AI Information Map: Hungary

AI Information Map · Hungary

The Hungary that comes back from the machines

20 questions, 10 assistants, two languages, some 8,000 phrasings apiece, and around 320,000 archived answers. This is an account of which Hungarian sources the systems put in a reader’s path, which they stand behind, and which they tell that reader to be wary of.

Contents

Research on Hungary is continuing. Visibility results are given as percentages of the slice under discussion; the counts that appear describe the design of the study, not its results, and are labelled where they occur. Figures are revised as further collection rounds are folded in. The full list of sources can be requested from MJRC.

AI Information Map · Hungary

How this was built

What was asked, of which systems, in which languages, and the rules by which every source that appeared on screen was logged, named and given a role.

The assistants and the prompt bank

10 widely used assistants were put to the same test: Perplexity, ChatGPT, Gemini, Copilot, Claude, DeepSeek, Meta AI, Grok, Mistral Vibe (formerly Le Chat) and Google’s AI Overviews. Each was given a fixed bank of 20 questions covering the ways people actually try to find out what is happening in Hungary, across five areas: media and information sources, politics and governance, economy and public services, society, rights and environment, and culture, sport and local life.

The bank deliberately mixes three kinds of question. Some are broad mapping requests of the “what are the main political issues” sort. Some ask the assistant directly which sources to trust or avoid. The rest are written in the voice of a person with a problem: a family relying on the public health service, a voter abroad, a parent following school reform, a visitor heading somewhere other than Budapest. Every question was put in both English and Hungarian.

Only the free tiers were used: what a person gets without paying. Paid and enterprise tiers may well behave differently, so what follows describes the ordinary user’s experience rather than the subscriber’s.

Scale, repetition and geography

The 20 prompts published alongside these findings are illustrative rather than exhaustive. Each underlying question was reformulated into roughly 8,000 phrasings, so that no single wording could drive a result, which yielded approximately 320,000 archived answers for this batch. Requests were issued through 30 IP addresses inside and outside Hungary, to reduce dependence on a single network location and to observe possible geographic variation. The collection window runs from June 2025 to April 2026, and a concluding round of validation prompting was carried out in June and July 2026.

The figures on these pages are drawn from the coded sample built out of that corpus. The final coded dataset contains 5,621 source-role records: 4,435 appearances in which a source was used or cited to build an answer, and 1,186 explicit judgements about sources, comprising 779 recommendations, 304 cautions and 103 disinformation-related flags. Of the 4,435 used or cited appearances, 3,254, or 73.4%, included a usable web address. Shares are computed on that coded material and will move as further rounds are coded and folded in.

How sources were coded

Each answer was stored with its prompt, the assistant, the language, the date, and whatever links, citation markers or source panels were visible. Every source the assistant produced was then lifted out and assigned one role:

• Used or cited: drawn on to build the answer
• Recommended: named to the reader as worth following
• Cautioned: named as something to avoid or treat carefully
• Flagged as disinformation-related: named as manipulation, propaganda or falsehood

Sources are counted by brand rather than by address. Wikipedia’s English and Hungarian editions count once. Telex’s Hungarian and English editions count once. European Commission directorates, agency portals and EU sub-sites are pooled as a single institutional entry, as are the town-hall sites of Debrecen, Szeged, Pécs and Győr, which the assistants use interchangeably when asked about life outside the capital. Public-media channels named jointly by an assistant (M1, Kossuth Rádió, Híradó.hu, and in some answers the national news agency MTI) are pooled under MTVA where the answer treats them as one thing.

Pooled institutional groups are broader than individual media brands and should not be read as directly equivalent units. “EU institutional sources (pooled)” covers Commission pages, Council and Parliament material, agency portals and other EU web properties; “Municipal sites (pooled)” covers four separate city governments. Reuters, Telex and HVG are single brands. A pooled group placing above a masthead means a whole institutional estate outweighs that masthead, not that one website does.

Two categories of warning, not one

The assistants distinguish between telling a reader to treat a source carefully and telling them a source carries manipulation or propaganda. In Hungary they used both, and all 10 systems applied the disinformation label to domestic sources at some point. Because the two verdicts differ in kind and in gravity, Finding 06 reports them side by side rather than collapsing them into a single warning list.

