The Chinese have been working on cognitive warfare for years. This is not particularly surprising. The People’s Liberation Army (PLA) has long understood that manipulating what people believe can be just as useful as destroying military equipment. Psychological warfare, propaganda, and disinformation have been around forever. What has changed is the technology available to conduct these operations and the possibility of automating much of the process.

Over the last several years, Chinese military researchers have been publishing papers on using artificial intelligence, large language models (LLMs), and autonomous agents to conduct what they call cognitive-domain operations (认知域作战). The objective is to influence how people perceive events, process information, and make decisions. This goes beyond convincing someone that a particular narrative is true. It includes creating doubt, exploiting existing divisions, undermining confidence in institutions, and manipulating the information environment in which people form their opinions.

The interesting part is how the PLA proposes to accomplish this. Rather than relying entirely on people to create propaganda and distribute it through social media, Chinese researchers are examining systems that could collect information about populations, identify psychological characteristics, generate tailored messaging, and monitor how audiences respond. The results would then feed back into the system to refine subsequent activity. Essentially, they are looking at turning influence operations into a continuous process supported by artificial intelligence.

In 2023, researchers from the PLA Air Force Early Warning Academy published a paper titled An Exploration of the Effectiveness Mechanisms and Key Technologies of Algorithmic Cognitive Warfare (《算法认知战的致效机理与关键技术探析》). This paper is particularly useful because it lays out how the researchers believe algorithmic cognitive warfare should work. They describe six stages beginning with profiling the target audience, capturing attention, shaping how information is interpreted, encouraging desired reactions, intervening as necessary, and monitoring the results.

There is nothing particularly revolutionary about the underlying psychological concepts. Intelligence services and propaganda organizations have used variations of these techniques for decades. What makes the Chinese research interesting is the proposed use of algorithms and AI to connect these activities and potentially conduct them at a scale that would be difficult to manage manually.

The paper discusses data fusion, behavioral analysis, deepfakes, recommendation algorithms, and generative models as enabling technologies. The goal is to understand an audience well enough to influence what information it encounters and how it responds. The system would not necessarily need to manufacture false information. It could take existing grievances, political disagreements, or legitimate concerns and amplify them in ways that serve the operator’s objectives.

This is where the more recent research on LLMs and autonomous agents comes into play.

The Advent of LLM Cognitive Warfare

In 2024 and 2026, Chinese military and academic researchers published additional work examining how language models could support cognitive operations. A May 2026 publication from the Chinese Institute of Command and Control describes cognitive warfare moving toward a stage involving large-model cognitive agents. These are systems that could maintain artificial identities, remember previous interactions, participate in conversations, and coordinate with other agents.

Consider the difference between that and a traditional bot network. Most bot networks rely on scripts, scheduled posts, and human operators to manage accounts and messaging. They can be effective, but they also produce recognizable patterns. Accounts repeat the same material, post at unusual intervals, and frequently lack the characteristics of genuine human interaction. An LLM-driven agent could potentially behave differently. It could maintain a consistent persona, respond to conversations, remember what someone said previously, and adjust its language to match the audience. Multiple agents could operate together while appearing to represent different people with different interests and opinions. The result could be an influence operation that looks less like a coordinated propaganda campaign and more like an ordinary argument unfolding across social media.

That is the theory, anyway.

Another area of Chinese research makes this more interesting. In February 2026, researchers associated with the PLA Academy of Military Science and National Defense University published papers examining the use of LLM-driven agents to simulate human behavior and social networks. These systems attempt to model how individuals interact, how groups form opinions, and how information spreads through larger populations.

The potential military application is obvious. If you can simulate a population, you might be able to test different influence strategies before using them against real people. Researchers could examine how a particular narrative spreads, which groups respond to it, and whether certain messages appear to produce the desired reactions.

Of course, there is a problem with assuming any of this will work as advertised. Artificial agents are not human beings. LLMs can generate convincing conversations without accurately reproducing human psychology. They hallucinate, behave inconsistently, and frequently reflect the assumptions of their developers and training data. The Chinese researchers acknowledge many of these limitations.

A simulated population agreeing with your propaganda does not mean a real population will follow along.

Still, the research shows where Chinese military thinking is headed. The PLA is not simply interested in generating more propaganda. Its researchers are examining whether artificial intelligence can help model the people they want to influence and predict how those people might respond.

There is also a commercial side to this.

Chinese technology companies have been developing platforms for social-media intelligence, public-opinion monitoring, behavioral analysis, synthetic-media generation, and automated information processing. Some of these capabilities overlap considerably with what the military researchers describe. GoLaxy, also known as Zhongke Tianji (中科天玑), is one company worth examining. Its products include systems for collecting and analyzing social-media information and characterizing online audiences. External reporting based on leaked company materials has described capabilities associated with its GoPro platform that could support AI-assisted influence operations.

