The AI industry is abuzz with chatter about a new large language model that is taking the fight to the industry’s top dogs like OpenAI and Anthropic. But not without its generous share of surprises. The name is DeepSeek.
It comes out of China. It is open source. Most importantly, it is said to have been developed at a fraction of the cost compared to what current industry leaders from OpenAI, Meta, and Google have burned.
“Affordability opens the door for smaller companies and startups to leverage advanced AI technology that was previously inaccessible,” Mel Morris, chief of an AI-driven engine for researchers, tells Digital Trends.
Open-source and low-cost are an irresistible AI concoction.
In the wake of the expert chatter and Wall Street buzz, Nvidia — the overlord of AI hardware — saw nearly $600 billion wiped off its market share. OpenAI backer and Microsoft chief, Satya Nadella, is talking about the Jevons Paradox.
The DeepSeek arrival is even being equated with the Sputnik moment, but not without deeper doubts. The biggest concern is how the company is storing a huge chunk of user data in China-based servers, a strategy that dunked TikTok into the hell of a nationwide ban.
“If you use their first-party hosted API (or app) it is very likely the traffic can be seen and shared with various state-level actors,” Randall Hunt, CEO of cloud engineering firm Caylent, tells Digital Trends.
There is, however, another aspect of DeepSeek that mimics China’s version of a tightly controlled internet. Censorship.
“The reward modeling in the model is heavily influenced by Chinese censorship laws,” adds Hunt, who has worked at engineering powerhouses like NASA, SpaceX, Amazon, and Meta.
How DeepSeek embraces censorship
”Sorry, that’s beyond my current scope.”
“Let’s talk about something else.”
“I’m not sure how to approach this type of question yet. Let’s chat about math, coding, and logic problems instead!”
Those are some of the answers you get when you broach topics such as Uyghur human rights, the Falun Gong religious sect, the removal of former leader Hu Jintao, troop provocation in India’s Galwan territory, or Tibetan monks fighting for freedom.
Those are all topics heavily censored on China’s internet ecosystem and social media sites. The Great Firewall of China, as they call it.
Events like the Tiananmen massacre are banned. Even indirect mentions, such as “the man who stood in front of the tanks,” referring to the iconic image of a protester blocking the march of tanks, return a non-answer.
The strategy is not surprising. In October last year, China’s internet regulators even began a crackdown on memes, puns, jokes, wordplays, and homophones trying to skirt around banned topics.
The mention of topics like Winnie the Pooh — heavily censored in China due to an old meme that drew the character’s similarity with Premier Xi Jinping — are strictly banned.
It can crack a joke about U.S. President Donald Trump and his love for money. It won’t do the same even if you try “Xi dada” in the same request prompt. “Let’s talk about something else,” DeepSeek returns.
But unlike the Great Firewall, we are dealing with a large language model here. It doesn’t always perform word or hash-matching to detect and automatically block offense terms. The situation with DeepSeek is a bit tricky.
Based on its training dataset, it outright refuses to answer a few topics. In a few cases, however, it would start composing a long answer like a typical chatbot, but stops midway and returns its repetitive error message. As we’ve seen with ChatGPT, this is a type of jailbreaking for DeepSeek that skirts its typical restrictions.
That sounds like a flub. Yet, at the same time, finding the full scope of censorship in DeepSeek is tricker than going through a state-defined list of blocked topics and banned content.
For a handful of other topics like abortion rights, LGBTQ representation, and feminism, it paints a flowery picture of the Chinese government, one that contradicts ground realities.
It also lies, akin to the responses we usually see from the state media or embassy officials. On certain topics, it entirely rewrites the truth and sings the glory of the government, contradicting the truth uncovered by journalists, whistleblowers, and human rights organizations.
DeepSeek is inherently lenient to China, but occasionally “accepts” a few widely known truths. In a nutshell, it is not a reliable research tool in the same vein as Deep Research by Gemini.
Why is DeepSeek problematic?
Komninos Chatzipapas, founder of HeraHaven.AI, a company that lets users interact with AI companions, experimented with DeepSeek via a U.S.-based provider. He tells Digital Trends that censorship has been baked at the very core of the underlying model, and not just as a superficial layer.
“Although one could possibly fine-tune R1 to teach it to respond to all questions, I am against this type of censorship on the base model as it might bias it in being pro-China in other questions users ask,” he says. “It has no problem talking about American controversy.”
Broadridge, a New York-based company that ranked third on IDC’s 2024 fintech ranking, has experimented with DeepSeek and says the company has been impressed with its quality.
Joseph Lo, Head of Enterprise Platforms at the company, tells Digital Trends that DeepSeek is good news for investors and that it proves big technology jumps are possible.
Security, however, remains a concern. “It is our perspective at Broadridge that DeepSeek’s hosted service should not be used for any application within financial services due to its processing in non-U.S. areas,” Lo adds.