The political moment this describes

Hungary’s parliamentary election of 12 April 2026 ended 16 years of Fidesz government and brought the Tisza Party to power. Public-service broadcasting has since been restructured, pro-Orbán media operations have seen personnel and programming changes, and Index.hu’s editor-in-chief was removed after the vote. Because coding and validation continued through that period, many of the warnings recorded here are retrospective: they refer to ownership, state-advertising and editorial patterns documented during the Orbán governments, even where the answer itself was collected after the change of government and after parts of the system had begun to be reorganised. Where these findings report an outlet as pro-government, that is the assistants’ description, anchored in the earlier period.

How the results are expressed

Because the mapping is continuous, visibility results are given as percentages: an outlet’s share of the visibility available inside whichever slice is under discussion, whether that is the whole coded corpus, one assistant, one language, one subject area or one role. Running totals of source mentions are withheld, since they rise with every round. Where counts do appear they describe the design of the study rather than its results: the number of assistants, languages and IP addresses, the approximate scale of the run, the number of distinct outlets behind a percentage, and how many assistants took a given position. These are labelled wherever they occur.

The work is not finished. Hungary’s database stays open and these figures are refreshed as new rounds arrive. Shares will move. Read every finding as the state of the evidence today rather than a settled verdict.

What this study does not do

It records where sources appear and what the assistants say about them. It makes no assessment of whether an answer was accurate, balanced or complete, and it does not classify outlets as public, private, state-held or captured: that classification is a separate strand of MJRC’s work and will be applied to this dataset later. Appearing anywhere in these findings, on any list, reports what an assistant said. It is not MJRC’s judgement of the outlet.

Where the visibility goes

Brussels answers first

The three sources most often behind an answer about Hungary are a bloc of EU institutions, an international wire agency and an encyclopaedia. The country’s most-used newsroom comes fourth, on 2.4%.

Nothing here is concentrated. The most-used entry, the pooled European Commission and wider EU institutional web estate, accounts for 4.7% of everything cited about Hungary, and it is an estate of many sites rather than a single one. The top three together reach 11.1%. The remaining 73.2% is shared out among 1,791 further outlets. Anyone expecting the machines to have settled on two or three Hungarian mastheads will not find that pattern.

The order of the names is the finding. Above every Hungarian newsroom sit Brussels, Reuters and Wikipedia, and international and institutional sources also remain prominent below Telex. Telex, at 2.4%, is the highest-placed domestic title, followed by 24.hu, and then, at 1.7%, by Daily News Hungary, an English-language outlet with a fraction of the reporting capacity of either. The assistants reach for the institution, the wire and the reference work before they reach for a Hungarian journalist.

How far down the domestic press sits is worth spelling out. HVG and Index.hu hold about 1.1% apiece. Origo, Magyar Nemzet and Népszava are well under one percent each. Broadcast news barely registers at all: RTL, TV2 and the public channels together account for a handful of appearances across the whole coded corpus, broadcast output being less readily represented in the citable web layer measured here. Hungarian television, which is where a great many Hungarians still get their news, is close to invisible in this layer.

Both halves of the picture deserve stating. The summit is institutional and international, and that shapes what a reader is handed on the questions that matter most. But the field below is genuinely wide: county dailies, statistical offices, party websites, city halls, specialist outlets, none of which a more concentrated system would surface at all. Hungary’s AI information layer is not captured so much as displaced.

The sources assistants draw on most

Each bar is one outlet’s share of all sources used to build answers, pooled across 10 assistants and both languages. Bars are scaled to the largest value in the chart.

One prompt bank, 10 results

10 machines, 10 Hungarys

Put the same 20 questions to 10 assistants and 10 different countries come back. One builds its answers almost entirely out of Brussels and Reuters; another names 500 outlets and links to hardly any of them.

The spread between systems is wider than anything else in this dataset. ChatGPT is the most institutional assistant by a distance: EU institutional sources and Reuters alone supply nearly a third of everything it cites about Hungary, and with the OECD and the government portal added, roughly two-fifths of its sourcing comes from four entries. Its range is correspondingly narrow: 138 distinct outlets across 40 coded runs, fewer than any system except Copilot.

