The first thing to note is how little any one source dominates. When AI systems answer questions about Romania, the single most-used source accounts for only about 4% of all citations, and the top three together for roughly 11%. Everything else, fully 72% of all visibility, is distributed across hundreds of different outlets. By the standards of search and recommendation systems, that is a genuinely diverse source base, and the systems deserve modest credit for not funnelling users toward a closed set of two or three names.

That said, the names at the top are worth reading. International institutions, the European Commission and the OECD, lead the list, which reflects how often AI systems answer questions about Romania through the lens of EU and OECD reporting rather than domestic journalism. English-language Romania coverage from Romania Insider and reference platforms such as Wikipedia and YouTube follow close behind, doing a good deal of the explanatory work. Romania’s own public-service and wire outlets, Radio România Internațională, Agerpres and the commercial newsroom Digi24, come next.

So the top of the distribution leans institutional and English-friendly, while domestic commercial newsrooms such as Libertatea, HotNews and Știrile ProTV sit closer to 1% each. This is a tilt rather than an exclusion: those outlets do appear, just lower down. The healthy sign is the long tail itself. The 72% spread across hundreds of outlets shows AI systems are reaching well beyond a core few, surfacing regional papers, niche sites and specialist sources that a more concentrated system would never show.

The fair reading is a balanced one. AI systems give disproportionate prominence to a small institutional core, and that pattern deserves scrutiny for what it means at the top. But the breadth underneath is real, and it is the more representative part of the picture.

There is no single AI view of Romania. When the identical prompt bank is run across systems, each one settles into a recognisable pattern of preferred sources, shaped by whether it browses the live web, which index it draws on, and how it weighs institutions against newsrooms.

Search-native systems such as Perplexity, Copilot and Google AI Overviews lean hardest on live Romanian outlets and platforms. Digi24, HotNews, Agerpres, Știrile ProTV and YouTube surface prominently, because these systems are effectively reading the current web in real time. Their source mix looks closest to an actual Romanian news feed.

Reasoning-first assistants such as Claude, ChatGPT and Gemini pull the other way, toward institutional and reference sources. The European Commission, OECD, Wikipedia and Reuters do more of the work. These systems more often explain Romania through stable, citable reference points than through today’s headlines. DeepSeek and Grok sit between the two, drawing heavily on international wires and reference sites.

The consequence is that which AI a person asks shapes which Romania they see. A Perplexity user is routed toward domestic commercial journalism, a Claude or Gemini user toward EU institutions and encyclopaedic summaries. Neither is wrong, but the divergence means source visibility is not a property of AI in general. It is a property of each specific system, and pluralism varies system by system.

Language is not a neutral setting. Posed in English, questions about Romania pull AI systems toward international institutions and English-language explainers. The OECD, the European Commission, Reuters, Wikipedia, the Reuters Institute and the Bertelsmann BTI index lead, alongside the English-facing outlet Romania Insider. The picture a non-Romanian-speaking user receives is filtered largely through supranational and foreign sources.

Posed in Romanian, the very same questions surface domestic newsrooms. Digi24, Libertatea, Știrile ProTV, HotNews and Adevărul rise to the top, with EU institutions still present but no longer dominant. The Romanian-language layer routes users into the actual national press in a way the English layer does not.

This divide has direct pluralism consequences. The English answer is more concentrated and more institutional. Its top three sources account for roughly 14% of citations versus about 11% in Romanian, and it leans on a narrower, more international base. Romanian answers are flatter and more locally grounded, distributing visibility across more domestic outlets.

For a diaspora reader, a foreign journalist, or anyone defaulting to English, the practical effect is that local Romanian journalism becomes harder to see. The same AI, asked the same thing, makes domestic outlets visible in one language and substitutes institutional summaries in the other. Language choice, not just topic, governs whose journalism reaches the user.

No single source list describes Romania. It changes with what is being asked. Economy and politics are dominated by institutional and international sources. The OECD, European Commission, Reuters and the BTI index lead, especially in English, where citizen-facing economic and political questions are answered largely through supranational reporting rather than domestic coverage.

Culture, sport and local life tell the opposite story. Here Romanian-language domestic and regional outlets surface most, alongside YouTube and dedicated event or festival sites, because this is content that international institutions simply do not cover. It is in these everyday, place-based topics that local journalism becomes most visible to AI systems.

Media and information questions surface the widest spread of named newsrooms, as systems list and compare outlets directly. Society, rights and environment sits in between, mixing EU and civil-society sources with domestic reporting.

The language split persists inside every topic. Even within economy or politics, the Romanian-language layer pulls in more domestic outlets than the English one. The takeaway is that pluralism is uneven across the map. A user asking about hard public-affairs topics in English sees the most institutional, least local picture, while cultural and local questions in Romanian surface the richest domestic-source diversity.

