PPPL to lead Genesis Mission project to create an AI operator for crucial fusion energy heating system

The research team will also create a digital twin of the beam-wave interactions inside a gyrotron

Princeton University

Facebook

XLinkedInWeChatBlueskyMessageWhatsAppEmail

 

image: 

PPPL's Syun'ichi Shiraiwa (center) will lead a Genesis Mission project to speed the development of gyrotrons needed for some fusion energy systems, working alongside PPPL colleagues Ahmed Diallo (left) and Álvaro Sánchez Villar (right). 

view more 

Credit: Photo credit: Michael Livingston / PPPL Communications Department

The U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) will lead a critical fusion energy research project and partner on seven other initiatives as a part of the Genesis Mission: the national initiative to harness artificial intelligence (AI) for scientific discovery. Announced July 22, 2026, these awards will support the Lab’s mission to harness on Earth the fusion energy that drives the sun and stars, while leading discoveries in plasma science and technology.

The Genesis Mission is a historic national initiative led by the U.S. Department of Energy, which is building the world’s most powerful integrated science discovery platform. By uniting government, industry, academia and philanthropy, it is accelerating breakthroughs in energy, scientific discovery and national security through a new platform that combines AI, supercomputing, quantum systems and advanced scientific instruments.

The goal of the Phase I Request for Application awards is to identify promising pathways toward transformative scientific capabilities and establish a foundation for future investment and scale. Project teams will design and demonstrate research workflows that integrate AI with scientific investigation, while rigorously evaluating whether those approaches can accelerate discovery, improve predictive capabilities, enhance experimentation or generate new scientific insights. 

Under the direction of Principal Research Physicist Syun’ichi Shiraiwa, the PPPL-led project will use AI to improve electron cyclotron heating (ECH): a crucial method for heating the plasma in fusion systems that use magnetic fields to hold the plasma. The research team will develop an AI system to autonomously operate the high-power gyrotrons used for ECH. The researchers will also build a digital twin, or virtual replica, to accelerate simulations of the physics inside a gyrotron and high-frequency gyrotron development needed by the fusion industry. Researchers from Rice University, the Naval Research Laboratory, Brookhaven National Laboratory and Sophelio will collaborate on this project.

In addition, two Genesis Mission projects will be led by researchers with joint appointments at PPPL and Princeton University:

• Toward a Facility for Laboratory Reconnection Experiments (FLARE) Digital Twin: Accelerating Experimental Research Using Neural Networks — Hantao Ji, a distinguished research fellow at PPPL, principal investigator on FLARE and professor of astrophysical sciences at Princeton University. 

• Foundation Model Platform for Fusion Energy: A Multi-modal Multidevice Learning — Egemen Kolemen, staff research physicist at PPPL and Princeton University professor of mechanical and aerospace engineering and the Andlinger Center for Energy and the Environment

PPPL was also selected as a partner on five other Genesis Mission projects. The following researchers will lead the Lab’s work on each:

• Core Accelerated Trajectories With Augmented Learning by Sim-to-experiment Transfer — Nathaniel Ferraro, deputy head of theory 

• Automatic Performance Portability for Distributed Scientific Workflows on the Genesis Mission Platform –– Shantenu Jha, head of computational sciences 

• AI-based Adaptive Low-temperature Plasma for Healthcare, Agriculture and Advanced Manufacturing –– Yevgeny Raitses, managing principal research physicist

• Data-driven Discovery of Multicomponent Plasma Fluid Dynamics –– Ammar Hakim, principal research physicist  

• AI-driven Rapid Discovery of Plasma Deposition for Novel Materials for Next-generation Semiconductor Devices –– Igor Kaganovich, principal research physicist

“The national labs and industry have a shared set of interests and priorities and very complementary capabilities, with laboratories delivering knowledge and industry delivering products. That really opens the aperture toward driving fusion and plasma science forward together,” said Laura Berzak Hopkins, PPPL’s Deputy Director for Research and Chief Research Officer.