At the other end, Mistral Vibe names 518 separate outlets, three times ChatGPT’s range, and reaches deep into the Hungarian regional and specialist press. Claude and Grok also range widely, and both lean more on Hungarian-language newsrooms than the institutional systems do. Gemini is the outlier in kind rather than degree: its most-cited source is Google’s own properties, followed by the business title Portfolio.

DeepSeek deserves a separate note. Its most-cited source about Hungary is LinkedIn, and the Russian state agency TASS sits third, a profile no other assistant produces. It is also the system that most often returned an answer with no links at all.

The practical consequence is worth stating plainly. Which assistant a person opens decides which Hungary they are shown. Open ChatGPT and Hungary is described by the European Commission. Open Google AI Overviews and the first voice is Telex. Neither is broken. Source pluralism simply is not a property of “AI”: it belongs to each product, and has to be judged one product at a time.

What each assistant reaches for

How much they cite against how widely they range

Citing a great deal and citing widely turn out to be different habits. ChatGPT is the clearest case: it produces a heavy citation load, roughly 13 sources per answer, but those resolve to very few distinct names, at 0.26 outlets per citation, the poorest ratio in the study. It returns to the same handful of institutions again and again. DeepSeek inverts this, offering fewer citations that lead to proportionally many more places.

The lesson is that a densely sourced answer can still rest on a very small number of places. Volume and variety have to be measured separately, because neither predicts the other.

Volume and range are indexed against the heaviest citer, Mistral Vibe, set to 100. “Outlets per citation” divides distinct outlets by total source appearances: the higher the figure, the less an assistant repeats itself. Whether those citations are actually linked is a separate matter, treated in Finding 08.

What the language of the question decides

In English, the Hungarian press nearly disappears

The same 20 questions, asked in Hungarian and then in English, return two different countries. Domestic newsrooms supply a fifth of the citations in one and under four percent in the other.

This is the sharpest divide in the Hungarian data. Taken together, Hungary’s own Hungarian-language newsrooms, among them Telex, 24.hu, HVG, Index.hu, 444.hu, Origo, Magyar Nemzet, Népszava, Portfolio and Pénzcentrum, account for 20.3% of everything cited when the question is asked in Hungarian, and for 3.7% when the same question is asked in English. That is not a thinning but a near-disappearance.

Something takes their place. English-language titles covering Hungary, such as Daily News Hungary, Hungary Today, The Budapest Times, Hungarian Conservative, The Debrecen Sun and Budapest Business Journal, supply 7.5% of English-prompt citations and 2.2% of Hungarian ones. Ask in English and this small group of outlets outweighs the entire domestic press by two to one. Telex, the country’s most-used newsroom overall, falls to 12th place in English on 1.1%; 24.hu falls to 0.4%; HVG and Index.hu to 0.2% apiece.

Above both groups the institutions hold steady. EU institutions, Reuters, Wikipedia and the OECD lead in either language, and are simply more dominant in English, where they face less competition. The Hungarian layer also pulls in the state’s own numbers, since the statistical office and the government portal each rise roughly fivefold when the question is put in Hungarian, because that is the language those pages are published in.

The people affected are identifiable. A correspondent in Budapest working in English, an official preparing a briefing, a member of the diaspora, a researcher without Hungarian, a company assessing the country: each asks in English, and each is handed Brussels, the wire services and a handful of English-language local titles, with the newsrooms that actually break Hungarian stories almost entirely absent. Nobody has been misinformed; they have simply been shown a Hungary described from outside it.

Asked in English

Asked in Hungarian

Bars give each outlet’s share of citations within its own language layer, and each chart is scaled independently.

Subject by subject

The subject decides who speaks

Ask about rights or the economy and the answer is assembled in Brussels. Ask about a festival, a match or a town outside Budapest and Hungarian voices finally arrive — from the town hall as often as the newsroom.

There is no single answer to who speaks for Hungary; it depends entirely on what is being asked. On society, rights and the environment, EU institutions supply 16.1% of all citations, more than three times the next source, joined by the OECD, Human Rights Watch and the Council of Europe. On the economy the pattern repeats with the Commission, the OECD and the statistical office. These are precisely the questions on which a citizen might most want domestic scrutiny, and they are the questions routed most firmly through supranational bodies.

