US, China Psy-Op Accusations Signal Deep Distrust on AI Regulation

Lately, in the US news cycle, the “psy-op” allegation in AI policymaking has cut multiple ways.

Individuals or organizations advocating to slow or halt frontier AI development and enact more regulatory oversight are accused of being psy-ops either backed by China or other domestic ideological groups. For instance, Elon Musk referred to former Anthropic employee Jacob Coxon’s viral resignation post from Anthropic as such, and White House science and technology advisor David Sacks called out the lack of similar “doomer” discourse groups within China as a primary tell that the campaign is manufactured.

The same allegation has befallen organizations and individuals pushing back on the aggressive local buildout of data centers. Sacks’ boss, US President Donald Trump, has emphatically stated that “there is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China.” Y Combinator CEO Garry Tan wrote in late August that “the datacenter discourse is a documented psyop” and connected dots from the Chinese government and state media to anti-data center messaging in America. Both of these movements are, according to prominent US figures inside of and adjacent to politics, psy-ops backed by—or at least benefiting—the Chinese government.

This piece does not adjudicate the truth of psy-op claims. Instead, it examines their political origins and impact. The US Department of Defense defines psy-ops as “planned political, economic, military and ideological activities directed toward foreign countries, organizations, and individuals in order to create emotions, attitudes, understanding, beliefs, behavior favorable to the achievement of… political and military objectives.” This term has caught on in online parlance and can be used broadly to refer to any statement perceived as inauthentic and in service of an ideological group’s interests.

US observers often assume China to be either the perpetrator or immediate beneficiary of these alleged psy-ops. This tendency is often part of a larger discursive trend of fearmongering about an AI race with China, which fuels domestic AI accelerationism.

However, this “psy-op” allegation has also been levied by Chinese commentators against their US counterparts. As the Carnegie Endowment for International Peace’s Matt Sheehan said in a recent interview, Chinese state media has implied that US calls for AI safety are part of an effort to unilaterally define the terms of Chinese—and rest-of-world—adoption of advanced technologies while letting US companies go unregulated. A commentary published on August 30 by a CCTV-affiliated account argued that efforts to rein in “dangerous capabilities” among other nations by the US would not be seen as sincere, as the US government had yet to ensure that American firms themselves face real scrutiny at the domestic level. Similarly, after Amodei's “Pacing the Frontier” essay, the state-owned outlet the Global Times published a piece titled "Targeting China's AI: US 'tech right' unfolds Cold War playbook," which denounced a “hidden agenda” of a “silent AI Cold War” whose "true purpose" was "to curb China's AI development through technological barriers and regulatory monopolies." On Sept. 14, the Chinese Foreign Ministry's spokesperson warned against "fearmongering” in response to a question about Amodei’s concerns.

These mutual psy-op allegations arise from readings and misreadings of hidden intent and sentiment both between countries and within countries.

And for these two broad-tent battleground issues—the public interest strand of contesting tech buildouts of infrastructure at the local level and regulating AI development at the frontier—claims of psy-ops indicate an underlying information ecosystem problem. Whether data centers are or aren’t worth it for local communities (probably not), as well as whether frontier models should be paced or restricted and how, all depend on adjudicating companies’ claims without a clear, agreed upon, third-party source of truth that is fully independent of company interests. The frequency of psy-op callouts in this debate is a symptom of this information gap.

Across both countries, psy-op narratives largely rely on a broader atmosphere of mistrust. China's media environment is far more shaped from the top down, whereas in the US company spokespeople often try to loudly shape both public and elite opinion at once. Each country's leadership might recognize that its own discourse is a mosaic of conflicting views, yet, as Sheehan also observes, each side tends to look at the other and see a grand adversarial conspiracy, a pattern that predates the AI issue and defines most competitive relationships between states.

