Artificial intelligence is becoming a bigger topic for each passing day. While AI has existed for a long time, generative AI has only been accessible to a greater public for a relatively short time. The release of ChatGPT in 2022 has not only had an influence on our everyday tasks but also changed out perception of AI and what it is capable of. Never before has it been so human.
In 2025, Chinese chatbot DeepSeek was released to a global audience and saw instant success, becoming the number 1 most downloaded app on the American Apple App store upon launch. What fuelled this success? What is the political background behind GenAI development? What role do GenAI models play in global politics? In this article, I will be looking into the significance of the ongoing US-China trade war for AI development, the development lanscape in China, and potential risks and concerns related to using Chinese GenAI.
The inherent political nature of AI and technology
With Chinese AI patents making up a hefty 40% globally, China has become a leading market in AI development. The Chinese tech sector thrives thanks to government support, having defined AI as an integral part of the “Made in China 2025” strategy. The domestic talent pool provides businesses with graduates from high-standard universities. Major industry pioneers such as Baidu and Alibaba have already included GenAI development in their business strategies to stay at the forefront of innovation.
However, the road to innovation has not been easy. The MIC 2025 strategy has been perceived as a threat by some Western countries. Chinese officials claim that high-tech industries have expressed hostility over the initiative as it would elevate China from being a low-cost manufacturer to a global competitor who simultaneously shuts out international competition domestically. In 2018, the United States launched a trade war against China by increasing tariffs, imposing sanctions, and a semiconductor chip ban aimed at halting AI and technology development, meant to address the “unfair advantage” China enjoys through the initiative – while the US and EU scramble to catch up.
A focal point of the sanctions is semiconductors. They play a central role in civil development and national security, as they are necessary for essentially every sector, including electric vehicles, fintech, smart cities, and not to mention military development. The chips are dual-use by nature, which means that US sanctions targeting military development will affect the development of the entire Chinese economy as it is not possible to target only military use semiconductors. Summarised, the semiconductor industry is dominated by a handful of companies from the Netherlands, Japan, Taiwan, South Korea, and the US. While these countries are allied with the US, they have joined in the sanctions to different degrees.
The sanctions propelled China into developing its own semiconductor production at a neck-breaking speed and massive state fund investments. 13 out of the 20 semiconductor companies with fastest growing revenue in 2024 focused on computing power, storage, and sensor research and development, which are key for AI training. Still, China remains reliant on parts and technology imports for the foreseeable future, as their locally produced AI chips still underperform compared to American semiconductor producer Nvidia.
The success of the US sanctions is debatable, as China has reached a leading position in 6G, EVs, and now AI development. The Chinese state and economy model allows for resource allocation to selected industries and companies, so-called “champions” which will spearhead innovation and development in conjunction with the state, and in the case of AI development, it is tech giant Huawei. Any major Chinese company or enterprise is bound to be tightly connected to the state.
The April 2025 Politburo study session, where Chinese politicians and officials study a topic chosen by the General Secretary Xi Jinping, was dedicated to AI. The emphasis was on China to become self-reliant and self-strengthening, and to build an ecosystem of hardware and software, making it independent and controllable. This marks a shift in Chinese policymaking in regard to AI, which previously called for international cooperation.
What do we mean by GenAI?
Generative AI (GenAI) is mostly designed as chatbots and responds to users’ prompts, questions, and requests. They are mostly used as an alternative to search engines, providing detailed answers as it combines search results from multiple engines and websites and presents them in easy-to-understand summaries. They can also be used to generate codes or images, and are popular among the general population.
This article focuses mainly on large language models (LLMs) and large reasoning models (LRMs), as they make up the most popular GenAI models such as ChatGPT, Claude, Gemini, and DeepSeek.
A brief overview of GenAI in China
While some of the most used GenAI models in the West, like OpenAI and Claude, are built as closed-source models, there are also open source models such as Llama2 by Meta. Open source means that under terms, an AI structure is available to use, study, modify, and share with or without modifications. Some model use is restricted by licenses and training data kept secret. Chinese DeepSeek, Kimi AI, MiniMax, and Ernie are or are planned to become released as open source.