Politics behaves differently again. Here the wire agency and the encyclopaedia lead, and one of the five leading sources in the Hungarian layer is a party website: Tisza’s own pages, at 3.4%. In this sample, party-controlled pages sometimes appeared ahead of journalistic reporting.

Culture, sport and local life produce the study’s most Hungarian answers, but not in the way one might expect. The leading source is not a newspaper: it is the pooled municipal websites of Debrecen, Szeged, Pécs and Győr. Local journalism does appear, in Debreciner, Szegeder, Szabad Pécs, Dehir and the county dailies, but it sits beneath the town halls, which publish listings in a form a machine can read easily.

The most self-referential category is the one about the press. Asked to describe Hungarian media, the assistants cite Telex, Reuters and 24.hu, and the Reuters Institute’s Digital News Report at 4.3%, level with the outlets themselves. A single annual comparative survey carries much of the weight when these systems account for Hungarian journalism, while the country’s own media scholarship, its regulator and its press bodies barely surface at all.

Leading sources in each subject, split by the language of the prompt

Bars give each outlet’s share of citations inside that subject. The tags beneath show which sources lead once the same subject is split by prompt language.

What the machines say to trust

Endorsed and overlooked

Asked which Hungarian sources to follow, every one of the 10 assistants names Telex. Left to answer anything else, they reach for institutions they never recommend and use the endorsed titles sparingly.

Using a source and vouching for one are separate acts, and the gap between them is the finding here. Invited to name reliable Hungarian sources, the assistants set aside the encyclopaedia and the Commission and reach for independent journalism. Telex leads on 6.0% of all endorsements, followed by HVG, 444.hu and 24.hu. Telex is recommended by all 10 systems, as are HVG, 444.hu and 24.hu, a rare point of unanimity in a study otherwise defined by divergence.

Set the two lists beside each other and the mismatch is plain. EU institutions, Reuters and Wikipedia, first, second and third among the sources actually used, collect almost no endorsements between them. Telex, first among endorsements, supplies 2.4% of citations. The assistants know which Hungarian outlets a reader should consult, and consult different ones themselves.

Endorsement is also unusually dispersed. The leading title takes 6.0% and the top five together 19.6%; the rest is spread across 331 further names. Local and investigative outlets do disproportionately well here: Debreciner, Szegeder, Szabad Pécs, Dehir, PécsMa, HAON, Átlátszó, Direkt36, Partizán and Lakmusz all appear far above their citation share. Asked directly for local sources, the systems produce a credible list of Hungarian local journalism. Left to answer a question about a Hungarian town on their own, they cite the town hall.

Two qualifications. English-language titles aimed largely at foreign residents, such as Daily News Hungary, Hungary Today and The Budapest Times, sit in the top 10 of a list purporting to name Hungary’s most reliable sources, a hierarchy unlikely to mirror that of a typical Hungarian-language news user. And endorsement lists are short and sensitive to phrasing, so the ranking below the leading four should be read as indicative.

The outlets assistants vouch for

Endorsement split by the language of the request

Where each assistant puts its confidence

Chips and bars show the titles most often endorsed inside each assistant’s own recommendation set, so shares are internal to that system.

What the machines say to avoid

A warning list organised around ownership

The assistants do not stop at telling readers to be careful: all 10 apply the disinformation label to Hungarian sources. The outlet drawing the most warnings, in both categories, is the public service broadcaster.

The assistants issue two grades of warning about Hungarian sources, and they use both freely. Alongside 304 cautions, every one of the 10 systems applied the outright disinformation label, producing 103 such judgements. Whether that reflects Hungary’s media terrain or the models’ priors about it is not something this dataset can settle. What it can record is that the charge was made, repeatedly, and by everyone.

Both lists are organised around ownership and political alignment rather than fabrication: the wording researchers logged returns again and again to state advertising, to the 2018 consolidation of some 500 titles under the KESMA foundation, and to editorial control following a change of proprietor, while almost nothing in the warned set concerns a specific false story.