For more information, visit PPPL's Genesis Mission project page.

PPPL is mastering the art of using plasma — the fourth state of matter — to solve some of the world’s toughest science and technology challenges. Nestled on Princeton University’s Forrestal Campus in Plainsboro, New Jersey, our research ignites innovation in a range of applications, including fusion energy, nanoscale fabrication, quantum materials and devices, and sustainability science. The University manages the Laboratory for the U.S. Department of Energy’s Office of Science, which is the nation’s single largest supporter of basic research in the physical sciences. Feel the heat at https://energy.gov/science and https://www.pppl.gov. 

 

USC leads national AI research project to accelerate scientific discovery

As part of the U.S. Department of Energy’s Genesis Mission, USC will lead a national research team developing a new kind of AI to predict turbulence — one of engineering’s oldest challenges — with applications across energy, transportation and man

University of Southern California

FacebookX

LinkedInWeChatBlueskyMessageWhatsAppEmail

USC is among the first universities selected to help lead the U.S. Department of Energy’s Genesis Mission, a national initiative that unites the country’s 17 national laboratories, universities and industry to explore artificial intelligence for scientific discovery.

As part of the initiative, USC will lead a multi-institutional research team developing a new kind of AI designed to tackle turbulence, a longstanding challenge in physics and engineering that affects technologies people rely on every day. More accurate predictions could help engineers build more efficient aircraft, strengthen the nation’s energy infrastructure and speed the development of technologies critical to U.S. competitiveness.

“At USC, we’re building an ecosystem where human-centered AI, scientific discovery and cross-sector collaboration reinforce one another,” said Gaurav Sukhatme, interim dean of the USC Viterbi School of Engineering and inaugural director of the newly named USC Mark and Mary Stevens School of Computing and Artificial Intelligence.

“The Genesis Mission reflects that vision by bringing together universities, national laboratories and industry to advance discovery while preparing the next generation of scientists and engineers,” he said.

Groundbreaking AI

Turbulence shapes both the natural world and the technologies that power everyday life, from airflow over airplane wings and wind turbines to fuel flowing through engines and pipelines. The same physics also drives storm systems, ocean currents and smoke rising from a fire.

Yet accurately predicting turbulence remains one of engineering’s toughest challenges. Although it follows the laws of physics, tiny differences in starting conditions can quickly grow into dramatically different outcomes, making it difficult for engineers to predict how turbulence will behave in real-world systems.

As part of the Department of Energy Genesis Mission, USC is leading a collaboration with the University of Michigan and Argonne National Laboratory to develop AI that can accelerate one of the most computationally demanding tasks in science: predicting turbulent flows.

“Consider the airflow around a commercial aircraft,” said Iván Bermejo-Moreno, associate professor of aerospace and mechanical engineering at USC Viterbi and the project’s principal investigator. “Predicting how turbulence evolves means following millions of tiny motions in the air as they interact — more than even today’s fastest supercomputers can calculate.”

Instead, scientists use mathematical models to estimate the smallest turbulent motions. Bermejo-Moreno's team is developing AI trained on the laws of physics and advanced computer simulations to recognize recurring patterns in turbulence and predict how they will evolve.

“Researchers have applied AI to turbulence before, but our approach is different,” Bermejo-Moreno said. “We’re teaching AI to recognize the flow structures themselves — the same kinds of swirling patterns you see in waterfalls or certain clouds — and use them to improve our predictions.”

The approach could make scientific simulations faster and more accurate, allowing researchers to solve problems that would otherwise take years to compute or remain beyond today’s computing capabilities.

Although the team’s first application is turbulence, the underlying AI could eventually improve scientific simulations across disciplines, enabling researchers to tackle problems that are too complex, too costly or too computationally intensive to solve today.

Trojans shaping the nation’s AI workforce

National initiatives like the Genesis Mission depend on more than breakthroughs in AI; they depend on the scientists and engineers who will develop and apply them. USC is helping strengthen that talent pipeline through research, training and collaboration.

