monitoring system operates on a stream that never closes. Topic definitions need to be recalibrated as the information environment evolves. AI processing pipelines need to be monitored for drift and adjusted as language patterns shift. Resource consumption needs to be managed, since continuous AI-assisted analysis at scale has real costs. These are not problems to be solved at launch — they are ongoing operational responsibilities.

The most durable value of a well-designed monitoring system is not any single output: not a particular investigation, trend report, or counter-messaging resource. It is the institutional memory and workflow integration it enables over time. Organisations that have embedded monitoring tools into their editorial processes, which have built their daily workflows around the intelligence these systems provide, respond faster to emerging threats, with more evidence, and with a richer understanding of how the narratives they are tracking have evolved. The cumulative knowledge that accrues in a well-maintained system becomes, in time, a competitive advantage in the broader information environment.

Your Feedback Matters

What did you think of this text? Take 30 seconds to share your feedback and help us create meaningful content for civil society!

This piece of resources has been created as part of the AI for Social Change project within TechSoup's Digital Activism Program, with support from Google.org.

AI tools are evolving rapidly, and while we do our best to ensure the validity of the content we provide, sometimes some elements may no longer be up to date. If you notice that a piece of information is outdated, please let us know at content@techsoup.org.

This content was created with AI assistance and has been reviewed and edited by Pablo Perez.