Today's Headlines

  • A hallucinated AI report brings US forces to the edge of boarding a Chinese ship, CNN reports, citing four sources familiar with the episode
  • The Federal Register website briefly offers an Alibaba Qwen model to search public comments, then removes it on Wednesday
  • The FAA prepares to launch SMART, an $875 million AI tool, over the Washington, DC, airspace before a nationwide rollout
  • California Gov. Gavin Newsom signs an executive order that puts a frontier-model "kill switch" on the table
  • Virginia Gov. Abigail Spanberger signs an executive order creating an AI task force and tightening the rules on data centers

All three of today's stories come back to how government is using AI. An intelligence analysis that hallucinated and nearly sent armed personnel onto a Chinese vessel, and a government site publishing federal rules that was running a Chinese model the FBI had named, landed on the same day.

On that same day, the government kept buying bigger AI. The FAA is moving to switch on a nationwide traffic-prediction system over Washington, DC, as part of a 12-year, $875 million contract.

Today's Top Three

A hallucinated AI report brings US forces to the edge of boarding a Chinese ship

CNN reported on September 18 that the US military came close to boarding a Chinese ship this spring on the basis of an intelligence report that was "entirely false" and had been generated with the help of AI.

The report described nuclear weapons components. It said a Chinese ship in the Middle East was carrying components of a nuclear weapons program, and it circulated across the US military in the midst of the war with Iran, according to the four sources CNN cites.

Preparations reached the boarding stage. Two of the sources told CNN that armed members of the US military were preparing to board the ship, and another source said military planes were already in the air.

The reversal came just before the planned operation. Officials dug into the report, found that a US Special Operations Command analyst had produced it with the help of AI, and found that a chatbot the analyst used had inaccurately identified the material the ship was carrying. CNN was unable to learn what the misidentified cargo was.

The chatbot blended two kinds of material. It fused open-source intelligence with secret signals intelligence in government holdings to reach its conclusion about the cargo, CNN reported, and the analyst then used AI again to package the finding into a standard intelligence report and disseminate it. The underlying reporting on the ship's manifest originated with US Special Operations Command Pacific, based in Hawaii.

Whether the tool was commercial or government-built remains open. CNN could not establish which it was, and a former senior US official familiar with the AI systems used by military and intelligence analysts told CNN that "the internal tools are mostly just copies of the commercial stuff wearing lipstick."

One source told CNN the episode "almost started a war." Any US operation against a Chinese vessel could have risked spiraling into an armed conflict between the two countries, CNN wrote.

The agencies involved have stayed silent. US Special Operations Command Pacific and the Pentagon did not respond to CNN's request for comment.

Adoption, meanwhile, is speeding up. Defense Secretary Pete Hegseth released an Artificial Intelligence Acceleration Strategy in January, with an accompanying memo describing the aim as "democratizing AI experimentation and transformation across the Department by putting America's world-leading AI models directly in the hands of our three million civilian and military personnel, at all classification levels." Multiple US officials told CNN the effort is decentralized, with different parts of government using different tools under different orders, and with no single standard for how the US verifies what those tools generate.

Targeting is where sources put the risk. "AI in targeting is definitely something that is ramping up and there is no real guidance for how having a human in the loop will prevent civilian casualties or fratricide," another source familiar with current military policy told CNN. One source also said hallucinations like this one have recurred across the intelligence community since the tools began proliferating.

The procurement continues. Ars Technica notes that the Department of Defense said last December it would build its GenAI.mil platform on Google's Gemini for Government and added Grok for Government as an option last month. In June a Pentagon representative told Congress the department uses generative AI to help produce congressionally mandated reports, and that 1.5 million active DoD personnel have used the military's generative AI tools.

The Federal Register briefly runs a Chinese model the FBI had named

The Federal Register website, run by the National Archives, briefly offered visitors an Alibaba Qwen model as a way to search public comments on proposed regulations, and officials removed it on Wednesday.

