Inside how two Hungarian newsrooms use AI to analyze documents, uncover patterns and accelerate investigations-while keeping verification and editorial judgment firmly in human hands.

Author: Orsolya Seregély

For Hungarian readers, 444.hu needs no introduction – but fewer people realize that the fact-checking site Lakmusz is also part of the Magyar Jeti universe, the same media group.

444 has long been known for its big investigative stories – plenty of the pieces that really rattled the government, whether corruption scandals or pedophilia cases, ran there first. So it’s not exactly surprising that Lakmusz grew out of the same stable. As they put it on their site, the goal isn’t just to publish reliable, verified information, but to push public debate toward facts and to expose how disinformation spreads.

And for that kind of work, AI turns out to be genuinely useful. We talked to Zsuzsanna Kékesi, who runs Magyar Jeti’s AI projects, about how they’re actually using these tools – in fact-checking, in investigations, and in the day-to-day of the newsroom.

“We use AI mainly for fact-checking and investigations, especially when we get documents back through freedom-of-information requests. Sometimes it’s years’ worth of financial declarations, contracts, grant paperwork – a huge pile of documents, and honestly we often don’t even know exactly what we’re looking for yet. That’s exactly when AI earns its keep,” Zsuzsanna says. Their most recent example: a research project into how the Hungarian government system actually operates, for which they requested grant application files – several thousand pages of PDFs. Along with the applications themselves came attachments and years of financial settlements.

“We didn’t know precisely what we were after, but we knew something was being hidden. We just had to figure out what. We ran it through Gemini – it has access to Google Workspace, so it could work through all the documents and pull out the relevant information from the PDFs. We asked it to organize everything into a spreadsheet. Once we looked that over, patterns started jumping out, and we could go back and dig into those specifically. That’s how we found what was being concealed, and where.”

They rely on this kind of approach a lot, partly because the material that comes back from data requests is often a nightmare to work with. At one point, for example, they received a stack of hand-filled-out asset declarations, which made things considerably harder.

“ChatGPT read the handwritten declarations for us. We worked from that, and it fed directly into the articles we wrote,” Zsuzsanna says – adding that anything AI reads or organizes gets checked by a human, every single time.

“We check everything the AI gives us, but even so, these tools are incredibly effective. It’s also good at things like tracing people through digital footprints, or figuring out who’s in a photo.”

The newsroom also uses AI for transcribing and subtitling audio and video, and for summarizing long documents.

“We have an English-language newsletter, Insight Hungary, where we already have an AI read-aloud feature. We haven’t rolled it out in Hungarian yet – the quality isn’t there – but in English it works really well.”

Tools built for their own needs

444 and Lakmusz don’t just use off-the-shelf AI tools – they also bend existing ones to fit their own purposes. One example: a system for tracking international news.

“Realistically, one person can only keep tabs on a handful of major Western outlets – German, French, Italian, Spanish, British, American. But plenty is happening elsewhere in the world too. We found a tool that tracks publishers’ social media activity and front pages and ranks stories based on that.”

The system also follows keywords and politicians’ names. Early on, they mostly used it to track Fidesz politicians, members of the Orbán family, and government figures – including how they were being covered in the Arab press.

“It’s the kind of task nobody would ever take on by hand. But if you get an automatic daily digest, it’s suddenly much easier to catch stories that would otherwise slip right past you.”

AI isn’t just for editorial work either – they use it on the management side too, to analyze data, track subscriber numbers and other results, and manage their dashboards.

“We adapted another existing tool for this. It flags anything on the dashboard that deviates from the norm – a sudden spike or a drop. That way we notice much faster when something needs a closer look.”

Some of these repurposed tools weren’t originally built for newsrooms at all. The news-tracking tool, for instance, was originally made for politicians – journalists needed something different out of it. The team learned the system through a workshop, then adapted it to their own workflow together with its developers. On top of that, they have a four-person dev team in-house, so some tools they can just build themselves.

“We experiment a lot. Plenty of it never turns into an actual public tool or product – there’s always more room to grow.”

The ground rules

Kékesi recently joined 444, after gaining first-hand experience with the early adoption of AI at other media organizations. As she tells it, traditional journalists were often wary of it at first – and even now, there’s still a lot of conversation needed about what it’s actually good for and how to use it responsibly.

“It doesn’t come naturally to a journalist to ask AI for help thinking through a piece. But on one thing we’re all in agreement: we don’t use it to write content. It doesn’t produce the article for us – though I do see magazines out there doing that. News media in general still tends to be more cautious about AI. But once a newsroom does start using it, it tends to just keep growing organically from there.”

A lot of outlets also use AI for illustrations. When that happens, both 444 and Lakmusz label the image as AI-generated or AI-assisted.

“It’s interesting to think about where exactly we draw that line – nobody ever labeled a photo as ‘edited in Photoshop’ before.”

AI and election campaigns

On the subject of elections, Zsuzsanna pointed out that AI-driven content isn’t some uniquely Hungarian phenomenon – it’s an international trend.

“A lot of people are riding this wave, and audiences are often gullible – they can’t tell AI-generated video from the real thing.”

The core problem, in her view, is a lack of critical thinking.

“People don’t stop to ask why a video exists, who made it, or who benefits if they watch it or share it. Why share something if you don’t even know what it actually is? That’s why media literacy and public education are such a priority for us at Lakmusz.”

Trust and transparency

Does being transparent about AI use actually build audience trust? Zsuzsanna thinks news consumption has increasingly become tied to individual people – influencers, well-known journalists – rather than to institutions. People often trust one familiar face more than they trust a newsroom as a whole.

In a media landscape that’s become heavily politicized, trust in traditional media can erode even as influencers gain ground.

“Wherever autocracy or populist politics is on the rise, trust in traditional media tends to fall – and trust in influencers and public personalities rises to fill that gap.”

That may be part of why so many 444 journalists are recognizable faces to their readers. The newsroom regularly asks its audience directly what they think about AI use, and whether it affects their trust.

“Turns out, it basically didn’t matter to them – as long as we kept delivering the quality and standards they already valued about 444. As long as we stay credible and it’s still the same journalists producing the work, they don’t really care what tools are running in the background.”

Staying skeptical, and where the real value is

“The way we use AI, tasks that used to take days can now take a few minutes. That’s the real point for us – it speeds up and supports the work, it doesn’t replace it.”

A good example: a grant-funding investigation where they dug through the financial settlements of dozens of civil organizations.

“Without AI, we honestly wouldn’t have been able to get through that dataset. It was already clear some of the settlements weren’t legitimate, so we also asked the tool to pull out the names of everyone involved.”

Once the information had been compiled, recurring names, addresses and administrative contacts began to emerge across records connected to several formally separate organizations.

“The same notary or clerk kept showing up, the same address kept showing up, the same people kept showing up in different places.”

This is where AI proved particularly useful: not as a source of conclusions, but as a tool for identifying patterns across a large volume of documents. The journalists then returned to the original records to verify the connections identified during the analysis.

The newsroom also works with smaller outlets, coordinating investigations and occasionally sharing documents obtained through freedom-of-information requests.

“Right now, data requests are basically a huge playground. We pull the information, and after that, what you get out of it really comes down to the journalist’s own creativity. We inspire each other too – sharing what kind of data is out there to get, and what stories you can actually build out of it.”