“Welcome to Vilnius, the city where we face cognitive warfare for real,” Robertas Kaunas, Lithuania’s minister of national defense, said in his opening remarks at this year’s GlobalFact.
On June 17, hundreds of fact-checkers assembled in Vilnius for the annual conference of the International Fact-Checking Network (IFCN), part of the Poynter Institute. Roughly 180 fact-checking organizations in more than 55 countries are signed onto IFCN’s Code of Principles, and many of them made the journey to Vilnius for the three-day event.
The historic capital of Lithuania was a fitting gathering place. Speakers throughout the program echoed Kaunas, detailing the barrage of Russian information manipulation and interference campaigns faced by countries across the Baltic region. (Lithuania shares more than 170 miles of its border with Russia.)
“Let me be clear about the role of the media today. You are democratic infrastructure as important as an air defense system or artillery,” said Kaunas in his speech. “No country can defend its information space alone.”
The Lithuanian government’s vocal support for fact-checkers received applause. It was a far cry from the recent rhetoric coming out of the U.S. government, and from many of the global social media companies that distribute the disinformation that fact-checkers fight daily.
In 2024, the Trump administration shut down the U.S. Agency for International Development (USAID), which had given grants to fact-checking organizations around the world. Last year, Meta ended its third-party fact-checking program in the U.S. — another significant revenue stream for IFCN signatories — claiming it was an effort to reduce “censorship.”
While Meta has yet to shutter its international third-party fact-checking program entirely, it has made clear its intentions to follow in the footsteps of X and disinvest in professional fact-checks in favor of crowdsourced “Community Notes.”
One signal of big tech’s waning support for fact-checkers was how few major tech companies were in the room at GlobalFact. While Meta and Google have historically sent representatives annually, only TikTok dispatched a proper delegation this year.
“The political wind shifted, and their support shifted with it. To the technology platforms not in this room today, rejoin us in the work of making high-quality, accurate information accessible to everyone,” Angie Drobnic Holan, director of IFCN and a 2023 Nieman Fellow, said in her own opening remarks. “Fact-checking is not censorship. It is not partisan. It never was.”
The grant model is unsustainable. What comes next?
The one-two punch of funding cuts from the U.S. government and big tech platforms has left fact-checking organizations at a crossroads. They can either compete for grants from a dwindling pool of philanthropic support, or they can explore new ways of making money.
“I’m not super at math, but I’m good enough to know you need more funds coming in than going out,” Holan said in a panel about alternative funding models. “Philanthropy has told me there is not enough philanthropic money to support your entire sector. There’s just not. But there’s also evidence to show that there’s a substantial pool of audience revenue.”
In April, IFCN’s annual State of Fact-Checkers report found that 62% of the 141 organizations surveyed had grown their audiences in 2025, but just 22% described their financial position as “sustainable.”
Peter Erdelyi, the founding director of the Center for Sustainable Media, made a clear-eyed pitch for revenue diversification during his mainstage talk. He argued that fact-checkers fall into the “public good trap”: because the benefits of their work are broad and diffuse, they are hard to monetize. He encouraged fact-checking organizations that can no longer sustain themselves on grant funding to build paid editorial products that serve targeted audiences, or even businesses.
On the consumer side, he suggested they could explore partial subscriber paywalls, or paid newsletters for professional audiences like lawyers. This revenue from paying readers could buttress the organization’s core, free fact-checks. On the business side, he threw out theoretical examples, like drug fact-checks for pharmacies, disinformation training for doctors’ offices, or monitoring products for health insurance companies.
“I need to tell you honestly that the market shows us that you need a mechanism to get people to pay,” said Erdelyi. “You need to push people, because a lot of people are meaning to support you, but they don’t. You need to give them an excuse.”
Across the conference, fact-checking organizations shared their own experiments in balancing serving the public with money-making endeavors.
Franco Piccato, the executive director of Chequeado, an Argentine nonprofit fact-checking organization, said his team custom-built an AI transcription tool optimized for Spanish that can be used by students for note-taking, journalists for event coverage, or academics for interview research. Chequeado is now selling the transcription software, called Desgrabador, as a service, with the goal of covering roughly 10% of its annual budget through tech product revenue by the end of this year.
Africa Uncensored, a Kenya-based outlet known for its investigative documentaries, has found its own balance. Roughly 80% of its revenue comes from philanthropic sources, while another 20% comes from commercial partnerships, according to Eric Mugendi, an editor-at-large for partnership and initiatives. That includes production deals with broadcast media and licensing archival footage. While this commercial revenue is limited, it allows Africa Uncensored to keep its dedicated fact-checking site, Piga Firimbi, free.
