Inside the rise of AI-native newsrooms run by autonomous agents

The idea of newsrooms run by AI agents might be a nightmare for many journalists. But these newsrooms are no longer hypotheticals. They already exist. While many outlets are still figuring out how to integrate AI into their workflows, a new breed of projects is building those workflows around AI from the ground up.

Unlike traditional news aggregators, which primarily collect and republish information from existing sources, these so-called AI-native newsrooms are attempting to replicate parts of the editorial process itself. These systems are built so they can monitor sources, research, cross-check information, write articles and even make editorial decisions themselves.

The extent of human involvement varies widely by project. Some of these outlets have editors reviewing every story, while others leave almost the entire editorial process to the machines.

Jindex.ai, a database of AI projects built by Ukrainian journalist Katya Gorchinskaya, has recorded approximately 30 of these outlets worldwide. For this piece, I spoke to five of the humans behind these operations in Canada, Croatia, Italy and the United States. They explained how they built these ventures, how they work, and what they might mean for the future of news.

1. The AI-native newsroom that broke the news about OpenAI’s hack

Earlier this year, news broke that over 1,200 AI agents within OpenAI went rogue. The agents had started to communicate with each other to autonomously hack Hugging Face. Details surrounding this incident were revealed by OpenAI itself at the Black Hat security conference in Las Vegas. But the outlet that broke the news to the public first didn’t have anyone at the conference.

RuntimeWire, a US-based AI-native outlet founded in May by technologist Ryan Merket, published the story three hours before WIRED and even before major global newspapers. Here’s how it happened.

At the Black Hat cybersecurity conference, Merket told me, he spotted a significant disclosure in a live presentation transcript and fed it into his AI newsroom. The synthetic team cross-referenced the speakers, official materials, and existing incident records. It was Merket himself who made the call that the disclosure was newsworthy. In his view, the public record was strong enough to publish the story. The first version went live roughly six minutes after he fed the transcript to the AI system.

“That process is fast because the underlying reporting infrastructure is built before the signal arrives,” he said. “However, the story still gets held if the evidence is thin, if the account cannot be verified, or if a central fact cannot be attributed.”

Merket pairs a two-decade career in tech startups with an early interest in journalism, after building and running his college’s first online newspaper. He is not just the founder of RuntimeWire, but also its editor-in-chief. He sees the outlet as a real newsroom with an editorial mission, a reporting process, sourcing standards, and a correction process. All the bylines are Merket’s because, as he explained, he is the one answering for every story. 

“[My editorial responsibility] begins before an article even exists: building the system, choosing its priorities, setting its standards, deciding what can be published automatically, and designing the safeguards that halt publication,” he said. “The operating model is unusual, but the accountability is straightforward: if a story is wrong, unfair, misleading, or insufficiently sourced, it is my problem to fix.”

2. The Italian newsroom with no human in the loop

A system of guardrails is in place for all the publications I spoke to. But human intervention is very thin or non-existent for some of them.

A good example is Redazione Zero, an Italian general-news AI-native outlet where news selection and writing are left entirely to autonomous multi-agent workflows. Its publisher is software developer and entrepreneur Emilio Salmeri, who explained to me how the publication works.

Redazione Zero is built around a series of independent AI agents, rather than a single model. Every morning, two agents using Anthropic and OpenAI models independently gather news, which is then filtered through automated checks.

Once curation is done, the system selects and writes stories following Redazione Zero’s public editorial charter, before separate AI verifiers fact-check the articles and block publication if they detect something wrong. A final AI editor writes the editorial explaining the day’s choices. Salmeri himself sets the rules and safeguards but does not select, review or approve stories before publication.

“Responsibility for what goes out today is mine,” Salmeri said. “The decisions are taken by the system; the consequences are borne by me. And that is why the hardest rules of the charter sit exactly where my responsibility is greatest.”

For example, Redazione Zero prohibits crime and court reporting altogether and imposes strict limits on health coverage, including a ban on medical advice and dosage information. Minors must never be identifiable, while stories generally require two independent sources, with any exceptions explicitly disclosed. Like the other publications covered in this piece, Redazione Zero allows readers to formally challenge articles, with a response promised within 48 hours.

