When One AI Feeds Another, the Newsroom Becomes a Fact-Checking Battleground

On a recent January morning, a newsroom dashboard delivered an unusual verdict: not one of the day’s top story pitches met the basic standard for publication.

The pitches hadn’t come from reporters. They had been generated by an artificial intelligence system trained to suggest news leads on global events. A second automated system, configured to flag anything that could not be traced back to primary documents or reputable outlets, rejected every one.

The clash between the two programs — one spinning out plausible stories, the other discarding them as unverifiable — captures the uneasy place generative AI now occupies in journalism. Around the world, editors are quietly experimenting with software that can draft headlines, summarize complex crises and even propose full articles about wars, peace plans and natural disasters. Yet research suggests those tools regularly invent facts, including fake United Nations resolutions, disasters that never happened and detailed weapons lists in ongoing wars.

An internal review of the rejected leads found the same pattern. Each idea mixed real developments with fabricated or misattributed details: a Security Council resolution on Gaza with the wrong number, a deadly earthquake in a real region with no record of such a quake, a Russian missile barrage described with a level of precision not supported by any public source.

The gap between “almost right” and accurate is where the risks are growing.

Gaza peace plan collides with invented resolutions

Nowhere are the stakes clearer than in Gaza.

One AI‑generated brief claimed the U.N. Security Council had adopted “Resolution 2812 (2026)” to hand sweeping authority over the Gaza Strip to a new “Board of Peace” chaired by U.S. President Donald Trump. The summary sounded plausible: the Security Council has endorsed peace frameworks before, and Trump has announced a Board of Peace focused on Gaza.

But when editors searched U.N. records, there was no such resolution on Gaza under that number or year. Resolution 2812 exists, but it deals with maritime security in the Red Sea, not the Palestinian territory.

The real diplomatic framework for Gaza is anchored in a different measure: Security Council Resolution 2803, adopted in November 2025. That resolution endorses what the United States has called a “Comprehensive Plan to End the Gaza Conflict,” welcomes the establishment of an International Stabilization Force and “welcomes the creation of a Board of Peace” to help oversee the territory’s transition.

In mid‑January, Trump announced the Board of Peace in a statement, saying it was his “great honor to announce that THE BOARD OF PEACE has been formed” and naming himself chairman. The body’s Gaza‑focused executive structure includes Secretary of State Marco Rubio, former British Prime Minister Tony Blair, World Bank President Ajay Banga and Trump’s son‑in‑law, Jared Kushner, among others.

Under the Security Council framework, the Board is meant to coordinate reconstruction and governance support in Gaza alongside a Palestinian technocratic committee that runs day‑to‑day services.

The arrangement has drawn sharply divergent reactions.

The European Union has welcomed Resolution 2803 and signaled interest in a seat on the Board, casting it as a pragmatic route to stabilize Gaza and restore basic services after years of war.

Palestinian civil society groups and some U.N. human rights experts, by contrast, warn that the structure risks sidelining Palestinian self‑determination. Francesca Albanese, the U.N. special rapporteur on the occupied Palestinian territories, said the resolution threatened to “entrench external control” over Gaza and could amount to a new form of trusteeship.

Hamas has also condemned the plan, accusing international actors of replacing Israeli occupation with “foreign guardianship.”

In such a contested landscape, even small distortions — a wrong resolution number, invented subcommittees or imagined powers — can warp public understanding. When an AI‑generated lead confidently pairs a real Board of Peace with a non‑existent resolution, it becomes difficult, even for specialists, to separate what has been agreed in New York from what exists only in training data and statistical guesses.

Disasters that happened — and ones that exist only in code

The same dynamic played out in coverage of natural disasters.

Some of the rejected AI leads described destructive earthquakes and landslides in earthquake‑prone regions such as Guerrero, Mexico, and Gilgit‑Baltistan, Pakistan. The stories read like wire copy: precise magnitudes, time stamps, casualty counts and links to supposed online encyclopedias.

A search of seismic catalogs, national disaster agencies and local news archives turned up nothing to match the most detailed accounts. The locations are real and vulnerable. The described disasters were not.

At the same time, two major catastrophes that did happen — a magnitude 6.9 earthquake in the central Philippines in 2025 and deadly landslides in New Zealand in January — showed how easily AI systems can misstate or relocate real events.

On Sept. 30, 2025, a powerful offshore quake struck near the municipality of Daanbantayan in northern Cebu province. The Philippine Institute of Volcanology and Seismology later revised the magnitude to 6.9, making it the strongest recorded in that part of Cebu. Official tallies put the death toll at 79 with 1,271 injured and more than 748,000 people affected.

Nearly 1,000 pieces of infrastructure were damaged, including roads, bridges, schools and tourism sites. Over 12,700 aftershocks rattled the region in the weeks that followed. For communities already familiar with typhoons and past displacement, the earthquake underscored how fragile building standards and disaster planning remain.

