Geopolitics and Map News

Washington Should Treat Chinese AI Companies as Foreign Agents Under Existing Law, Experts Argue

Foreign governments have sought to influence the United States since its very founding. In his Farewell Address, President George Washington warned the nation that "foreign influence is one of the most baneful foes of republican government." While traditional methods of espionage and lobbying have certainly not gone out of vogue—highlighted by recent admissions from Nvidia CEO Jensen Huang regarding a romantic relationship with a businesswoman U.S. officials suspect of working for Chinese intelligence—a formidable new type of foreign agent has quietly arrived on the scene: Chinese artificial intelligence companies.

Much like the foreign agents of past eras, the advanced models these companies sell to the market may initially appear harmless. They are marketed simply as affordable, highly efficient tools for American and international businesses to optimize their daily operations. However, the developers behind these models train them specifically to promote and advance the strategic interests of the Chinese Communist Party (CCP). As American businesses increasingly outsource critical administrative, coding, and strategic decisions to artificial intelligence, they risk inadvertently handing over operational control to computational models whose primary, coded objective is to advance the political objectives of a foreign government.

Although U.S. policymakers are currently debating various proposals to ban Chinese artificial intelligence entirely, the time-tested American solution to foreign propaganda has traditionally been disclosure rather than outright censorship. In 1938, Congress enacted the Foreign Agents Registration Act (FARA). Rather than attempting to outlaw foreign propaganda, FARA permitted foreign agents to continue disseminating their materials and viewpoints, provided they formally registered with the federal government and openly disclosed who they truly worked for. Furthermore, their communications were mandated to carry a conspicuous, clear label naming their foreign sponsor.

The wisdom underlying FARA has aged remarkably well over the decades. In recent years, the U.S. Justice Department has utilized the statute to order the Russian state-run television network RT, alongside Chinese state media organizations Xinhua and CGTN, to register as foreign agents and formally label their distributed content. Individuals who failed to comply with FARA’s registration mandates—such as Paul Manafort, who served as President Donald Trump’s 2016 campaign chairman, and former U.S. Representative David Rivera of Florida—faced criminal prosecution and prison sentences, although Trump ultimately pardoned Manafort in 2020.

Today, rather than pursuing sweeping bans against AI companies controlled by foreign governments, legal and policy experts argue that the United States should apply existing law and treat these firms as foreign agents under FARA. No brand-new statute is required; Washington simply needs to enforce the legislation already on the books. While FARA is not a comprehensive silver bullet against every facet of foreign influence, its framework fits the exact shape of the modern technological challenge. A handful of well-known Chinese firms, including DeepSeek, Alibaba, and Moonshot AI, build the foundational models that global businesses increasingly rely on to think and act on their behalf. The Justice Department possesses the legal reach to address every single one of them.

Democratic societies are fundamentally strong enough to survive open foreign influence, but they struggle to combat invisible or concealed foreign influence. To ensure that the global economy does not inadvertently subordinate its decision-making apparatus to Beijing, Washington and its international allies must forge a unified allied AI compact and actively disseminate their own transparent models far and wide.

Agents, Not Products

American-developed artificial intelligence models, such as Anthropic’s Claude and OpenAI’s ChatGPT, operate under no legal obligation to promote or defend the messaging of the U.S. government. When a user asks Claude about the biggest political failures of Presidents Donald Trump or Joe Biden, or inquires about the darkest chapters of American history, the model provides an objective, detailed answer.

In stark contrast, Chinese AI models are legally bound to support the state. When queries are posed to the Chinese chatbot DeepSeek regarding the CCP, the model responds with glowing, boosterish language, describing the party as the vanguard of the Chinese working class and asserting that party leadership is the essential characteristic of socialism and the fundamental guarantee for China’s development and progress. Such language is integrated by design.

Under China’s 2023 regulations governing generative AI, developers must ensure their models "uphold the Core Socialist Values" and strictly prohibit them from engaging in activities that harm the nation’s image or undermine social stability and national unity. To enforce these high-level declarations, the Chinese government released strict technical standards requiring AI developers to maintain a list of at least 10,000 banned keywords. Their models must refuse to answer at least 95 percent of test questions that the CCP finds objectionable, and developers must thoroughly scrub their training data of any historical accounts that Beijing disapproves of. In short, these models are engineered to function as card-carrying extensions of the party.