“The decision – correction, withdrawal or reasoned rejection – is mine and never the system’s,” Salmeri said. “The system writes, but it does not rule on its own conduct.”

3. Can AI-native oases alleviate news deserts?

Both Merket and Salmeri created their newsrooms out of an admiration for journalism and an interest in the bounds of artificial intelligence. But some of the founders I spoke to were motivated by the lack of local news in their own communities.

Tihomir Jauk is a Croatian AI developer and software engineer who created the AI-native outlet Moj Kraj (‘My area’ in Croatian) out of a combination of technical curiosity and a desire to fill out the gaps of hyperlocal news coverage in the country. The outlet covers over 124 localities by scanning general news but also over 2,000 smaller sources such as municipal and district websites, government announcements, utility providers, traffic updates and weather information.

“All that information is collected, geotagged and then processed to ensure it’s relevant, reliable and not just AI slop,” Jauk told me. “Large, established news portals have much higher costs, with journalists and entire teams to pay, so they often don’t have the time or resources to cover smaller sources and communities. That’s where we come in.”

Before ingesting the content, there’s an AI agent that pre-screens and analyses each source’s terms of service and privacy policies for usage restrictions and attribution requirements. Draft articles then undergo automated checks for grammar, tone and neutrality, with those falling below thresholds rejected. As a result, only 20–30 of around 100 daily drafts are published.

“Once an article is drafted, it goes through a series of verification steps. Each article is assessed against a set of questions, such as whether it is sufficiently neutral, and given numerical scores. We then apply thresholds: for example, if an article scores below 80% for neutrality, we don’t publish it,” he said.

Having guardrails in place is something veteran technology journalist Nicholas De Leon takes very seriously. His outlet, Tucson Daily Brief, has a two-tiered publishing model. Routine RSS aggregations and simple public directory logs publish automatically while higher-stakes coverage undergoes manual review before going live.

Like many of his counterparts, De Leon launched this project out of a mixture of personal tech curiosity, a desire to test AI’s reporting capabilities, and a drive to fill local news gaps in southern Arizona, where he is based.

“Tucson has one newspaper left and a couple of TV stations,” he told me. “A lot of the debate around AI journalism is about whether it’s taking jobs. But the point of this isn’t to prove that I can take anyone’s job or do it better. It’s about what more I can do.”

De Leon’s initial focus was covering town hall meetings. He developed a “live AI reporter” to automatically capture livestream audio, transcribe it in real time, and draft articles based on those transcripts.

The system could cover multiple meetings simultaneously, expanding reporting capacity without additional journalists. De Leon still edits and fact-checks every draft of this coverage before publication, and he said that AI-written articles are already around 90% publication-ready.

De Leon has also developed a daily two-minute podcast using an AI clone of his own voice, a Spanish-language TikTok channel featuring an AI anchor named “Patricia”, weekly local crossword puzzles, and a RAG search engine called ChatTDB, which retrieves relevant information from the Tucson Daily Brief archive and uses AI to generate answers grounded in that information.

Despite the amount of time and effort, the operation is not yet profitable. Like all of the other outlets I mentioned so far, this is a side project driven by curiosity over what AI can do. De Leon stresses how big outlets often prioritise national and international news over everyday local information, such as new restaurant permits, liquor licence filings, zoning changes, and health inspection logs, which directly affect the community he lives in.

“To me, that’s always been the most interesting part of journalism. I can’t really affect what the president says or what’s happening elsewhere in the world, but I find it useful to know when a new restaurant is opening or a street is closed for construction,” De Leon said. “That kind of information has always been incredibly relevant and interesting to me.”

4. The Canadian news brand revived by AI

In 2023 Canadian media group Métro Media declared bankruptcy, bringing more than 30 hyperlocal news operations to a halt. One of them was its flagship newspaper, Journal Métro, and its 16 print weeklies dedicated to covering local government and community news in Montreal and Quebec City.

Two years later, technology entrepreneur Pierre-Antoine Fradet acquired Journal Métro’s intellectual property with the aim of restoring local news coverage in Montreal. To make the relaunch financially sustainable, the publication adopted an AI-driven model, automating routine reporting and keeping overhead costs low.