Proponents of open source models argue that it makes innovation more accessible, safer, and more transparent. Building open source models could improve the (especially Western) public’s trust of Chinese GenAI. However, concerns regarding data handling, security, and censorship are still key areas to address. The Cyberspace Administration of China, the national internet regulation agency, has emphasised the need for AI to be reliable and controllable, and bots must uphold “core socialist values”.
The Data Security Law of the People’s Republic of China classifies data and defines protection measures for different categories. However, its vague definitions have discouraged foreign investment and co-investment in different research areas because of concerns regarding access to data and potential conflicts with local data laws. This leads to a decline in knowledge exchange between countries.
DeepSeek // HighFlyer
Perhaps the most well-known GenAI outside of China, DeepSeek released its rival to ChatGPT, DeepSeek R1, in early 2025. Just like the OpenAI o1 model, R1 is a reasoning model, meaning responses are produced incrementally, just like humans think. DeepSeek claimed that it was able to train its model at a cost of USD 5.6 million, compared to the USD 100 million it took OpenAI to reach the same model quality. The launch rattled investors across the globe, and Nvidia, tightly linked to AI training, lost billions in value on the stock market. Upon launch, DeepSeek’s AI assistant became the most downloaded app on the Apple App store in the US.
DeepSeek is financed and managed by Chinese hedge fund HighFlyer. Through neural networks, natural learning processing, and machine learning, DeepSeek has launched different tools specialising in coding, content creation, and conversational AI. “DeepSeek’s key innovation was its pioneering use of a technique called “sparsity,” which makes AI models more efficient by activating only relevant portions of the model to answer a specific query, much like the human brain fires up specific neurons for a task.“
Kimi AI // MoonShot (Alibaba)
In July 2025, Moonshot released its Kimi K2 model. Low-cost and open source, the large language model’s strength lies in coding. The only use condition is that “Kimi K2” is to be displayed on a user interface if a commercial product or service it was used for has more than 100 million monthly active users, or reaches USD 20 million in monthly revenue.
Qwen // Alibaba
Tongyi Qianwen (Qwen) is a series of large language models and multimodal models developed by Alibaba Cloud. Built on Meta’s Llama, it was released to the public in September 2023 after governmental approval. Alibaba has released more than 100 open source models, yet some remain closed. The latest Qwen 3 model family is licensed under Apache 2.0, which is an open source license.
MimiMax-M1 // Alibaba, Tencent
MiniMax was founded in 2021 and has attracted investments from tech giants such as Alibaba and Tencent. The MiniMax-M1 was launched in 2025 and is an open weight (refers to the final weights and biases of a trained neural network), large-scale, hybrid-attention reasoning model. The M1 is also said to be more effficient than DeepSeek.
Ernie Bot // Baidu
Similar to ChatGPT, Ernie is a chatbot developed by tech giant Baidu, which also hosts the major search engine in China. In development since 2019, it was released for testing in 2023 and made available to the public in August the same year.
SenseNova // SenseTime
SenseNova 5.5 was marketed as a direct substitute to Chat-4o upon its release in 2024. In June 2025, SenseTime claimed that the new SenseNova 6.0 outperformed Chat-4o on several parameters.
Foreign GenAI models not available in China
OpenAI’s ChatGPT is not officially available in China. Access has been blocked by the government, but it is noteworthy that OpenAI blocked users in China from using their tools in 2024, even taking steps to block API traffic. This once again raises concerns about the right to access technology, and who is controlling other countries’ developments, as this is perceived as an extension of the tech trade war. But OpenAI’s definite exclusion of the Chinese market has provided ample opportunities for local companies.
Claude and Gemini are also officially unavailable. However, some restrictions are bypassable with a VPN.
Apple AI is not available in devices purchased in mainland China. Earlier this year, Alibaba teamed up with Apple to bring AI to iPhones, but the US Trump Administration has scrutinised the deal, citing potential national security threats. Not only could the deal help Chinese firms improve their AI, but positions Apple in a precarious situation when it comes to censorship and data-sharing laws in China.
Is Chinese GenAI safe to use?