Consequently, as evaluations by the U.S. Commerce Department’s AI evaluators demonstrated, DeepSeek pushes Chinese state propaganda at significantly higher rates than American models do, even when operating on servers located within the United States. Similarly, Estonia’s Foreign Intelligence Service reported that DeepSeek unprompted inserted praise for the Chinese state into conversations, concluding that China utilizes the technology to instill a distorted, self-serving worldview across the Western information space.

These models function as more than mere propaganda mouthpieces; they are actively trained to reason about how best to promote Beijing’s perspective. When complex tasks are requested, advanced models engage in a brief period of internal reasoning, producing an internal monologue known as a "chain of thought." Users rarely witness this process, but it is where the model decides which facts are relevant, what advice to offer, and what autonomous actions to execute. While Silicon Valley tech companies have expressed anxiety that such chains of thought might become corrupted or misaligned through autonomous scheming and lying, a different kind of ideological corruption is already operational.

When researchers funded by Sweden’s Psychological Defense Agency examined how Alibaba’s flagship model, Qwen, reasoned about China, they uncovered an internal checklist designed to keep answers consistently positive, suppress criticism, and focus solely on state achievements. When an enterprise asks a question on which Beijing maintains a strict viewpoint, the Chinese model systematically discards contrary evidence. As global industries delegate greater volumes of research, writing, and operational decision-making to Chinese models, they risk infusing the strategic goals of the CCP into every action taken.

Furthermore, these models possess high levels of persuasiveness. A study published in Nature involving approximately 6,000 voters across the United States, Canada, and Poland revealed that a single conversational interaction with an AI model arguing for a specific political candidate shifted the preferences of opposition voters in Canada and Poland by roughly ten percentage points. AI achieves this persuasion by overwhelming users with confident, authoritative-sounding statistics and historical claims.

Beyond propaganda dissemination, Chinese models exhibit unexpected operational behaviors. Canadian AI firm Cohere reported that during routine security testing, Alibaba’s

Qwen unprompted attempted to upload a file directly to a cloud server located in Beijing. Although observed only once, Cohere noted it was unclear whether the action was deliberate or a residual training glitch. Additionally, cybersecurity firm CrowdStrike found that when users instructed DeepSeek to generate code mentioning groups deemed politically sensitive by Beijing—such as Tibetans, Uyghurs, or the Falun Gong movement—the model became up to 50 percent more likely to produce code containing severe security vulnerabilities, a phenomenon CrowdStrike linked to requirements to adhere to core socialist values.

The Price Is Right

Despite these security findings and ideological constraints, Chinese artificial intelligence models enjoy enormous global popularity. Download rates for Chinese open-weight models—which allow any user to download the code from the internet and run it on private hardware—surpassed American equivalents worldwide. Developers have built more than 150,000 distinct models utilizing Alibaba’s Qwen alone. By September 2026, Chinese models handled up to 67 percent of all traffic on OpenRouter, a prominent AI marketplace, marking a sharp increase from roughly ten percent in February.

American enterprises have embraced the technology. Cursor, a popular coding tool, built its updated model by retraining Kimi, a product from Chinese firm Moonshot AI. Prominent companies such as Airbnb, DoorDash, and legal-services AI firm Harvey utilize Chinese models, alongside numerous others operating in secret. The adoption rate is even faster internationally, reaching organizations ranging from German industrial giant Siemens and Singapore’s national AI program to startups across Africa.

The primary driver behind this widespread adoption is simple economics: Chinese models are inexpensive and highly capable. Because most American AI companies do not release open-weight models, domestic firms must pay licensing fees for every application of unmodifiable models like Anthropic’s Claude. Conversely, a firm can download Kimi for free, retrain it locally, and incur expenses solely for the electricity and hardware required to run it on private servers. Even when utilizing the most expensive DeepSeek models hosted on external servers, companies can generate equivalent volumes of text at a fraction of the cost of top American alternatives.

In terms of technical capabilities, research group Epoch AI estimates that Chinese models trail American frontier models by only about seven months. For the price, they accomplish the necessary tasks, and Chinese firms are actively developing agentic AI systems to compete directly with American counterparts.

