When AI Pitches the Scoop, the Fact-Check Still Wins

On a recent afternoon on a foreign desk, an editor thought she had found a scoop that tied several of the world’s most volatile crises together.

An artificial intelligence tool had produced a polished pitch: a secret United Nations–backed plan, brokered by Oman and the United Arab Emirates, to defuse an escalating confrontation between Iran and Israel. The lead came with names, dates and even a URL for what it described as a UN document outlining the deal.

When researchers went to verify it, the URL did not resolve. The document was nowhere in the UN’s database. Major wire services and official government sites showed no trace of the supposed summit. The more they checked, the less of the story existed.

That phantom agreement was one of seven AI-generated geopolitical leads that a newsroom recently put through a standard verification process — and ultimately rejected, one after another, for failing the most basic test of accuracy.

A convincing voice — with invented specifics

The episode illustrates both the power and the limits of generative AI in hard-news reporting. The system was adept at mimicking the patterns of real events in some of the world’s most sensitive regions. But it also fabricated specifics: dates, documents, vote counts and diplomatic formats that evaporated under scrutiny.

Editors involved in the review treated the AI’s work like any other unvetted tip. Each lead was checked against primary or near-primary sources, including official government releases, UN records, court rulings and major wire reports. None could be fully corroborated.

In some cases, the AI got basic political facts wrong, such as identifying the wrong U.S. president for the date of a supposed diplomatic initiative. In others, it cited documents and web addresses that did not exist. Several leads blended real geopolitical dynamics — Iranian protests, Israeli settlement expansion in the West Bank, mediation by Gulf states, and contested elections in West Africa — with invented details, such as a named “revolution,” a trilateral summit format or a precise parliamentary seat split.

Real-world patterns made the falsehoods plausible

The patterns the AI leaned on are well documented.

In the Israeli-occupied West Bank, for example, Prime Minister Benjamin Netanyahu’s government has significantly expanded settlement activity over the past several years. In May 2025, the cabinet advanced plans for 22 new settlements, including legalizing previously unauthorized outposts. In December of that year, it approved 19 more. Early this month, Israel’s security cabinet moved to extend Israeli authority over land, environmental and archaeological issues in parts of the territory that the Oslo accords had placed under Palestinian civil administration. Critics inside and outside Israel have described the steps as “annexation in all but name.”

Any formal declaration of annexation, or new UN sanctions in response, would be a major development. The AI system suggested just such measures, attaching them to specific dates and votes. But researchers found no record of those decisions in cabinet communiqués, Knesset archives or UN Security Council proceedings.

A similar pattern emerged around Iran.

On Dec. 28, 2025, protests broke out across Iran over a collapsing currency and soaring prices, quickly widening into demands for political change. Human rights groups say security forces responded with live ammunition, metal pellets and mass arrests. They have described the ensuing crackdown as the deadliest protest repression in the Islamic Republic’s history. Authorities imposed a near-total internet shutdown starting Jan. 8, making independent reporting difficult and casualty counts uncertain.

Against that backdrop, AI-generated leads portrayed what they called a named “revolution” and alleged that UN bodies had adopted specific resolutions on precise dates, using language that does not appear in any publicly available UN document. Amnesty International, Human Rights Watch and UN officials have condemned the violence and urged investigations, but there is no record of the resolutions or legal findings the AI described.

The tool also drew on the Gulf’s well-established role as a diplomatic go-between. Oman, in particular, has long hosted quiet talks between Iran and the United States, including back-channel meetings that preceded the 2015 nuclear agreement. Since 2025 it has convened multiple rounds of indirect nuclear discussions in Muscat, some of them canceled or delayed after flare-ups involving Iran and Israel. The UAE has brokered prisoner exchanges and played a visible part in normalization deals and regional economic outreach.

Those real patterns made it easier for the system to produce a convincing — but ultimately fictitious — UN-UAE-Oman plan with a formal title and document number. None of those details could be confirmed in the UN’s treaty and document repositories or in official statements from the countries said to be involved.

In West Africa, the AI’s leads echoed a region struggling with repeated coups and fragile elections. Since 2020, military juntas have seized power in Mali, Burkina Faso, Niger and Guinea. Niger, Mali and Burkina Faso have announced their departure from the Economic Community of West African States and formed an alliance of Sahel states. In Guinea-Bissau, a November 2025 election was followed by unrest and another coup attempt. Benin’s Jan. 11, 2026 parliamentary election resulted in pro-government parties winning all 109 seats amid changes that extended presidential powers.

The AI tool reflected these realities, but also invented exact turnout rates, vote tallies and constitutional court rulings that local electoral commissions and courts have not recorded. In states where press freedom is limited and data are sparse, those false specifics are harder to spot at a glance — and potentially more inflammatory.

How newsrooms are drawing the line

News organizations worldwide are wrestling with how to integrate generative AI without sacrificing the verification standards that underpin their credibility.

The Associated Press, which has experimented with AI for internal tasks such as summarizing transcripts, instructs its journalists to treat any generative output as “unvetted source material.” AP’s guidance says AI “is not a replacement of journalists in any way” and warns staff about “the models’ tendency to generate inaccurate responses.” The agency does not allow AI tools to write publishable copy.

The New York Times has developed internal AI tools for summarizing documents and brainstorming headlines, but bars their use for drafting or significantly rewriting articles. In internal guidance, the Times has told staff that journalism must remain “reported, written and edited” by people, even when AI assists with background tasks.

Concern extends beyond newsrooms. UN Secretary-General António Guterres has warned that social media and artificial intelligence are helping to fuel what he has called “a tsunami of falsehoods.” In remarks launching a set of global principles on information integrity in 2024, he said that unchecked AI systems and opaque algorithms risk “destabilizing democratic systems” by amplifying disinformation and hate speech.

Lawmakers are beginning to respond. In New York, state legislators have proposed the New York FAIR News Act, which would require outlets to disclose when AI significantly assists in producing news content and mandate human editorial review before publication. Media unions have backed the measure as a way to safeguard jobs and standards. Some First Amendment scholars have argued that regulating internal newsroom tools could infringe on press freedom, underscoring the difficulty of writing rules for rapidly changing technology.

The lesson from seven killed stories

The newsroom that rejected all seven AI-generated leads did not publish a formal account of the episode. But the internal review offers a snapshot of how generative systems behave when they are asked not to summarize known facts, but to originate news in fast-moving, opaque environments.

In each case, the AI started from something real — Omani back-channels, Israeli settlement expansion, Sahel coups, Iranian crackdowns — and extrapolated forward into specifics that no official record supports. The result was content that sounded authoritative yet could not be tied to any verifiable event.

For the editors who killed the stories, the lesson was not that AI has no place in the newsroom, but that its proper role remains limited. As a way to map patterns, flag undercovered regions or assemble timelines from existing coverage, the tools may save time. As a source of original “scoops” on wars, revolutions and secret accords, they still demand the same skepticism applied to a tip from an unknown stranger.

In this case, after hours of checking records from New York to Muscat and Cotonou, the most accurate choice was to publish nothing at all.