The timing matters, and readers should hold it in view throughout this section. The parliamentary election of 12 April 2026 ended 16 years of Fidesz government and brought the Tisza Party to power, and the Hungarian media system has been reorganising since: public-service broadcasting has been restructured, pro-Orbán media operations have seen personnel and programming changes, and Index.hu’s editor-in-chief was removed after the vote. Many of the assistants’ warnings are therefore retrospective. They refer to ownership, state-advertising and editorial patterns documented during the Orbán governments, even where the answer was collected after the change of government and after parts of that system had already begun to change. “Pro-government”, as the assistants use it here, describes an arrangement anchored in the earlier period, and in several cases that arrangement no longer holds. What the systems carry into the present is a reputational assessment formed over years, applied to a landscape moving in months.

The public broadcaster heads both lists. Pooled as MTVA, covering M1, Kossuth Rádió, Híradó.hu and the associated channels, it takes 11.2% of all cautions, from all 10 assistants, and 12.6% of disinformation flags, from seven of them. Origo follows on both, cautioned by all 10 systems and flagged by six. Magyar Nemzet, TV2, Mandiner and the KESMA group complete the core.

Beneath that sits a smaller and different group: the far-right Kuruc.info, the campaign operation Megafon, sites such as Vadhajtások and 888.hu, and, thinly but present, a Kremlin-adjacent tail of Hidfő, NewsFront Hungarian, Sputnik and RT in Hungarian, plus a handful of conspiracy blogs. Together these account for under a tenth of the flags. The assistants are far more exercised about ownership concentration than about pro-Kremlin content.

Two asymmetries are worth naming. There is no equivalent grouping on the other side of Hungarian politics: apart from two isolated mentions of Partizán, nothing critical of the government appears in the warned set. And Index.hu, cautioned by seven systems, is simultaneously one of the 10 most-cited sources in the whole study, a contradiction taken up at the end of these findings.

Sources the assistants say to handle carefully

Sources flagged as disinformation-related

Warnings split by the language of the request

Cautions and disinformation flags are counted separately and each chart is scaled to its own maximum. Warnings are a small share of all source mentions and pick out recurring subjects rather than every problematic source. Appearing on either list records what an assistant said during the collection window; it is not MJRC’s assessment of the outlet, and it may not describe the outlet as it stands today.

The proof behind the judgement

Verdicts without evidence

Across 1,186 rulings on which Hungarian sources deserve trust, fewer than three in 10 arrived with anything the reader could open. Warnings were substantiated half as often as praise.

Once an assistant has told a reader whom to trust and whom to doubt, the question is on what authority. Where a judgement carried a citation or a link, that is the assistant’s reasoning made visible. In Hungary, most of the time, there was nothing to see: 28.4% of all source judgements came with any supporting reference at all.

The imbalance within that figure matters more than the headline. Recommendations were substantiated 34.9% of the time. Cautions managed 16.1%, and disinformation flags 15.5%. The graver the accusation, the less likely the assistant was to show its working.

That asymmetry would be less troubling reversed. An unsupported recommendation costs a reader little, since they can go and look. A statement that a national broadcaster is a vehicle for disinformation, delivered with nothing attached, cannot be traced, tested or argued with; it can only be swallowed or dismissed. And the bodies built precisely to make such judgements auditable, among them the Media Ownership Monitor, the CEU media centre, Reporters Without Borders, IPI and the Hungarian fact-checking desk Lakmusz, appear only at the margins of the evidence that does exist.

Between systems the gap is total. Copilot supported seven judgements in 10, and Grok six. ChatGPT supported none. It delivered dozens of verdicts on Hungarian media, including several disinformation flags, and gave the reader not one reference by which to check any of them. Claude was barely different at 3.3%.

How often a verdict came with proof

How much proof each assistant offered

The share of each assistant’s own endorsements, cautions and disinformation flags that arrived with a citation or link. A striped bar marks a system that supplied none. Systems making few judgements move more easily between rounds.

The citation without a link

The unlinked citation

A quarter of all sources in this study are named in the text and never linked. One assistant supplies an address for barely one source in eight, and 24 runs displayed no links whatsoever.

Of the 4,435 sources used to build answers about Hungary, 73.4% came with a usable web address. The remaining quarter were named in prose and left there: an attribution that reads exactly like sourcing and cannot be followed.

