AI Pluralism Matters. Meet the Project Mapping It Across the World

AI is becoming a gatekeeper for public information. The Media and Journalism Research Center’s AI Information Map offers something regulators, publishers and the public have lacked: a way to see whose journalism these systems make visible, and whose they leave behind.

When people turn to an AI assistant for an explanation of an election, a public-health debate or a country’s politics, the answer does more than summarize information. It chooses an information environment: which newsrooms, public institutions, databases and reference sources will be visible, and which will not.

That choice is now being made at scale. Yet independent evidence of how AI systems make it is still scarce. The Media and Journalism Research Center (MJRC) began building such evidence in January 2025, when it launched the AI Information Map. The project tracks ten widely used systems (ChatGPT, Gemini, Copilot, Perplexity, Claude, DeepSeek, Meta AI, Grok, Mistral and Google AI Overviews/Search) and records the sources they cite, link, recommend, caution against or simply name. It tests questions in English as well as national and local languages.

Each response becomes part of a growing empirical record of AI information pluralism. By comparing the sources behind generated answers across products, languages and countries, the Map turns an otherwise opaque process into something that can be examined. The debate has moved beyond whether AI can write. The urgent question is whose writing AI makes findable.

The value of staying with the story

The Map also reflects MJRC’s distinctive strength: long-term research. Its major projects are built to follow a changing information system over years, and, in some cases, across two decades, rather than capture a single moment. The Media Influence Matrix, launched in 2017 and now described as a permanent project, offers one example of that sustained approach. It is particularly valuable for AI: a one-off test can reveal an odd answer; repeated observation can show whether a pattern endures, whether a source steadily gains or loses visibility, and whether the information landscape changes as companies alter their products.

MJRC’s wider research programme already treats media power as something that must be mapped over time: across ownership, funding, regulation, audiences and, increasingly, the technology companies that shape how information circulates. The AI Information Map carries that same long-view method into the systems that are beginning to mediate how people encounter the news.

One country, two information systems

The first published country studies, Romania, Spain and Hungary, show why measurement matters.

In Romania, MJRC analyzed about 360,000 archived AI answers. At first glance, the source mix looked diffuse: the European Commission was the most visible source but accounted for only about 4% of citations. The sharper finding emerged when researchers changed the language of the prompt. In English, the systems leaned toward the European Commission, OECD, Reuters, Wikipedia and other international or English-language sources. In Romanian, Digi24, Libertatea, Știrile ProTV and HotNews became much more visible.

The implication is straightforward: an assistant is not necessarily translating one shared picture of Romania. It can construct different versions of the country for Romanian-speaking and English-speaking users.

Spain reveals a different aspect of the same problem. Across about 400,000 archived answers, El País was the most visible individual source, yet it accounted for just 4.9% of citations. More than three quarters of source visibility was spread among another 1,924 sources. Wikipedia, Cadena SER, the government portal La Moncloa, Reuters and the European Commission all ranked near the top. Spanish journalism was present, but so were institutional and reference sources that shape the information layer around it.

The Spanish results also point to an accountability problem. RTVE, the public broadcaster, appeared near the top of both the systems’ recommended and cautioned lists. Nine of the ten assistants described it as reliable, while several also warned users about it, usually without explaining the apparent contradiction.

Hungary makes the language effect harder to ignore. MJRC’s study drew on about 320,000 archived answers. When questions were asked in Hungarian, Hungarian newsrooms supplied about one fifth of citations. When comparable questions were asked in English, their share fell below 4%.

For a user who does not read Hungarian, the AI-mediated version of Hungary can therefore be shaped far more by international institutions and external sources than by the country’s own journalism. The study also found that 26.6% of named sources in the coded Hungarian sample had no usable link. Mistral linked 12.1% of the sources it named; Google AI Overviews linked 39.4%.

That matters to publishers as much as to readers. A newsroom may lend authority to an AI-generated answer without receiving a click, or giving the user a practical route back to the reporting itself.

A measurement tool arrives as Europe starts regulating

The project’s relevance became more immediate on 31 August 2026, when the European Commission designated ChatGPT a Very Large Online Search Engine (VLOSE) under the Digital Services Act. OpenAI reported an average of 159.1 million monthly active users in the European Union; Google Search and Microsoft Bing had already received VLOSE designations.

The designation brings additional duties to assess and mitigate systemic risks associated with the service and its algorithmic systems, including risks to fundamental rights, electoral processes and public security. ChatGPT must meet those additional obligations by January 2027.

Not every system in the AI Information Map is itself designated as a VLOP or VLOSE. But a major system is now within that regime, and several others are closely tied to already regulated search engines and platforms. The regulatory question follows: how can authorities assess effects on media pluralism if they cannot observe which sources an AI system repeatedly elevates, overlooks or names without attribution?

Platform transparency reports are part of the answer, but they cannot replace independent observation. The AI Information Map does not claim to be a DSA compliance audit. It does provide the outside evidence regulators need to ask better questions: whether a small set of sources dominates answers; whether local journalism vanishes across languages; whether politically influenced outlets gain unusual visibility; whether citations arrive without usable links; and how sharply results vary from one AI product to another.

Why publishers and companies are watching

The audience for that evidence is already broader than media researchers. MJRC says think tanks and regulatory authorities in six European countries are using the project and its data in work on AI, media pluralism and the obligations of large information intermediaries. Commercial companies have also begun to use it.

Their question is practical: which sources do AI systems pick up? As information seeking shifts from link lists toward generated answers, visibility inside an AI response becomes commercially consequential. Organizations want to know which publishers, databases and institutional pages the systems treat as authoritative, why a company or subject appears through one source but not another, and how those patterns vary by market, language and product.

AI source visibility is becoming part of the information economy. For publishers, the stakes include audience and attribution. For businesses and public institutions, they include discoverability and reputational context. For regulators, they include whether the new gatekeepers broaden public access to information or quietly narrow it.

From three countries to ten, and from free AI to paid AI

MJRC and its partners are now extending the first phase of the project beyond Romania, Spain and Hungary to Germany, the United Kingdom, Serbia, Brazil, India, South Africa and the Philippines. The group spans mature European markets, a non-EU country with a contested media environment, large Global South information markets, and multilingual settings where platform dependence and disinformation have been central concerns.

The expansion should help distinguish local patterns from features of AI-mediated information that travel across markets.

Explore the AI Information Map and its country programme

A second question is emerging alongside the geographic expansion. The project has focused mainly on free versions of AI systems because they reach the widest public. But paid subscriptions may offer different models, search and browsing functions, context windows and other capabilities; enterprise products add another layer. MJRC plans to compare paid and free versions to see whether payment changes not just the answer, but the sources behind it.

That comparison could expose a new form of information inequality: whether people who pay for AI encounter a meaningfully different information ecosystem from those who do not.

The first phase of the AI boom was obsessed with whether machines could produce convincing prose. The next phase should examine what happens when those machines decide which journalism, institutions and voices the public sees. The AI Information Map makes that question measurable.