AI did not invent the false claim. It gave the false claim a faster press — and it handed the fact-checker the same machine. In 2025, the Brazilian fact-checking organization Aos Fatos watched AI-generated content climb from 7% to 16% of the claims it investigated in a single year [1]. The same organization spent that year building its own AI tools. Nothing about the underlying discipline changed. What changed is the volume running through both sides.That distinction — volume vs. substance. It is worth working through carefully, because the year Aos Fatos recently finished is a test case being played out in newsrooms globally.What Actually Doubled?Start with the number itself. Aos Fatos fact-checked 619 claims in 2025. 16% of them involved AI-generated content, up from 7% the year before — more than double in one year, driven by fabricated images and video rather than fabricated text. Reuters Institute researchers, reporting the figure at a 2026 industry gathering on AI and the future of news, framed it as one data point inside a larger shift: AI-powered disinformation in Brazil reached over 32 million views on TikTok in 2025, with 2.1 million additional likes and shares across Facebook and Instagram tied to the same wave of content [1].Those are real numbers, and they describe a real change. Producing a convincing fake image at scale has become radically cheaper over the past two years. That is not in dispute, and nothing in this piece argues otherwise. Newsrooms that ignore this shift are negligent.But notice what the number does not show. It does not show that falsehood itself got any easier to sustain once someone actually looked. A fabricated image is not true because it is well-made, any more than a forty-footnote report is more accurate than a four-footnote one. The image still fails the same test it always has: does it hold up when someone traces its origin and checks what it claims? Volume changed the odds that any single fake reaches a reader before it is checked. It did not change what checking finds when it happens.Same Instrument, Opposite VerdictsHere is the part of the story that gets less attention than it deserves. Aos Fatos did not just absorb the increase — it built against it. The organization developed Fátima, an AI chatbot that answers audience questions about claims already in its database, and is building a companion tool, Busca Fatos, for real-time verification during live coverage [1]. Two other fact-checking organizations cited at the same gathering, Spain's Maldita and the UK's Full Fact, have gone further: both run large language models that scan and classify claims across millions of sentences of public discourse, flagging the ones worth a human's time before a person ever sees them.That is the same underlying technology on both sides. One version of it manufactures a fabricated image cheaply enough to flood a feed. Another version of it reads a million sentences overnight and hands a fact-checker the 12 worth investigating. Neither use case is more "natural" than the other. Technology has no preference for truth or falsehood; it accelerates whatever instructions and judgments sit behind it. A newsroom that trains a model to prioritize harmful narratives ends up with a filter. A bad actor who prompts the same class of model to generate a convincing fake gets a pollutant. The mechanism does not decide which one it becomes. The person operating it does.As is my practice, I give credit where it is due. Maldita, Full Fact, and Aos Fatos are doing genuine, difficult, good-faith work, and the tools they have built represent real progress against a real problem. None of it would exist without practitioners choosing to point the same capability at detection instead of deception. That choice is the entire story — not the capability itself.Who Still Has to Push Publish?Chris Morris, CEO of Full Fact, and one of the panelists presenting the Aos Fatos figures, described the risk well: newsrooms are "in danger of getting to a place where no one believes anything they'd read or see or hear anywhere" [1]. That is the failure mode worth taking seriously, and it is not a technology failure. It is a discipline failure, and it can happen with or without AI in the loop.Consider the two ways this plays out. In the first, a fact-checking team runs a Maldita-style classifier overnight, gets a shortlist of twelve claims worth investigating out of a million sentences, and still sends a person to trace each one back to a source, test the method behind it, and ask who benefits from it being believed. The tool changed how the team spent its morning. It did not change what counted as verification. In the second, a newsroom — or a reader, or a company evaluating a vendor's claim — sees an AI-generated shortlist, a citation, a polished report, and treats the artifact's existence as proof of the claim it contains. The tool did the same work in both cases. Only one of those newsrooms is still doing its job.That test applies at any scale. A reader deciding whether to share a video, a fact-checking desk triaging a million sentences, a company evaluating a benchmark slide before it reaches the board — all three are answering the same question with the same three ingredients: is the origin traceable, does the reasoning hold up under a second look, and does whoever is asking benefit from the answer being believed. AI has made all three questions faster to ask and faster to dodge. It has not changed which one actually needs to be answered.The PointAos Fatos's number will keep climbing. So will the sophistication of the tools built to answer it, on both sides of that fight. Betting that better detection technology eventually wins the race outright is a mistake — Maldita and Full Fact's own systems still hand the final judgment to a person, by design, because the technology was never built to make that call alone [1]. The organizations doing this well are not the ones with the most advanced models. They are the ones who never stopped treating the model's output as a lead to check rather than a verdict to publish.That is not a hopeful prediction or a cautious hedge. It is the same principle discussed in previous articles … opened with and now tested against a full year of real data from a newsroom that lived through both sides of it at once: the questions we ask determine the answers we get, and no volume of AI-generated content — real or fabricated — changes who is responsible for asking them.Citation(s)Citation 1 — Reuters Institute for the Study of Journalism, "AI and the Future of News 2026: what we learned about its impact on newsrooms, fact-checking and news coverage." Primary source for all Aos Fatos figures (619 claims fact-checked in 2025; AI-generated content share rising from 7% to 16%; the 32 million TikTok views / 2.1 million Facebook–Instagram interactions figures), the Fátima and Busca Fatos tools, Maldita's and Full Fact's claim-classification systems, and the Chris Morris quote.