Being used and being recommended are not the same thing. When systems are asked outright for reliable or trustworthy sources, the picture shifts away from institutions and reference platforms toward independent and investigative Romanian newsrooms. Digi24, HotNews, G4Media, Recorder and PressOne lead the recommended set, alongside the national wire Agerpres and the English-facing Romania Insider.

This is one of the more encouraging patterns in the data. The outlets AI systems actively endorse lean toward accountability journalism. Investigative non-profits like Recorder and PressOne appear far more prominently here than in the raw citation mix, which suggests systems can distinguish sources they pulled from in passing from sources they consider worth trusting.

Recommendations are also less concentrated than citations. The top source captures only about 3% of all endorsements and the top five around 11%, spread across a wide field of national, regional and investigative titles. By system, search-native tools name more live newsrooms while reasoning-first assistants offer shorter, more curated reliability lists. The language split returns too: Romanian-language prompts yield more domestic recommendations, English prompts more international and English-facing ones.

The caveat is consistency. Because endorsement lists are short and vary by system and phrasing, no outlet is universally recommended. Reliability guidance depends heavily on which assistant a person asks and in which language. Still, the centre of gravity is clear. Asked to vouch for Romanian journalism, AI systems point first to independent and investigative outlets.

The systems draw a clear line between sources that are merely unreliable and those that are actively manipulative. The caution list, outlets to treat carefully, clusters on a small set of sensationalist news-TV and tabloid brands. România TV, Antena 3 CNN and Realitatea Plus dominate, joined by social and messaging platforms such as TikTok, Telegram and Facebook, and tabloid titles such as Cancan and Click!.

The disinformation list is different in kind. It points to a recognisable pro-Kremlin and conspiracy ecosystem, including Sputnik, ActiveNews, Național, SolidNews, Flux24, News-Pravda and TopWar, with Telegram recurring as the distribution channel that ties much of it together. These are not flagged for sloppiness but for coordinated manipulation and foreign-influence content.

That the systems separate these two categories at all is notable. They distinguish loud and partisan from engineered falsehood, and the named disinformation outlets map onto sources independent researchers have long associated with Romanian-language influence operations.

Two cautions apply. First, these flags are far rarer than recommendations. Systems are much readier to endorse than to warn, so the warned set is smaller and more volatile. Second, the lists skew toward well-known repeat offenders, so newer or lower-profile manipulation sources may pass unflagged. The pattern is directionally credible and aligned with the public record, but it should be read as the most visible tip of the problem rather than a complete map of it.

Findings on trust and distrust raise a harder question: on what basis? When a system attached a supporting link to its verdict, that link is the evidence behind the judgment, and those links point overwhelmingly to established independent Romanian newsrooms. Digi24, HotNews, Romania Insider, Știrile ProTV, Agerpres, Recorder and G4Media lead, with fact-checking and media-monitoring outlets such as Veridica and RISE Project in the supporting tail.

In other words, AI systems largely rely on a small circle of credible mainstream outlets to adjudicate the trustworthiness of everything else. That is reasonable as far as it goes, since these are real, reputable sources, but it also means the same handful of newsrooms quietly underwrites both the answers and the meta-judgments about which sources are good or bad.

The more striking finding is how often no evidence is offered at all. Reliability recommendations carried a supporting link only about 38% of the time, caution flags about 43%, and disinformation flags just 17%. The most serious accusation a system can make about a source, that it spreads manipulation, is the least likely to come with a citation.

This is a transparency gap with real stakes. When an AI tells a user that an outlet is unreliable or a propaganda vehicle and shows no source, the user has no way to audit the claim and must take the verdict on trust. The pattern suggests AI systems are more comfortable asserting source quality than demonstrating it, and that the confidence of a judgment is not matched by the evidence behind it.

Most of the names at the top of the Romania findings are easy to explain. The European Commission, the OECD, Reuters and Wikipedia are vast, globally indexed institutions, and Digi24 or Agerpres are among the largest newsrooms in the country. Romania Insider is neither. It is an independent, English-language website with a small editorial team and a readership measured in the low hundreds of thousands per month. Yet it ranks as the third most-used source overall, ahead of every Romanian-language commercial newsroom.

The pattern fits its design. Romania Insider is strongest in English-language answers, where it reaches about 4.6% of all citations, against 2.2% in Romanian, and strongest on economy and business questions, the subjects it was built to cover for an international audience. For a system answering an English query about the Romanian economy, it is an obvious match: credible, in English, and easy to parse.

None of this reflects badly on Romania Insider itself. By every available account it is a serious, editorially independent outlet, and a reader who lands on it gets a real and reasonable view of the country. The concern is not the source but the concentration. When AI systems lean heavily on a single English-language explainer as their main window onto a non-English country, they inherit that one outlet’s choices about what to cover and how to frame it, in place of the many domestic newsrooms reporting in Romanian. Romania Insider is a good outlet to surface; the risk is in surfacing it so much that it stands in for the rest. Healthy pluralism means no single source, however reliable, becomes the default lens.