Chinese AI could subtly shape how Americans understand history, January AI cofounder and CEO Noosheen Hashemi, one of Oracle’s first vice presidents — who helped increase the company’s annual revenue from $26 million to $3 billion — warned in an exclusive interview with Breitbart News.
Hashemi, a former Silicon Valley Democrat who became a supporter of President Donald Trump and his America First agenda, told Breitbart News that Chinese AI models do not enter American systems as neutral technology. “To the extent that we implement Chinese models in our work, those vessels don’t travel empty,” she said, noting there is now “empirical evidence” that AI systems can affect “what the user remembers” and “what the user receives as information.”
Hashemi pointed to four mechanisms she said are reflected in empirical research — “data, rules, filters, and language” — through which AI outputs can be shaped. She described data as “what the model learns from the training material and information environment” and said a 2026 Nature paper found Chinese state-scripted coordinated media in large-language-model training data. According to Hashemi, when researchers deliberately subjected an open-weight model to additional pretraining on that material, “its answers became more favorable to Chinese political institutions and leaders,” which she said demonstrates that “the data that goes in essentially impacts the answers that come out.”
The second mechanism, Hashemi said, involves the “rules” developers operate under, including “legal requirements, post-training instructions, refusal policies, and behavioral goals.” She pointed to another 2026 paper from Stanford and Princeton published in PNAS Nexus that, as she described it, posed 146 questions about Chinese politics to models originating in China and models originating elsewhere. China-originating systems were “substantially more likely to refuse basically sensitive questions, or give shorter answers, or give inaccurate answers,” she said. Hashemi cited examples involving dissident Wei Jingsheng, the Great Firewall, and Nobel Peace Prize laureate Liu Xiaobo, saying one model inaccurately described Liu as “a Japanese scientist associated with nuclear weapons.” She said such responses can “sound like an authoritative correct answer, except it’s wrong.”
Hashemi said “filters” are controls applied before an answer reaches a user and can “block, shorten” or “rephrase material.” She described a 2026 DeepSeek audit in which researchers tested nearly 650 politically sensitive topics and found only 12 produced what she called “hard censorship,” such as an explicit error or empty response. She said researchers also found what she described as “semantic suppression,” in which a model still produced an answer but information was filtered. According to Hashemi, information could be present in the model’s reasoning before the final answer was revised or reframed around topics such as “transparency or government accountability or civic mobilization.”
Language can also change the response, Hashemi said, pointing again to research she said found “more favorable answers about Chinese political institutions and leaders if the questions were asked in Chinese rather than in English.” She also cited two studies from 2025 in which people were exposed to liberal-biased, conservative-biased, or neutral models. According to Hashemi, participants’ “opinions and decisions were affected by the model’s bias, even when their own views were counter to the model,” while another study involving almost 5,000 people found attitudes on policy areas changed based on how the model formulated the question. “The overall risk to changing history is sort of forward-looking,” she said, “but the evidence that AI can change and shape answers is very real.”
For ordinary users, Hashemi said identifying systematic foreign influence is difficult because AI often delivers a single synthesized response rather than a visible collection of sources. “A library used to give you 10 books, Google gave you 10 links, but AI more often than not is giving you one answer, and you can’t really see how it’s coming up with that answer,” she said. China, she said, “doesn’t have to really change the historical record” if AI becomes “that front door to the historical record,” because “shifting the answers enough” could influence what information users receive.
Hashemi urged users — including students — to practice what she called “AI literacy.” Students receiving an AI-generated answer “should look up the sources,” “open the links,” compare information, and avoid assuming that “the one answer that they’re getting is the answer,” she said. Hashemi pointed to differences among AI companies, noting that Anthropic has published Claude’s constitution while xAI has promoted “truth seeking.” “Consumers have choice,” she said, adding that the United States also has “private capital,” “entrepreneurs,” “transparency,” and a free press that can evaluate, test, and question models. “In a free society, no answer has to be final,” she added, warning against taking answers from AI “as gospel.”
Asked whether the larger threat comes from Americans directly using Chinese AI models or from Chinese technology becoming embedded in products without users necessarily realizing it, Hashemi said the latter poses the greater danger. “The standard is not one fact or lie. It’s millions of small distortions that nobody notices,” she said. “Political power presents the model, and the model shapes the user.” At scale, she said, that process “can shape institutions,” potentially affecting “classrooms, journalism, government, medicine, politics, and everything.” That is why, Hashemi said, she is “so obsessed with getting AI, American AI, into allied systems,” including factories, government, business, and medicine, because Chinese models “will not travel empty.”