In Tauranga, on New Zealand’s North Island, a severe storm on Jan. 22, 2026, saturated hillsides above coastal suburbs and campgrounds. Parts of Mount Maunganui, Welcome Bay and Papamoa were hit by sudden landslides. At a popular campground near the beach, caravans, tents and vehicles were buried or swept away. At least two people, including an overseas tourist, were killed there; more bodies were later recovered from a home in Welcome Bay. Authorities reported multiple people missing, including children, and declared local states of emergency.

Government ministers compared the battered east coast to a “war zone.” The disaster renewed debate over zoning, tourism‑campground safety and how to manage growing landslide risk as climate change brings more intense rainfall.

In testing, some AI systems replicated the broad outline of those events but supplied incorrect dates or magnitudes, moved them to nearby provinces or embellished casualty figures and damage based on generic patterns rather than specific reports. Others invented new quakes and floods in equally vulnerable locations, borrowing from the language and structure of real tragedies.

Experts say such fabrications are not merely hypothetical errors. Experimental studies have found that when language models generate explanations alongside false claims, readers are more likely to believe the misinformation than when they see the false claim alone.

For residents in regions like Guerrero or Gilgit‑Baltistan, where past disasters have left deep scars, a fictional catastrophe described in concrete detail can feel less like a technical glitch and more like an exploitation of real trauma.

Ukraine’s skies, and the lure of perfect numbers

On the war in Ukraine, the pattern was similar: real events framed with invented precision.

Rejected AI leads on recent Russian air attacks described elaborate “barrages” with exact numbers of each missile type — hypersonic Kinzhals, Kalibr cruise missiles, Iskander ballistic missiles, various drones — and a precise list of destroyed targets in several cities. Those details went beyond what Ukraine’s armed forces or Western defense officials had put on the record.

The underlying story was not in doubt. Since late 2025, Russia has repeatedly launched large, complex air campaigns against Ukraine’s energy grid and cities, combining waves of Shahed‑type drones with cruise and ballistic missiles.

On Jan. 9, 2026, Ukraine’s air force said it tracked 278 “air attack assets” in a single night, including 36 missiles and 242 drones, most of them aimed at Kyiv and surrounding regions. Ukrainian officials said they intercepted the vast majority, but 18 missiles and 16 drones got through, striking 19 sites and leaving nearly half of the capital’s apartment blocks temporarily without heat. Hundreds of thousands of residents lost power in the middle of winter.

Ten days later, on the night of Jan. 19–20, President Volodymyr Zelenskyy said Ukraine had expended nearly $100 million worth of interceptor missiles to repel another large‑scale assault involving 18 ballistic missiles, 15 cruise missiles and more than 300 drones.

“The cost of this defense is enormous,” Zelenskyy said in an address, arguing that Ukraine needed more air‑defense systems and ammunition from its allies to sustain its defenses.

Military analysts note that even basic figures in such attacks — total missiles launched, types used, number intercepted — are often disputed and refined over time. Exact breakdowns by weapon are difficult to verify publicly.

Language models, however, are designed to produce fluent, specific answers. Given a prompt about a Russian “combined missile and drone barrage,” they often output tables of weapons and impacts that look authoritative but are extrapolated from patterns in past coverage rather than sourced from new, verifiable reports.

In a war where daily attrition is measured in both lives and billions of dollars in equipment, a single misleading detail about how many hypersonic missiles were fired or how much the defense cost can color debates in foreign parliaments and living rooms far from the front.

A verification bottleneck — and a shifting competitive edge

Studies of large language models used on document‑based reporting tasks have found that roughly 30% of outputs contain at least one hallucinated detail. More open‑ended systems tend to be several times more error‑prone than tightly constrained tools.

At the same time, an analysis of more than 180,000 articles from about 1,500 U.S. newspapers found that nearly 9% showed signs of being partially or fully generated by AI systems, often without any disclosure to readers.

Major national outlets have adopted strict internal rules. Some permit reporters to use AI tools for tasks such as transcribing interviews, summarizing documents or experimenting with headlines, but they bar software from drafting or substantially rewriting copy without human oversight. Most policies stress that all factual claims must be checked against primary or established secondary sources.

Smaller newsrooms, particularly in the Global South, face a harder trade‑off. With fewer staff and limited budgets, they may be more likely to lean on generative tools to fill gaps in coverage or produce volume, while lacking the resources to systematically verify each AI‑suggested fact against agency cables, court filings or U.N. records.

The internal test that produced a slate of unusable leads — and a machine saying “no” to another machine — suggests one possible path forward: pairing generative systems with verification engines tuned to reject anything that cannot be cross‑checked.

That approach, news executives acknowledge, is slower and more resource‑intensive than simply publishing AI drafts with minimal editing. But on stories involving war, diplomacy and disaster, newsroom leaders say the alternative is worse.

In Gaza, the question is who will ultimately govern a devastated territory. In Ukraine, it is whether cities can keep the lights on under relentless bombardment. In Cebu and Tauranga, it is whether communities will be safer when the next quake or storm comes.

Those are not abstract problems, and they do not unfold on the same terms as a predictive text model. In an era when machines can describe an invented Security Council vote or a non‑existent landslide as if they had been there, the oldest rule in journalism — show your sources, or it does not run — is becoming not a nostalgic slogan but a competitive advantage.