USC graduate students regularly participate in the Argonne Training Program on Extreme-Scale Computing — one of the nation’s premier high-performance computing programs — and many go on to internships and careers at the Department of Energy’s national laboratories. This year, two students from Bermejo-Moreno’s lab will participate.

USC is also helping prepare the next generation of AI and advanced computing researchers through intensive summer schools led by USC Viterbi’s Department of Aerospace and Mechanical Engineering and the USC Center for Advanced Research Computing (CARC). The programs give students hands-on experience with the AI and computing tools that will power tomorrow’s scientific discoveries.

Science across sectors

USC’s role in the Genesis Mission builds on years of investment in artificial intelligence, interdisciplinary research and cross-sector collaboration.

Under President Beong-Soo Kim, USC has expanded those efforts with the launch of the USC Stevens School and broader investments in human-centered AI connecting engineering, computing, medicine, the sciences and the arts — creating new opportunities for researchers to work across disciplines and alongside partners in private industry and government.

For Bermejo-Moreno, the Genesis Mission is a natural extension of USC’s longstanding collaborations with the national laboratories and industry, bringing together the expertise needed to translate cutting-edge research into real-world impact. His Computational Aerospace Lab at USC Viterbi has relied for years on the labs’ world-class supercomputing resources to tackle scientific problems that would otherwise be impossible to solve.

Now, through the Genesis Mission, USC joins a network of 17 national laboratories, universities and industry leaders working to harness AI for scientific discovery — an effort that advances the USC Stevens School’s mission while helping train the researchers who will carry that work forward.

“Every actor, every player, brings expertise,” Bermejo-Moreno said. “National laboratories contribute world-class expertise in high-performance computing and increasingly in AI. Industry is advancing AI technologies at an extraordinary pace, while universities contribute fundamental research, innovation and the next generation of scientists and engineers.”

“The sum of those parts is much greater than what any one of us could do alone,” Bermejo-Moreno said.

About the Genesis Mission

The Genesis Mission is a historic national initiative led by the U.S. Department of Energy, which is building the world’s most powerful integrated science discovery platform. By uniting government, industry, academia and philanthropy, it is accelerating breakthroughs in energy, scientific discovery, and national security through a new platform that combines AI, supercomputing, quantum systems, and advanced scientific instruments.

The goal of the Phase I RFA awards is to identify promising pathways toward transformative scientific capabilities and establish a foundation for future investment and scale. Project teams will design and demonstrate research workflows that integrate AI with scientific investigation, while rigorously evaluating whether those approaches can accelerate discovery, improve predictive capabilities, enhance experimentation, or generate new scientific insights.

 

AI and quantum chemistry identify efficient blue OLED materials

Researchers designed an end-to-end workflow from molecular design to blue OLED device fabrication and evaluation

Nagoya University

Facebook

XLinkedInWeChatEmail

 

image: 

Illustration of end-to-end workflow from molecular design to blue OLED device evaluation

view more 

Credit: Issey Takahashi

Organic light-emitting diodes (OLEDs) have become a standard in modern devices with incredible contrast and sleek designs. While initially an expensive luxury, OLEDs are gradually becoming more financially accessible as the technology improves. Now, researchers at the Institute of Transformative Bio-Molecules (WPI-ITbM) at Nagoya University and the Institute for Advanced Study at Kyushu University have combined quantum chemistry with machine learning to identify new materials for blue OLEDs to incorporate in next-generation ultra-high-definition displays. Their research was published in Angewandte Chemie International Edition on July 21, 2026.

Blue pixels waste energy

While OLEDs are primarily composed of organic molecules, many high-efficiency red and green pixels incorporate heavy metals, such as iridium, to improve their efficiency. Such phosphorescent emitters can achieve nearly 100% internal quantum efficiency; however, conventional fluorescent blue pixels are capped at 25%. Blue pixels often use completely organic emitters because blue light requires significantly higher excited-state energy. Iridium-based emitters could also raise the efficiency of blue pixels; however, the high excited-state energy can accelerate molecular degradation, shortening the pixel lifetime.