Psy-op discourse is also not unique to this time period. A common delegitimation tactic is to cast discourse you don’t like as being in some way foreign or inauthentic so as to be more discreditable. In an Oct. 1982 address, Reagan alleged that the nuclear freeze movement was fueled not by "the sincere, honest people who want peace, but by some who want the weakening of America, and so are manipulating many honest and sincere people”; he repeated this charge at various other speaking engagements. While the Soviet-backed World Peace Council did fund some Western peace groups, the House Intelligence Committee ultimately stated via its Chairman that Soviet agents had "no significant influence on the nuclear freeze movement.”

To ponder the clear analogy to the “Cold War era” further, amid conditions of mistrust, meaningful US-Soviet treaties were still gradually possible. The Intermediate-Range Nuclear Forces Treaty, for instance, eliminated a class of missiles and introduced on-site inspection of each side’s facilities by the other. But notably, conditions today are different: AI is not a good analogue for nuclear weapons, as private companies are the key owners and actors here; capability is not as discretely countable as individual warheads are; and all of the technology is dual-use.

These discontinuities harden the information environment, too, making it more challenging for countries to make headway on international compacts amid mutual mistrust. Frontier AI expertise now lies outside the natural checks and balances of public-funded research, national laboratories, and public-sector knowledge, due to systemic problems that have self-perpetuated a brain drain from the public sector and declining independent inquiry due to the high costs of AI inputs and the resource-gated concentration of these research ecosystems within frontier firms.

We are stuck looking for proper third-party evaluators that do not fully exist. AI companies are inherently profit-seeking; profits fund the quest to build better AI, while independent researchers report difficulty obtaining the funding and access needed to meaningfully adjudicate claims. But AI companies’ duties to shareholders, motivating aggressive scaling and expansion, will always outweigh even the most well-meaning disclosure or press release on slowdowns, both in actual practice of what gets released, as well as how these disclosures are perceived nationally or internationally regardless.

And at the international level, the question of trustworthy interpretation and evaluation of industry claims haunts any attempts at navigating an agreement between countries like the US and China. As Zilan Qian argues, mistrust is a lasting social problem that cannot be techno-solutioned away, as the two state systems define a trustworthy institution differently: to sum up a wider set of complex considerations described by Qian, while independence from government reads as more credible to Washington, state assurance makes conclusions more credible in Beijing.

Who is an objective third party? How do we know that they are? Is there capacity within government or among broader swathes of the public to really evaluate whether claims being made by companies are true? Part of the difficulty is that AI has been, in Lene Hansen and Helen Nissenbaum's term, "technified,” or seen as something only technical experts can judge. Due to the qualities of frontier AI research, to be a technical expert is to be part of an extremely unrepresentative elite, mostly in a few San Francisco-based firms. Public opposition to an unchecked tech buildout, for instance, suggests many Americans reject that deference.

But outside of deferring to companies as that compromised source of truth, there is no consensus of what slowing down the frontier should look like in practice, or what third-party evaluation or verification protocol could be generally widely accepted—not least among the public and competing companies in the US, before even getting into the question of what would be acceptable internationally. To be clear, I don’t think that some imaginary singular source of truth is going to save us—as knowledge itself is inseparable from power structures—but, if the broader ecosystem made room for many kinds of knowledge production outside private industry, from better-resourced public agencies and independent researchers to local communities being empowered to wield checks and balances against the industry, we would be in a better place to test claims and decide, collectively and consultatively, how to act.

Wherever one sits within the bigger tent of Big Tech pushback, the question of a trusted evaluator and the captured information environment is clearly a common problem area. This lack of consensus fosters more difficulty at the international level. Coordination is necessary, but legitimacy and signaling credibility require trust within both domestic and international realms.

In both the US and China, an agreement abroad depends on legitimacy at home. Until both countries build some way to reliably and independently check what their companies claim, such as through public evaluators with real independence and resources distinct from labs, the “psy-op” allegation is an easy way for AI boosters of all varieties to cloud the waters. By doing so, they forestall efforts to think more critically about the rushed deployment and adoption of AI systems across increasingly critical state functions and the destructive drumbeat driving an unrelenting, resource-hungry data center buildout.