The shortfall is heavily concentrated. Mistral Vibe, the widest-ranging system in the study, attached an address to 12.1% of the sources it named. Google AI Overviews managed 39.4% and DeepSeek 63.6%. The other seven systems linked almost everything they named, five of them above nine in 10.

This recasts Finding 02. Mistral Vibe’s 518 distinct outlets looked like the strongest showing for pluralism in the dataset. Most of those outlets are names in a sentence. The reader is told that a Hungarian regional title reported something and given no way to reach it, which for that title is worse than not being mentioned: its name lends weight to an answer while no reader is sent to it, no claim can be traced back, and no visit is recorded. The citation becomes ornamental: borrowed authority with the borrowing hidden.

A smaller failure sits underneath. In 24 of the 400 coded runs, an assistant produced a full and confident account of some aspect of Hungary with no links displayed at all. 17 of those were DeepSeek, the rest scattered across Perplexity, Gemini and Grok. Those blanks skew by language, too: 17 fell in the Hungarian layer against seven in English.

The share of each assistant’s named sources that carried a link

Bars give the proportion of each assistant’s source appearances that included a usable web address. Naming without linking, rather than an absence of sourcing language, is the dominant pattern.

A single outlet, examined

Index.hu, used and distrusted

The 10th most-cited source in this study is also one that seven of the 10 assistants tell readers to handle with care. They use it constantly and warn about it in the same breath.

Most names near the top of these findings need no explanation. The European Commission publishes enormously and is indexed exhaustively. Reuters is a wire service. Telex is the only Hungarian newsroom in the overall top four and the most-recommended outlet in this study. Index.hu is the one the machines cannot hold a single position on.

It is the 10th most-used source in the study at 1.1% of all citations, and the fourth most-used Hungarian newsroom. It is also the sixth most-cautioned source at 4.9% of all warnings. Of the 17 judgements passed on it, 15 are cautions and two are recommendations, and both recommendations are hedged: one suggests it for quick headlines with cross-checking, the other for readers using browser translation.

What the assistants actually say

The wording is remarkably consistent across systems that agree on little else. ChatGPT logs it as requiring care because of ownership change and dependence on state advertising. Claude describes it as pro-government following a change of proprietor. Grok twice notes a reputation for independence described as diminished. Meta AI, Copilot, DeepSeek and Google AI Overviews all reach for the same explanation. Not one warning concerns a false story. Every one concerns who owns the outlet and what followed.

Hungarian readers will recognise the reference. The 2020 departure of the editor-in-chief and the resignation of most of the newsroom, who went on to found Telex, is among the most documented episodes in recent European media history. The assistants have absorbed it accurately.

Why the contradiction persists

The two behaviours never meet, because they follow from different questions. Asked which Hungarian sources should I trust, an assistant produces the ownership story and issues a caution. Asked what is happening in Hungary, it produces whatever Hungarian-language reporting matches the query, and Index.hu, still large, still publishing constantly, still widely indexed, is among the first things it offers. The outputs do not visibly reconcile the caution with the later citation. Nothing in either answer acknowledges the other.

Note the asymmetry in who sees which. The caution surfaces mostly when a reader thinks to ask about sources. The citation surfaces on every ordinary question, and 47 of its 50 appearances came from Hungarian-language prompts, that is, from readers most likely to be in Hungary and least likely to be asking an assistant to vet the Hungarian press for them. The people most exposed to the outlet are the least likely to be shown the warning attached to it.

The outlet

Origins

Launched in 1999 out of the earlier Internetto, and for two decades among the most-read news sites in Hungary

The 2020 rupture

The editor-in-chief was dismissed in July 2020 and the great majority of the editorial staff resigned; many founded Telex, now the most-recommended Hungarian outlet in this study

Standing at the close of collection

Still a large general-news operation with substantial traffic and a full daily output; its editor-in-chief was removed following the April 2026 election

Where it sits in this data

Tenth most-cited source overall; sixth most-cautioned; recommended twice, both times conditionally; never flagged as disinformation by any system

Details of the outlet are drawn from the public record. Its presence among cautioned titles reports what the assistants said during the collection window and is not MJRC’s assessment of Index.hu.