Next generation blue pixels

The limited efficiency of conventional fluorescent blue OLEDs is due to their different excited states (i.e. higher energy state) formed when an electric field is applied across them. Only 25% of the resulting excited states can directly emit light. Therefore, next-generation OLED displays may incorporate thermally activated delayed fluorescence (TADF) molecules into blue pixels. TADF molecules are advanced organic molecules that use ambient thermal energy to convert non-emissive excited states into light-emitting states. That said, these TADF molecules need to be readily synthesized and have high color purity and high emission efficiency.

Artificial intelligence and quantum chemistry

Many leading TADF emitters use boron-containing molecular frameworks, which can be synthetically demanding to construct. To overcome this, the researchers focused on “boron-free” 13-ring frameworks systematically enumerated from combinations of five- and six-membered rings under defined chemical constraints. By strictly restricting the molecules to contain only carbon, hydrogen, and nitrogen atoms, they generated a virtual library with more than 19,000 molecules that satisfied all their chemical constraints and analyzed them with a combination of quantum chemistry and machine learning.

Quantum chemical calculations can accurately predict molecular properties; however, applying them to an enormous number of molecules incurs substantial computational costs. In contrast, machine learning enables rapid evaluation of large candidate sets but require training data. To generate these data, they randomly selected 1,000 molecules from the over 19,000-candidate library and calculated energy parameters relevant to TADF using quantum chemistry, generating the labeled dataset for machine learning.

They applied the trained model to over 17,000 molecules whose 3D structures could be prepared, enabling large-scale screening with substantially reduced computational cost. Among the over 17,000 analyzed molecules, they first selected 50 promising candidates for higher-level quantum chemical calculations. After further considering their calculated properties and synthetic feasibility, they selected two molecules to synthesize and proceeded with experimental evaluation.

Both molecules exhibited vivid blue emission with narrow bandwidths (i.e., high color purity). Additionally, their photoluminescence quantum yields in thin films reached nearly 100% (93–99%), indicating highly efficient light emission. The researchers fabricated OLED devices using the two molecules, and the resulting devices produced highly pure blue emissions. A device based on Cz-PAH-1 approached the blue-primary color defined by the next-generation Rec. 2020 ultra-high-definition display standard, while a device using Cz-PAH-2 achieved an exceptionally high maximum external quantum efficiency of 35.2%.

Broader significance

A key innovation of this work lies in integrating large-scale virtual molecule generation, quantum chemical calculations, machine learning-based screening, molecular synthesis, photophysical characterization, and device evaluation into a single continuous discovery pipeline. Using two newly developed molecules, the researchers demonstrated that machine learning-based candidate selection can be successfully translated into experimentally validated materials development.

The methodology demonstrated here can readily be extended to the efficient discovery of other classes of organic functional materials with enormous molecular design spaces, facilitating the identification of promising candidates for future OLED applications.

---

Journal

Angewandte Chemie International Edition

DOI

10.1002/ange.2731055 

Method of Research

Experimental study

Subject of Research

Not applicable

Article Title

Machine-Learning-Guided Discovery of Boron-Free Narrowband Blue Thermally Activated Delayed Fluorescence Emitters

Article Publication Date

21-Jul-2026

 

Combined quantum chemistry and machine learning approach.

Credit

WPI-ITbM

 

Fabricated AI animal videos are distorting how we see wildlife, researchers warn

Copyright (AP Photo/Michael Probst) Copyright 2026 The Associated Press. All rights reserved

By Una Hajdari

Published on 21/07/2026 - EURONEWS

AI-generated animal footage is racking up millions of views online, but researchers warn it is also distorting public understanding of wildlife behaviour and could be undermining support for conservation.

Fabricated animal footage on social media is distorting conservation efforts and could be undermining public support for endangered species, a study has found.

Generative AI tools are now capable of producing convincing wildlife videos and photographs that never happened, according to a paper published in Conservation Biology by researchers at the University of Cordoba.