The publication hired just one human: Olivier Robichaud, a former Journal Métro journalist who had been laid off a year before the bankruptcy and who was appointed editor-in-chief.

Journal Métro uses AI tools to monitor trusted sources for new press releases, reports and statistics, automatically generating draft articles from the information it finds. Robichaud then verifies the facts, adds context and, where necessary, seeks opposing viewpoints before approving the stories for publication.

“The idea is to use AI to do the easy stuff so that a flesh-and-bone journalist can do the hard stuff,” Robichaud said. “With AI, I could produce 20 or 25 of those [straightforward] articles and still have time to work on a piece of original reporting. The goal is to spend fewer resources on simple news stories that don’t necessarily offer much return on investment.”

Like all the publications covered in this piece, Journal Métro clearly labels whether an article was written by Robichaud or its AI system. As Robichaud puts it, there is always a “human in the loop” before, during and after publication.

5. What does this mean for the future of news?

Despite the guardrails, editorial standards and disclosures, the rise of AI-native newsrooms raises uncomfortable questions. Are they further muddying an already polluted information ecosystem, where audiences increasingly struggle to distinguish human-generated content from AI-generated material? What happens to the journalists whose jobs these systems might displace? And, perhaps most importantly, what does all this mean for journalism itself?

Robichaud, who has been a journalist for over 15 years, acknowledges that AI will probably replace jobs in the industry, but he argues that the outcomes depend entirely on employers’ choices.

“Some newsrooms might decide they are cutting journalists,” said Robichaud. “But news outlets might decide that’s not how they want to use AI. They might use it differently and keep their journalists focused on the work they consider important, and that readers and listeners expect from their particular brand.”

Robichaud points to Journal Métro itself as an example. When the publication went bankrupt, its journalists lost their jobs and local communities lost a source of news. Now, with AI helping to make the operation financially viable, at least one journalism job has been restored (his own) along with access to local reporting.

“Montreal was losing a source of local and hyperlocal news, and that was creating not a news desert, but a certain lack in the type of coverage that people used to have access to,” he said.

De Leon from the Tucson Daily Brief, who’s been a journalist for more than 20 years, wants to reframe his project as “additive” rather than subtractive. In a shrinking local media market where no one is covering surrounding municipal town halls, he is expanding total local coverage rather than taking away anyone’s job.

“There’s a lot of stuff here that goes uncovered, like local town hall meetings in Marana or Oro Valley. These meetings aren’t necessarily covered week in, week out,” explained De Leon. “So I’m trying to add to the local news coverage. I’m not taking away anyone else’s ability to cover those meetings.”

The limitations of these systems are not just theoretical. All the people I interviewed for this piece acknowledged mistakes made by their AI newsrooms.

At Journal Métro, Robichaud sometimes receives AI-generated drafts that fail to grasp the right angle or tackle subjects which are too complex or sensitive for the system. Meanwhile, Jauk's system discards 70% to 80% of the roughly 100 articles it drafts daily because they fail automated quality checks, including assessments of neutrality and tone.

Even Redazione Zero, which operates without human review before publication, has published four corrections involving references to unpublished content, a misinterpreted fact-check, a column that failed to meet editorial standards and outdated information. Each error has prompted new safeguards and verification checks.

“What I can do is make every error observable and make sure it leaves a scar in the machine,” Salmeri told me.

Despite these shortcomings, the technologists I spoke to believe AI-native newsrooms have a role to play. They argued that their aim is not to replace journalists or suggest that AI can replicate everything human reporters do. Even Salmeri, whose newsroom is perhaps the most autonomous, acknowledges AI’s limitations and the irreplaceable value of journalists who cultivate sources and report from the field.

“It would not replace the journalist who finds a source or goes to the scene. It would be another form of reporting, and we are still at the stage of trying to understand how well it can work,” he said.

Similarly, Croatia's Jauk and RuntimeWire's Merket argue that machines should take on the repetitive work of gathering and processing information, freeing journalists to focus on reporting that AI cannot replicate.

“I do not believe the future is an empty newsroom where software produces an endless stream of pages,” Merket said. “It is a newsroom where machines carry the mechanical load, holding humans to a higher standard for the decisions that actually matter.”