Another company, Xianjin AI, openly markets a product called Advanced Cognitive Warfare (先进认知战). There are other firms working on public-opinion analysis, automated commentary systems, and military AI applications. This does not mean every company selling these technologies is conducting covert influence operations. Public-opinion monitoring has legitimate commercial and government applications. Neither does the existence of a military LLM automatically mean it is being used to manipulate foreign populations.

But the overlap between the commercial technology and the military research is worth examining. The PLA papers describe systems that collect information about audiences, profile their behavior, generate tailored messaging, distribute information, and monitor reactions. Chinese companies are developing technologies capable of supporting many of those functions.

The question is whether the Chinese military or intelligence services have connected those capabilities into an operational system.

The procurement records offer some clues.

The PLA has advertised requirements for LLM infrastructure, specialized AI agents, information-domain simulations, recommendation software, and AI-assisted decision-support systems. Some of these records identify budgets, technical specifications, suppliers, and development schedules. This is important because it demonstrates that the military is investing in actual systems rather than simply publishing research about what might be possible.

One project that deserves additional attention is the Hybrid Intelligent Cognitive Analysis and Decision Platform (混合智能认知分析与决策平台). Chinese project-registration reporting explicitly references military cognitive-domain operational requirements, while separate reporting associates a platform of the same name with Zhongke Fanyu. That is a potentially significant connection between military requirements and commercial development. However, the available evidence does not establish a confirmed PLA contract, operational customer, or completed deployment for the platform.

The same limitation applies to the broader procurement activity. The PLA purchasing LLM infrastructure or agent-based software demonstrates investment in those technologies. It does not establish that the military has deployed a fully autonomous cognitive-warfare system.

Meanwhile, Chinese-linked influence operations are already using generative AI.

Spamouflage, also known as Dragonbridge, has operated networks of inauthentic social-media accounts for years. Its activities have included coordinated political messaging, propaganda, and attempts to manipulate online discussions outside China. Public reporting in 2024 described operators associated with Spamouflage using ChatGPT to summarize and analyze social-media posts from overseas Chinese dissidents targeted for harassment.

That finding is particularly interesting because the AI was being used to analyze people rather than simply produce propaganda. It aligns with the audience-profiling stage described in the PLA’s algorithmic cognitive-warfare research. Other reporting has documented Chinese-linked influence activity using generative AI for content production, translation, and synthetic media. Taiwan remains a major target of Chinese information operations, with Taiwanese intelligence reporting extensive coordinated disinformation activity involving political, military, and social issues.

These operations demonstrate that generative AI is already being used in the influence environment. What they do not demonstrate is that all the capabilities described in Chinese military research have been integrated into one autonomous system. That distinction is important because the technology being proposed is considerably more ambitious than what has been publicly verified.

The real potential change is not the creation of better deepfakes or the ability to generate millions of propaganda posts. It is the possibility of combining population-level intelligence collection, artificial identities, personalized interaction, and continuous feedback into a system capable of maintaining influence operations over long periods. Such a system would not necessarily need to convince everyone of the same narrative. It could exploit existing disagreements, amplify distrust, and encourage different groups to become more entrenched in their positions. The operator might achieve its objectives simply by making it harder for a population to agree on what is happening or whom to trust. And because the system could potentially adapt its behavior based on audience reactions, the operation might become less dependent on constant human direction.

After reviewing the Chinese military publications, academic research, commercial technology, and procurement records, the evidence points toward a sustained effort to develop the components of AI-enabled cognitive warfare. The PLA has articulated the concepts, military-affiliated researchers are investigating the technical requirements, commercial companies are developing overlapping capabilities, and government procurement demonstrates investment in relevant technologies.

What remains missing is proof that the PLA or Ministry of State Security has deployed a single operational platform combining all six stages of the 2023 framework. We also lack reliable evidence that autonomous cognitive agents can consistently produce the psychological and behavioral effects envisioned in the research. The next important discovery would be a Chinese military contract, technical specification, or acceptance report identifying a cognitive-warfare platform, its sponsoring organization, the contractor building it, and its operational purpose. That would help establish whether these concepts have moved beyond research and experimentation. Until then, we should treat this as an emerging capability rather than a confirmed operational system.

There is a larger issue here that tends to get lost in discussions about AI-generated propaganda. Disinformation does not always require an outright lie. Sometimes it works by amplifying legitimate disagreements, selectively presenting authentic information, or encouraging people to distrust everything they encounter. The Chinese military understands this. Its research treats human perception and decision-making as something that can potentially be modeled and manipulated through technology. The ambition is to move beyond controlling the message and toward influencing the process through which people interpret the message. That is a much more interesting problem than another deepfake video making the rounds on social media. And it is one we should be paying considerably more attention to.