Hashemi tied that concern directly to industrial competition with China and Trump’s reindustrialization agenda. “President Trump has made reindustrialization a key tenet of his administration,” she said, saying that “industrial power is really defense” and “really national power” because “if you can’t make things, you can’t defend yourself.” Calling China the “first factory to the world” and Germany the “second factory in the world,” Hashemi said a coordinated U.S.-German effort could expedite American reindustrialization while ensuring American AI enters German factories. China already competes at a scale and price others struggle to match, she said, and adding AI to Chinese manufacturing could further compound that advantage through factories capable of thinking, self-correcting, redesigning products, and watching for errors. “We need to have a competitive power in physical industrial AI,” she said.
Asked what the Trump administration should prioritize to preserve America’s AI advantage over China, Hashemi said the United States needs energy, strong chips, compute, and models, but said the competition goes beyond technical performance. “It’s not enough for our models to win benchmarks,” she said. “It’s important that we spread American AI globally,” because “inventions become power when technology is actually adopted.”
Hashemi described the central contest as one of “diffusion” — which country’s technology becomes widely used throughout the world. “The fight is whose models are diffused all over the world, which models are getting used, to what extent, in what countries and what industries,” she said. “The biggest prize is diffusion: whose AI the world uses.” She also called for improvements in “provenance and attribution and repeatable evaluation,” saying research evaluating AI behavior can help explain how systems arrive at their answers.
Hashemi, who said she has spent the past eight years building AI after years in Silicon Valley, said the experience changed how she views the stakes surrounding the technology. Political concerns in Silicon Valley often centered on issues “like climate,” she said, but “once you live AI essentially in an AI world day in and day out, it becomes really clear what the stakes are.” Hashemi said she now does not believe climate is “the biggest threat at all,” adding, “I feel that the biggest threat is living under authoritarian AI.”
Looking toward the midterm elections, Hashemi said she hopes voters “keep their eyes on the real contest” while acknowledging that concerns surrounding AI infrastructure are legitimate. “The local issues are real: water, noise, electricity. All of that is true,” she said. But because AI depends on compute, data centers, chips, and electricity, she said policymakers cannot ignore strategic competition with China. Hashemi pointed to “PRC-linked influence activity” that she said has been documented and discussed by companies including OpenAI and xAI. “We know these things are real,” she said, while adding that “legitimate local concerns” must also be addressed. She characterized the U.S.-China AI competition as “absolutely a war” because the contest encompasses “economic power, military capabilities, science, medicine, manufacturing — all of it.”
Hashemi expanded on the industrial competition in an August 19 Townhall column, saying that Germany’s manufacturing base presents an opportunity for the United States to spread American AI into allied economies. She warned that if “American AI does not get embedded into German manufacturing first, Chinese AI will,” pointing to Beijing’s AI Plus initiative and its goal of 70 percent AI penetration across key sectors by 2027 and 90 percent by 2030. Hashemi called for American AI to run in German factories while American energy powers it and German manufacturers invest in U.S. reindustrialization, writing that “trade was the warm-up; AI is the real fight” and that “America cannot afford to finish second.”
Hashemi’s warnings come as U.S. officials have accused Chinese AI companies of attempting to obtain American AI capabilities through knowledge distillation. The FBI, NSA (National Security Agency), and CISA (Cybersecurity and Infrastructure Security Agency) said Chinese firms had “extracted billions of tokens across millions of exchanges/requests” from American frontier models since at least late 2024, while Anthropic separately accused DeepSeek, Moonshot, and MiniMax of conducting “industrial-scale campaigns” to “illicitly extract” capabilities from Claude. China rejected the allegations as “groundless” and described distillation as a “common practice.”
Alliance for a Better Future CEO Janet Kelly also warned about the geopolitical stakes of the United States-China AI competition during Breitbart News’s September 8 “State of the Economy” policy event. Kelly said China would “cheat and steal and lie” to beat the United States and warned that if America hesitates, it risks handing “the most powerful technology in the world” to a foreign adversary. She also said America is the only country in the world that can build AI “the world trusts,” while warning that “the window for this is closing.”
Treasury Secretary Scott Bessent also addressed the United States-China AI race at the Breitbart News event, saying, “If they were to pull ahead of us on AI, then nothing else matters.” Bessent said the United States benefits from an ecosystem built around technology, chips, and financing, but also criticized AI companies for doing a “horrendous job reaching out to communities” affected by data centers. He also warned against pausing American AI development because “the Chinese won’t pause.”
Read the full article here