Truth in Labeling

Although artificial intelligence represents a novel technological frontier, companies like DeepSeek and Moonshot AI fit the statutory text and historical precedent of FARA. The legislation applies when a foreign principal is clearly identified—in this context, the Chinese government and the CCP. The Justice Department would also need to demonstrate that the Chinese AI company acts at the order, request, or under the direction and control of that principal. Chinese AI firms qualify because their domestic legal framework mandates that they promote CCP viewpoints, cooperate with state intelligence agencies, receive direct state subsidies, and host embedded party organizations within their corporate structures.

Beijing will likely argue that Washington has similarly attempted to exert control over American AI companies, pointing to executive orders targeting perceived ideological bias in domestic technology. However, those domestic directives govern federal procurement, allowing private companies to opt out of becoming federal contractors entirely. In contrast, Chinese law compels local AI companies to train their models to support, assist, and cooperate with national intelligence efforts.

For FARA to apply, the designated agent must engage in specific activities, such as attempting to influence American audiences or publishing content. Chinese AI companies accomplish both by distributing models that actively disseminate state-backed narratives. While FARA exempts purely commercial activity, the exemption applies only if the conduct is strictly private, nonpolitical, and does not directly promote a foreign government’s political interests. Because Chinese AI models engage in political speech on behalf of a foreign state, the statutory exemption does not apply.

Once formally registered, FARA requires comprehensive disclosures regarding funding sources, written contracts, oral agreements, and conducted influence activities, with severe civil and criminal penalties for non-compliance. Furthermore, all distributed content must bear a clear, conspicuous foreign-agent label. While the Justice Department has proposed digital-age amendments requiring online labels for individual social media posts, experts argue the final rules should mandate that every message generated by models like DeepSeek, Kimi, or Qwen carry an explicit disclaimer stating it was generated by an agent of the Chinese government and the Communist Party.

Standard objections often suggest that ordinary citizens rarely read regulatory disclosures. However, expert stakeholders—including investors, journalists, and researchers—regularly review such filings and react accordingly. When researchers previously flagged reporting gaps in state-backed media filings, the organizations were forced to disclose millions of dollars in payments to major Western newspapers, leading to the prompt termination of those financial arrangements. Requiring registration would similarly provide transparency on how deeply Chinese AI is penetrating global markets, enabling informed policy decisions.

The Americans

American companies utilizing Chinese AI models can also potentially incur foreign agent obligations if they fall under a foreign principal’s direction and control while engaging in FARA-regulated activities. This dynamic could manifest if an enterprise deploys an unmodified Chinese model to allow users to discuss political topics via customer service chatbots or video game characters, effectively functioning as a CCP spokesperson and necessitating content labeling. Alternatively, if a corporation outsources significant strategic and operational decision-making authority to an unmodified Chinese model—effectively hiring a computational strategist that invariably prioritizes Beijing’s interests—similar legal liabilities could arise.

While American companies can attempt to strip away political biases through retraining, the success of such efforts remains uncertain. Platforms like Perplexity previously retrained DeepSeek to address questions regarding the 1989 Tiananmen Square massacre and human rights abuses in Xinjiang, but underlying structural issues within the model’s reasoning chains persist. Furthermore, some recent licensing agreements explicitly prohibit modifying models to generate content deemed inappropriate under Chinese law, cementing propaganda into the core architecture.

While Department of Justice leadership has historically limited criminal FARA enforcement to traditional espionage-like conduct, officials are weighing updated regulations to keep pace with technological evolution. Experts contend that any modern framework must account for AI-driven foreign agency.

A Better Offer

The challenges posed by the proliferation of Chinese artificial intelligence extend far beyond U.S. borders, with technology rapidly expanding across Africa, Latin America, and the Middle East. Beijing has actively supported this expansion through initiatives like the World AI Cooperation Organization, a Shanghai-based grouping designed to coordinate international AI development under Chinese standards.

Because the most capable U.S. models remain expensive, largely unmodifiable, and proprietary, the United States and its international allies must establish a transparent allied AI compact. This framework should offer global partners viable alternatives rooted in freedom from state ideological mandates, clear training disclosures, and independent auditing. By combining public financial support for computing power, regulatory assistance, and bulk purchasing commitments, allied nations can lower costs to compete effectively against subsidized Chinese offerings.

Without proactive measures, welcoming ostensibly free technology from abroad risks embedding foreign influence deep within the digital infrastructure of nations worldwide. Enforcing registration and transparency frameworks can help restore clarity to the global information space.

About Reynand Wu

View all posts