The authors, José Guerrero-Casado, Tamara Murillo-Jiménez, Antonio Carpio, Francisco Tortosa and Rocío Serrano-Rodríguez, argue this content is increasingly shaping how the public understands animal behaviour, often incorrectly.

One example the researchers highlight is a viral AI video showing various bird species sheltering chicks from the rain, widely shared with captions describing it as "true mother love."

The framing misses an established fact: in 90% of bird species, males also take part in raising young, while many reptiles, amphibians and fish provide no parental care at all.

The paper also points to invented interactions between species, such as fabricated footage of predators and prey, or parasites and their hosts, behaving with implausible affection toward one another.

Separately, the authors highlight videos showing fictional bonds between humans and wild animals, including one clip of a polar bear being rescued by fishers and reacting with exaggerated gratitude.

The authors warn such scenes risk giving people a false sense of security around wild animals and could encourage demand for exotic pets, fuelling illegal wildlife trade.

Conserving 'cute' animals

The researchers also expect AI-generated content to skew toward mammals, since these species already tend to perform best on social media.

They warn this could reinforce existing funding imbalances, with conservation projects for less popular animal groups losing out.

They also warn that fake, location-tagged wildlife footage could drive tourists to sites where the animal shown was never actually present, adding pressure on ecosystems.

The study stops short of proposing a fix, conceding that global regulation of AI content is unlikely in the near term, and calls instead for wider media literacy education so audiences learn to question what they see online.

Relate

Citizen science records under threat

The findings echo a separate warning issued this week by researchers writing in Nature Ecology and Evolution, who argue that AI-manipulated photographs, audio and video submitted to citizen-science platforms could contaminate the data researchers rely on to track where species occur and how they behave, potentially leading to flawed ecological conclusions.

The researchers point to over-enhancement as the more common problem, rather than outright fabrication.

Editing tools can strip out or alter the physical features used to identify a species, sometimes causing it to be misidentified altogether.

They cite a real case in which a photograph believed to show a red-winged blackbird, a North American species never before recorded in Brazil, was submitted to iNaturalist as a potential first sighting.

The bird was in fact an epaulet oriole, a species common to the region. The image had been "rebuilt" using Google's AI image editor, which added markings resembling the North American bird, likely reflecting that species' heavier representation in the tool's training data.

The researchers stress the contributor had no intention of misleading anyone. Their goal, they say, was purely cosmetic: a better-looking picture.

The team says they managed to replicate the same mistake independently using AI editing tools. They conclude that contributors urgently need to be made aware of how much damage this kind of editing can do to scientific data.

 

'Unprecedented': OpenAI models autonomously hacked a rival firm, fuelling fears of rogue agents

Copyright Copyright 2023 The Associated Press. All rights reserved.

By Una Hajdari

Published on 22/07/2026 - EURONEWS

OpenAI has admitted one of its models exploited a hidden flaw to escape a controlled test and break into Hugging Face's servers, in what its CEO called an autonomous, first-of-its-kind breach.

ChatGPT maker OpenAI said late Tuesday that its artificial intelligence system hacked into another AI company on its own in what the company called an "unprecedented cyber incident."

"We had a significant security incident during evaluation of our models," OpenAI CEO Sam Altman said in a statement posted on social media.

AI startup Hugging Face said last week that it had detected an intrusion into its data processing systems that it suspected was caused by an AI agent autonomously acting on its own.

"We suspected last week's cyberattack might have come from a frontier lab, given the sophistication of the agent," Hugging Face co-founder and CEO Clément Delangue said in a statement. "Turns out it did!"

This means the attack was so advanced and well-executed that Hugging Face suspected it came from one of the top AI companies' systems, not a random hacker.

What happened?

OpenAI was running an internal test to measure how good its AI models are at hacking — a benchmark called ExploitGym.

To see the models' maximum hacking ability, they deliberately switched off the safety filters that normally stop the models from doing dangerous cyber activity.

The test was meant to run in a sealed-off "sandbox" or an isolated environment with no real internet access, apart from a tool that lets the models download software they might need to complete the task.