Reuters reported the removal. Ars Technica says the change followed social media users pointing out the contradiction, and Reuters found that an archived version of the site's source code confirmed the Qwen model came down on Wednesday.

The contradiction runs through an FBI allegation. Earlier this month the FBI named Alibaba among six leading Chinese firms it says are conducting "industrial-scale distillation," which the agency argues is helping America's biggest competitor cut costs and development time in the AI race.

The start date remains unknown. It is unclear when the National Archives began offering the option, and the agency, the White House and the FBI have all stayed silent, with the Archives also declining to respond to Ars Technica.

A screenshot brought it to light. A user with the handle "tleilax___" posted a widely shared screenshot of the option displayed on the government site on September 15, which means the service was live for at least a day before Wednesday's removal.

Experts put the security risk low. Georgetown University Law Professor Anupam Chander told Reuters that the answer depends on how the model was trained, and that the site's content is "already public, so the model was not working with sensitive government information." The Chinese news site Sina.com reported that the Federal Register used a small open-weight model at the "Qwen3 0.6B level" that serves solely to retrieve documents and keeps government data away from Alibaba and any third party.

Reaction splits along policy lines. Daniel Castro, president of the Information Technology and Innovation Foundation, told Reuters it is an "insane" disconnect for a US agency to use an Alibaba model while the FBI urges stakeholders toward American models. Rep. John Moolenaar (R-Mich.), who chairs the House China Committee, holds that no federal government entity should use a Chinese AI model, telling Reuters that doing so only deepens federal dependence on Chinese AI against the national interest.

Some analysts question the policy frame itself. Research submitted to the US-China Economic and Security Review Commission in March found that US export controls are calibrated to constrain frontier training by restricting access to advanced semiconductors, and leave the small-model deployment cycle untouched. If the models that matter most for industrial AI are small, specialized and open, the researchers wrote, the current US policy framework could be targeting the wrong layer of the competition.

The FAA lines up an $875 million AI tool, starting over Washington

The FAA is preparing to switch on SMART, an AI system that advises air traffic controllers, in the congested airspace above the Washington, DC, area.

The system's job is prediction. The FAA describes SMART as using AI models to predict air traffic flows and identify potential conflicts from operational factors such as airline schedules, weather, airport capacity and airspace conditions.

The launch date comes from The Wall Street Journal. US government and industry officials told the Journal that SMART could debut for the three major airports in the Washington, DC, area as soon as Monday, September 21.

This is the first step toward a national rollout. The FAA oversees 29 million square miles of US national airspace, and the DC deployment is positioned as the opening stage of covering all of it.

The contract runs 12 years and $875 million. SMART is part of an award made in June to the Boston-based company Air Space Intelligence, which also covers a Flow Management Data and Services system meant to replace the current one at the FAA's Air Traffic Control System Command Center in Virginia.

Narrowing the scope drew approval. Philip Mann, principal consultant at Vector Strategic Consulting, called starting small before going nationwide the "right call." Mann worked at the FAA in multiple roles for 17 years and now advises on aviation safety and AI governance.

He also located the risk. "SMART's risk was never any single prediction—it is a national-scale system with AI components carrying more unknowns than anything the FAA has fielded," Mann wrote in an email to Ars Technica. "Every cut in scope shrinks the unknowns."

Airlines spent weeks confused. Politico reported that the industry struggled to follow the agency's deployment plans, and anonymous airline officials said their concerns eased after the FAA said SMART would leave controller and airline procedures as they are and instead produce "alternative route information" delivered through existing FAA systems.

What runs underneath stays undisclosed. The types of AI models in SMART have gone unstated, and Ars reached out to the FAA for comment. Mann says he wants to know whether the rollout remains limited to aircraft flying at 24,000 feet and above, as the FAA described in June.

He is watching what unlocks the next expansion. "What I will be watching is what evidence gates the next expansion: performance under real workload and degraded data rather than demonstration conditions, and a written answer to who owns the outcome when a prediction is wrong," Mann told Ars.