“LLM poisoning” and how to test for it
With a record number of people globally turning to social media for news, the work of fact-checking disinformation on social platforms is as important as ever. Still, several speakers called attention to an emerging vector for disinformation that some have coined “LLM poisoning.”
Bad actors are flooding the web with articles containing false or misleading narratives in the hope that commercial generative AI products will reflect them in their responses. A Russian network of over 180 sites — dubbed the Pravda network — found early success with the strategy by publishing false articles in Nordic languages.
“If you ask ChatGPT who was the aggressor in the Russia-Ukrainian war, who gets to decide the answer? Or is it neutral? Or is there a bias?” said Pekka Kallioniemi, a Finnish researcher and content creator who runs the Twitter account Vatnik Soup, which documents Russian disinformation campaigns. “ChatGPT, Claude — that’s the next battlefield. That’s where we see the battle of narratives.”
“Poisoning” can occur at several different stages in the LLM life cycle. It can happen when large swaths of the internet are scraped for data to train a model, or when a product uses real-time web search to answer a specific user query. Kallioniemi argued these attempts at narrative control have been aided by the rollout of generative AI tools.
“If you look at 2013, there were these infamous troll farms where people were working in shifts, like a factory producing commentary online,” said Kallioniemi, talking about the old Russian disinformation playbook. “Now most of that is being automated.”
Another session took on the natural follow-up question: How do you even begin to track bias in AI models? The research lab Gazzetta previewed a beta tool to monitor, document, and compare bias in LLMs. AIdas allows users to plug in the same prompt into up to four of the latest commercial models, including OpenAI’s GPT, Anthropic’s Claude Opus, xAI’s Grok, and Deepseek. Users code the responses on criteria like source diversity or whether the response is “neutral” or leans toward one perspective. The coded data can then be copied into a spreadsheet or exported as a CSV file.
Factnameh, a fact-checking organization for Persian news and social media, has already published some early research using the tool. Its team audited narratives about the U.S.-Iran War in popular models and sketched out some helpful frameworks for talking about model bias.
That includes the idea of “political cueing” — whether a model takes on the ideological premise of a user’s prompt. Factnameh found that Gemini was the most likely to mirror a prompt, labeling it an “echo” that “turns up the volume” on bias embedded in a user’s query. Meanwhile, the researchers labeled Claude as the “prosecutor” for pushing back on ideologically loaded prompts and citing the most diverse sources.
Borderless disinformation requires a cross-border response
Fact-checking organizations from across Central and Eastern Europe took center stage throughout the conference. One message was repeated by several speakers, including journalists from Hungary, Lithuania, Poland and Estonia: Disinformation campaigns cross borders, and so should fact-checking initiatives.
“We need to remember that Russia doesn’t treat Estonia as one specific target. It doesn’t take it country by country. It takes us as part of a much larger theater,” said Holger Roonemaa, the head of investigations and fact-checking at Delfi Estonia. Roonemaa said that in recent years, the coordinated campaigns Delfi tracks have spread beyond the Baltic region. “What’s important in Estonia can be equally important in Norway, Sweden, Germany, or France.”
Cross-border investigations, co-publishing, knowledge sharing, and centralized resources were all named as strategies essential to countering Russian disinformation across Europe.
New models for cross-border collaboration were also at the core of a panel I moderated about deepfakes and the challenges newsrooms in Global Majority countries face in accessing quality AI detection tools. (I did not receive compensation or travel support from IFCN.)
On the panel was Pamposh Raina, the head of the Deepfakes Analysis Unit (DAU), an India-based initiative that investigates suspected AI-generated audio and video material. IFCN-certified fact-checking organizations can submit potential deepfakes directly to the DAU, an offshoot of the Trusted Information Alliance, and the general public can submit them via a WhatsApp tipline. The nonprofit project has centralized verification expertise and access to detection tools for newsrooms across India, including partners like Factly, The Quint, and BOOM. When the DAU identifies a deepfake or AI-manipulated media, it publishes a report and full methodology of its analysis on its site for anyone to read or cite.
Last year, the DAU opened its escalation channels to any IFCN-certified organization around the world. Two journalists from Southeast Asia on our panel and have already collaborated with the DAU — Celine Samson from VeraFiles in the Philippines, and Ika Ningtyas from Tempo in Indonesia. Both newsrooms leaned on the DAU in the process of reporting high-profile fact-checks, including VeraFiles’ attempt to authenticate a damning video of the president of the Philippines snorting a white powder. Raina said that since opening up its channels, the DAU has also supported newsrooms as far away as Japan and Georgia.
On another panel, Szabolcs Panyi, an investigative editor for VSquare in Hungary, summed up the sentiment from speakers across the conference: “These bad guys are cooperating. We should be cooperating, too.”