Many military and political concerns have risen in regard to Chinese GenAI. Australia has banned DeepSeek from government devices due to security risks. Additionally, Chinese GenAI models are known for being subject to censorship under Chinese supervision. This includes political and historical subjects. Chinese chatbots are subject to censorship in accordance with Chinese laws and regulations. This means that they do not provide any information about sensitive topics such as political leaders, drawbacks of socialism, or historical events related to Tiananmen Square and Hong Kong protests. According to Wikipedia, Ernie even claimed that Covid-19 originated among American vape users.
In 2024, Baidu refuted claims that Ernie is linked to Chinese military research, as an academic paper had claimed that the People’s Liberation Army cyberwarfare division had simulated military response plans using Ernie Bot.
In June this year, Reuters reported that Chinese GenAI DeepSeek is supporting Chinese military and intelligence operations, according to a US military official. The allegations also include claims that DeepSeek has provided user information and statistics to the Beijing surveillance apparatus in accordance with Chinese law. But it is worth noting that ties between tech sectors and military use is not exclusive to China. In the same month, American OpenAI (the company behind popular ChatGPT) won a USD 200 million defence contract after quietly removing any anti-military use clauses from their user guidelines earlier this year. But it doesn’t end there – also in June, senior figures from Meta (Facebook, Instagram, WhatsApp), Palantir, and OpenAI were sworn into the US military, as it aims to “fuse cutting-edge tech expertise with military innovation”.
How might this affect us if we don’t use it for research purposes? Investigative journalist Christo Grozev tried using DeepSeek to summarise his investigation on Russian activity in the cyberspace aimed to destabilise Ukraine before the full-scale invasion in a narrative tone for a YouTube video script. But DeepSeek changed large parts of the content, completely rewriting certain parts and suddenly talking about the West’s cancellation of Russian culture, and how it will backfire on the West.
All chatbots like DeepSeek and ChatGPT are known to hallucinate and sometimes get things completely wrong. The possible reasons are plenty, be it incorrect decoding, inaccuracies and biases in training data, overfitting (when a model can’t draw logical conclusions from anything except its training data), and model complexity. This can lead to nonsensical answers and patterns that are not in line with human reasoning.
So to check if this was another instance of typical GenAI shenanigans, I submitted the same prompt that Grozev used (without taking credit for the article) into ChatGPT to see what it could produce.
From my understanding, ChatGPT has provided a summary that is truer to the contents of the original article. Below is the full response ChatGPT provided:
Here’s a refined, podcast-friendly version of the article. It’s been smoothed out for clarity, flow, and listener engagement—without any other changes.
—
**\[Intro music fades]**
**Host:**
Welcome to today’s deep dive: *Hidden Bear: Inside Russia’s Unit 29155 Cyber Squad*. We’re uncovering how the same GRU cell behind infamous poisonings and bombings also ran high-stakes cyber-operations long before Russia’s full-scale invasion of Ukraine.
—
### 🎙️ Episode Outline
1. **Unit 29155 in the Shadows**
GRU’s Unit 29155—a kill-and-sabotage squad—did much more than physical ops: they built a cyber wing. Six hackers were indicted in August 2024 for attacks on Ukraine’s infrastructure and NATO systems, including Poland’s transport network—under the ransomware guise “WhisperGate.”
2. **Origins and Key Players**
The unit debuted in 2012, led by Chechen-born hacker “Key” (Timur Stigal), who was tapped by GRU to coordinate false-flag cyberattacks. His teenage son was later wrongfully tied to the work, but evidence suggests it was the elder Stigal at the helm.
3. **False-Flag Campaigns Galore**
* Polish telecoms and medical data leaked under “Right Sector” persona.
* Credit-card data blamed on “Anonymous Poland” to smear Bellingcat.
* Early disinformation campaign targeting WADA, FBI, Boeing, and the U.S. Paralympics.
They even ran a Bulgarian “Anonymous” account to spread doctored documents and staged protests.
4. **Journalist Puppet**
They recruited Bulgarian investigative journalist Dilyana Gaytandzhieva to publish hacked embassy emails, amplifying them through her outlets. She later founded *ArmsWatch.com*, their own disinformation portal, mimicking Bellingcat’s style to control leaks.