A controller shortage sits behind all of it. The air traffic controller workforce shrank by 6 percent over the past decade, according to a US Government Accountability Office report published in December 2025, and the FAA has only just started replacing hundreds of radar systems dating to the 1980s along with radios, telecommunication lines and voice switches.

Other Developments

Policy and Regulation

  • California Gov. Gavin Newsom issued an executive order on September 18 positioning the state to lead on AI oversight. The order directs the state to convene a group of experts that will deliver recommendations within two months on strengthening AI safety measures in state law. Newsom wants the group to consider requiring AI companies to embed independent verification groups onsite for regular audits, subjecting transparency reports and risk assessments to independent auditor standards, creating a "kill switch" that is routinely verified as effective, and requiring companies to report "loss-of-control incidents" like the OpenAI attack on Hugging Face as critical safety incidents. He is calling on Congress and President Trump to review and adopt the state's framework, or use its regulation as "a floor, not a ceiling, for the benefit and safety of all Americans." Gavin Newsom is pushing for an AI kill switch (The Verge)
  • Virginia Gov. Abigail Spanberger signed Executive Order 22 on the same day, addressing data centers and AI risks in the state that hosts the data center capital of the world. The order bans executive branch officials from signing nondisclosure agreements for data center projects, requires expedited noise regulations, and orders a review of the backup generation that data centers run. It also establishes an AI task force to evaluate how state government can address risks to Virginians such as workforce displacement and data privacy, and how existing law applies to AI harms. Spanberger paired it with a Data Center Accountability Framework that calls for eliminating by-right approval of the kind Loudoun County used for years, removing some state subsidies, setting environmental guardrails and shielding residents from data-center-driven energy prices. Virginia governor creates an AI task force and moves to restrain data centers (The Verge)

Models and APIs

  • TypeSafe AI, founded by Diogo Almeida, an OpenAI researcher who helped build ChatGPT and invent reinforcement learning from human feedback, released a transformer-based model called Jev this week that outputs no text. Jev returns probabilities, which the company calls "calibrated decisions," and because users define the possible outputs in advance, the model cannot hallucinate. Output tokens are free and input tokens are metered by the billion rather than the million. Demand ran high enough that the company briefly lost the ability to serve its API, and a Vercel software engineer said that swapping OpenAI's Luna 5.6 for Jev in a classifier that reviews commands for safety returned results 5 to 18 times faster and with greater accuracy. A new kind of AI model from a ChatGPT inventor is thrilling developers (TechCrunch)

Business

  • Disney hired Karandeep Anand, the former CEO of Character.AI, as its first chief technology officer. Disney sent Character.AI a cease and desist letter in September 2025 accusing the company of infringing on its characters, and the executive who ran that startup now takes Disney's top technology job. Variety reports that Anand was chosen by Josh D'Amaro, who became CEO in March after Bob Iger stepped down. Anand advised Character.AI's board before becoming CEO in May 2025, worked at Facebook from 2015 to 2021, and spent 15 years at Microsoft before that. Disney's first CTO led an AI startup it once accused of copying its characters (TechCrunch)

Research

  • Google expanded its AI & Economy Research Program, which tracks how AI affects jobs and productivity, by bringing in a group of economists. Philippe Aghion, the 2025 Nobel Laureate in Economics and a professor at INSEAD and the Collège de France, joins as an Academic Advisor, and Ajay Agrawal of the University of Toronto's Rotman School of Management joins as a Visiting Fellow. Anu Madgavkar, formerly a partner at the McKinsey Global Institute, and Daniel Rock come in as Directors of the program, working alongside Alex Imas, Director of AGI Economics at Google DeepMind. The expanded team will feed future updates of the AI & Economy ATLAS, the open-access dataset on how people use Google's AI tools that the company released as v1.0. New experts join Google's AI & Economy team (Google)

Source: selected by the editors from the AI news inbox collected on September 19, 2026 (29 items, 16 primary and 13 secondary).