However, the models became fixated on solving the test. Rather than solving it the intended way, they went looking for a shortcut and found a way to claw themselves into the open internet — which they were never supposed to reach.

Getting there involved a chain of steps, quietly gaining more and more access inside OpenAI's own systems until they hit a point with an internet connection.

Once online, the models reasoned that Hugging Face — a big platform hosting AI models and datasets — probably held the answers to the very test they were supposed to solve.

So they broke into Hugging Face's servers to steal those answers, essentially to cheat, using stolen login credentials and more flaws to get in.

Chinese models to the rescue?

As an open marketplace that anyone can publish to, Hugging Face hosts a huge volume of Chinese-developed models.

When Hugging Face's team tried to analyse the attack, they fed the raw attack data — the code and commands used to exploit their system — into commercial AI models to help reconstruct what happened.

But those AI models have built-in safety filters designed to block anything that looks like hacking — and to those filters, the evidence of an attack looks exactly the same as an attack itself.

So the models refused to help, unable to tell the difference between a hacker doing harm and a company defending itself.

Blocked, Hugging Face switched to an open-weight Chinese model — Z.ai's GLM 5.2 — which it could run locally, inside its own systems, and which processed the material without refusing.

Chinese labs such as DeepSeek and Alibaba's Qwen have become some of the most downloaded model families on the platform, and by some measures, Chinese developers now account for a larger share of Hugging Face's downloads than their US counterparts.

Major security concern

The disclosure comes amid heightened concerns about the cybersecurity capabilities of powerful models that led US President Donald Trump in June to sign an executive order creating a framework for the federal government to vet the national security risks of the most advanced AI systems for up to a month before their public release.

"AI is accelerating the discovery and exploitation of vulnerabilities," OpenAI said in its statement Tuesday. "The primary lesson from this incident is that model security and safety must keep pace with rapidly advancing capabilities."

Delangue said he spent the past 24 hours working with OpenAI, "and we strongly believe there was no malicious intent on their part. It's quite mind-blowing that all of this happened autonomously!"

Delangue added that it "might be the first incident of its kind."

OpenAI said the intrusion was caused by a combination of its AI models, including its newly released GPT-5.6 Sol and an "even more capable" model that is still being tested internally.

OpenAI said its AI used stolen credentials and discovered a previously unknown vulnerability to access Hugging Face servers.

It went to "extreme lengths to achieve a rather narrow testing goal" and "found ways to gain access to secret information that it could use to cheat the evaluation," the company said.

Trump DOJ's 'stunning' admission about ICE deaths appalls CNN anchor

Robert Davis
July 22, 2026 

RAW STORY

Masked law enforcement officers, including HSI and ICE agents, walk into an immigration court in Phoenix, Arizona, U.S., May 21, 2025. REUTERS/Caitlin O'Hara/File Photo

CNN anchor Jake Tapper stunned one of his colleagues on Wednesday while discussing the findings of a new investigation into the killings of Renee Good and Alex Pretti in Minneapolis

Tapper investigated the Trump administration's Operation Metro Surge, which sent about 3,000 immigration agents into Minneapolis in January 2026, and the corresponding departure of Trump administration prosecutors assigned to the case. He said the investigation found that the prosecutors had already raised alarms about the operation and former Customs and Border Protection commander Greg Bovino's actions when Good and Pretti were killed.

Prosecutors also expressed consternation that the statements of ICE agents were found to be false repeatedly, according to the investigation.

"They get away with a lot," Tapper said about the Trump prosecutor's concerns over CBP and Bovino going to Minneapolis. "Them coming here is a bad idea. Bovino coming here is a bad idea. And the quote was, 'Someone's going to get killed.' That was in early January."

"January 7th, Renee Good is killed, and then Bovino calls mad about the lack of prosecutions against protesters, and says a whole bunch of stuff about shutting down the city, and says some anti-semitic things," Tapper continued. " ... And then after that call, one of the prosecutors says, 'They're going to kill someone else.' And then a few days later came the Alex Pretti killing."