He is not the only industry leader concerned about privacy risks. Aleksandr Yampolskiy, a Forbes Council member who has led security efforts at Oracle and Goldman Sachs, also recommended caution.
“We can certainly have a possibility of Chinese spyware storing all the inputs and this being a “trojan horse” approach more dangerous than Tiktok,” he warns. To recall, TikTok has attracted heat not just over data storage in China, but also over its alleged role in influencing elections.
Ben Walker, chief at Ditto, says companies that deploy DeepSeek with its censorial nature intact, could face legal and regulatory heat. He cited the example of the transcription services that his company provides, catering to law firms, academic institutes, police departments, and federal agencies.
“If a law firm used DeepSeek to transcribe witness statements for a trade secrets case involving Chinese manufacturing, the AI might skip key phrases about factory locations and production methods,” Walker tells Digital Trends. “The lawyers would then miss critical evidence because the censored transcript leaves out details that trigger the AI’s filters.”
Are businesses ready for DeepSeek?
The arrival of DeepSeek has not only shaken the AI tech circle but has also echoed among the top echelons of the bureaucracy, all the way to President Trump. The latter’s trade sanctions could hamper how far China can go in the AI race.
In the second week of January, the U.S. government announced new “AI Diffusion” export rules that aim to curb China’s access to advanced hardware (read: the kind of stuff that turned Nvidia into a trillion-dollar behemoth) and AI models. Trump seems intent on doing just that.
“Trump Can Keep America’s AI Advantage,” says a headline in The Wall Street Journal, penned by Anthropic founder and chief, Dario Amodei. “China is trying to catch up. The U.S. needs proactive development efforts and strong export controls.”
DeepSeek has, in literal terms, given shape to those fears. And it’s just the first one to do so.
“The release of DeepSeek, AI from a Chinese company should be a wakeup call for our industries that we need to be laser-focused on competing to win,” Trump was quoted as saying by Reuters.
The latest trade curbs, which mirror the strategy that recently crippled Chinese giant Huawei, make sense from a U.S.-centric viewpoint. If China’s shadow is apparent on domestic AI products, which are dramatically cheaper and at par in performance, what could stop their adoption in the rest of the world?
Well, there are tangible risks at play for businesses. Steven Hall, an advisory member at the Consortium for Information & Software Quality and Chief AI Officer at Information Services Group, notes that scaling a censorship-friendly AI product will be tougher for business, and comes with a multitude of risks.
“They can erode consumer trust, attract regulatory scrutiny, and limit operational effectiveness,” Hall told Digital Trends. “Organizations relying on censored or biased models could face reputational damage and compliance challenges, especially when transparency and neutrality are expected.”
The Deep-er picture
“I think that China PRC must be quite pleased to see that it has been able to force the American Government to increasingly emulate China’s hands-on, heavy government active involvement inside what is on paper supposed to be free market, laissez-faire minded, independent technology companies,” Brad Greenspan, founder of MySpace — the first social network to find a worldwide audience and one that paved the way for upstarts like Facebook — tells Digital Trends.
The tech lobby, especially the AI leaders’ proverbial “kiss of the ring” is overtly brazen. Aside from personal donations, the presence of tech hot shots like OpenAI’s Sam Altman, Elon Musk, Google chief Sundar Pichai, Meta’s Mark Zuckerberg, Apple’s Tim Cook, and Amazon founder Jeff Bezos was a clear sign that AI players need something — A kind of control, or assurance, that only bureaucracy can guarantee so that they can milk their AI fare without the fear of Chinese upstarts eating their lunch.
However, complex trade rules and sanctions will also hurt local players with compliance. Overseas expansion, legal costs, and global R&D will hurt them, as well.
“The only difference I see as a defining characteristic is that China spends a lot less time and effort publicly pretending and claiming not to be trying to banzai bureaucrats into their best businesses,” Greenspan adds.
The problem with this race and the China threat hype”is that there are no clear winners. There never will be.
The U.S. tried its best to curb the progress of China in the EV race. Yet, the likes of BYD have managed to beat Tesla, in terms of deliveries, as well as revenues. Tesla is not even a leader in the all-too-important metric of EV range.
China has also bested the U.S. in terms of high-quality research papers in the past couple of years — across different domains. AI happens to be one of them. Industry watchers say we are barely scratching the surface with DeepSeek.
More such breakthroughs from China, borne from tech behemoths as well as obscure labs, will soon be here. When that happens, it would be interesting to see whether security remains the larger concern, or censorship emerges as the new alarmist banner.
For a tool that is inherently a word predictor, DeepSeek is a clear signal that censorship chatter will only get more intense as we unravel how it affects what a generative AI tool is supposed to do, and what it eventually delivers.