5. **Recruits & Infrastructure**
By 2019–2020, Unit 29155 shifted to scouting talent at hacking contests. They paid top recruits, like student Vitaly Shevchenko, hefty salaries. One of them compromised Estonia’s Defense Ministry. Dmitriy Denisov, a core hacker, even had multiple online identities and openly used extremist language in Telegram chats—leaving a massive digital trail.
6. **Hybrid Warfare Tactics**
In late 2021, targeted propaganda and sabotage campaigns ramped up: fake graffiti, false narratives stoking Ukrainian nationalist tensions, even a Molotov cocktail attack. They set up spoofed government websites and encrypted-payment schemes, recruiting saboteurs across Europe via Telegram.
7. **Personal Corruption in Cyber Command**
Stigal fell ill and passed leadership to Roman Puntus. Puntus, meanwhile, funded a shell company—named after his mistress—to drive hacking ops and even drone production. Their setup was riddled with nepotism, mismanagement, and personal scandal.
8. **Final Shift and Exposure**
In early 2022, Unit 29155 widened its cyber campaign across Europe—including attacks on Ukrainian government, energy firms, and universities. But their earlier physical sabotage teams—like bomber squads—received much more attention. By 2023, many of their servers were abandoned or exposed by hacktivists, revealing the full scope of their operations.
—
**Host:**
From poisoning ex-spies to planting digital bombs, Russia’s Unit 29155 proves there’s no line too thin for hybrid warfare. A GRU black-ops unit that kills indiscriminately—and hacks relentlessly.
Thanks for joining us. If you want a deeper unpack on any twist—just let us know. Until next time, stay sharp.
**\[Outro music plays]**
—
That’s our podcast-style edit: clearer structure, listener-oriented flow, and emphasis on story arcs and characters. Let me know if you’d like tweaks on tone, length, or additional narrative elements!
ChatGPT showed no hesitation in relaying the information in Grozev’s article, quite a stark contrast to DeepSeek’s creative retelling thereof. From a very surface-level comparison, it does seem that DeepSeek might follow stricter rules in terms of ideological position when it comes to Russian-Ukrainian politics.
What to keep in mind
Chinese censorship laws and regulations are a clear example of how politics influence technology available to the public. But restrictions in responses from chatbots is not the only concern when it comes to GenAI, instead, it could even be risky to limit critique towards AI models to their reasoning capabilities. We should keep in mind the close ties of major tech companies, including Western and perhaps most notably American ones such as Meta and OpenAI, which have close ties to the military, and how we may contribute to the development of military technology when it does not align with our own democratic morals and ideals.
Removing user agreements forbidding any military use and working together with the US military blurs the lines between cilivian users and training data used for military purposes. OpenAI’s close ties to Microsoft can be seen as problematic as Microsoft has supported the IDF during the war in Gaza. Similarily, China is supplying Russia with drone components and tools used in the invasion of the Ukraine. The Chinese army is already using DeepSeek in its hospitals, and Chinese academics have reportedly claimed that it can be deployed for military use as it is capable of running military simulation scenarios.
There is no transparency of what happens with the data collected by DeepSeek, but due to local laws and regulations, we can assume that collected data can be used for any purpose by Chinese companies and even governmental organs. According to DeepSeek’s privacy policy by January 2025, user data will be stored on servers in China unless the software is run locally. What’s more, security vulnerabilities on the platform might also expose collected data to external cyber threat actors or expose data in data breaches.
Of course, advanced GenAI models publicly available are quite a new phenomenon and we will have to see how the landscape evolves over time. In the meantime, a good rule of thumb is: if it’s free, you’re the product.
Want to read more on this topic? Here are some suggestions:
Assessing the risks and opportunities posed by AI-enhanced influence operations on social media by Rolf Fredheim, James Pamment
How China’s AI Development Remains Resilient Despite Escalating US Chip Export Restrictions by Andy Liao
China’s drive toward self-reliance in artificial intelligence: from chips to large language models by MERICS
Ethics and regulation of AI in defence technology: navigating the legal and moral lanscape by Adrian Toutoungi and Dr. David Klein LL.M.