CNN's Brianna Keilar seemed appalled by the report during "CNN News Central."

"That is stunning," she said. "When a lot of people say that who are critical of what happened in Minneapolis, they would be dismissed as hyperbolic."

Tapper added that prosecutors thought Bovino and CBP were intentionally trying to incite a riot to go after protesters.

"Just keep in mind, we're not talking about a bunch of hippie communists, right? These are prosecutors. These are people who joined the government, want to lock away bad guys, and keep communities safe," Tapper said. "And the accusations, the impression they got was ... these tactics were overaggressive and actually counterproductive. They were not going after the worst of the worst. They were going after protesters, and they were trying to start a riot."

Trump's food-stamp 'fraud' scheme unravels as red states cry foul

Travis Gettys
July 23, 2026 

RAW STORY


A man holds a sign reading "SNAP Feeds Families," as food aid benefits will be suspended starting November 1 amid the ongoing U.S. government shutdown, during "A Rally for SNAP" on the steps of the Massachusetts Statehouse in Boston, Massachusetts, U.S., October 28, 2025. REUTERS/Brian Snyder/File Photo

An Agriculture Department push to root out "fraud" in the nation's largest food aid program is facing unexpected resistance from two Republican-led states that say the administration's own numbers don't add up, according to a new report.

Ohio and Georgia, the first states examined under the initiative led by Agriculture Secretary Brooke Rollins, told USDA's inspector general that its fraud-detection methods were flawed and its findings inaccurate, according to documents obtained by Politico.

“The basis for Ohio’s disagreement is twofold — proper audit procedures were not followed, and 2) the specific data requested by OIG was insufficient and as a result does not accurately reflect Ohio’[s] administration of SNAP," the state said in a response to the inspector general's initial findings.

The pushback complicates months of public messaging from Rollins, who has repeatedly claimed "significant fraud" in the Supplemental Nutrition Assistance Program, including unverified assertions that beneficiaries are driving luxury cars.

In Ohio's case, the inspector general initially flagged $59.7 million in potential fraud. After the state disputed the findings, that figure was cut to $13.3 million — and Ohio is contesting even that number. State officials say at least 93 percent of the original claims weren't fraud at all, but routine administrative matters. One flagged case involved a person over 110 years old, wrongly suspected of being a fraudulent recipient; Ohio provided Social Security records confirming the person was alive and eligible.

Georgia's dispute was even sharper. The inspector general initially estimated nearly $300 million in potential fraud in the state — a figure later reduced by about 95 percent after Georgia challenged the underlying data, according to a former USDA employee familiar with the review.

Former department officials say the pattern suggests legitimate payments and clerical errors are being swept into fraud totals. Alan Shannon, a former USDA payment accuracy specialist, said he's never seen the agency characterize routine errors this way, calling the approach "highly unusual," and another official with the agency criticized the administration's efforts.

“It concerns me when they’re potentially inflating the findings and undermining public trust in the programs,” said Stacy Dean, the former deputy undersecretary of USDA’s Food, Nutrition and Consumer Services division.

“They’re there to enhance public trust in the programs by being that watchdog, but when they don’t take the time to check their work and to be sure of its accuracy, they can erode public confidence in these critical nutrition programs.”

USDA Inspector General John Walk defended the process, saying preliminary findings are shared with states as part of a normal back-and-forth before final conclusions are reached, and pushed back on the characterization of the exchanges as "disagreements and disputes."

The dispute lands amid a broader administration campaign — backed by Vice President JD Vance and formalized through a March executive order — to publicize fraud in federal benefit programs. It also coincides with SNAP enrollment dropping by roughly 4.3 million people over the past year, following stricter eligibility rules passed by Republicans last year.

With inspections still pending for states including Texas, Florida, North Carolina and Pennsylvania, Ohio and Georgia's experience suggests the administration's final tallies could fall well short of the sweeping fraud claims driving its messaging campaign.