DeepSeek, China’s potentially ground-breaking AI model, has been making headlines and raising eyebrows.
DeepSeek is the name of a free AI-powered chatbot that has many of the same features as ChatGPT. But while “all ink is good ink,” the publicity has been a two-edged sword for the controversial AI upstart.
Like when any new AI model is released, many organizational leaders are asking, “Should my company use DeepSeek?”
Security and privacy issues may complicate the answer.
Potential Risks Associated with DeepSeek
DeepSeek’s models can collect and process data free from legal fetters. For instance, an AI model built by a company in the U.S. generally can’t collect personal identification information (PII) without the user’s consent. In contrast, those built abroad don’t have to concern themselves with U.S. state laws.
DeepSeek’s privacy policy doesn’t explicitly say they won’t collect data without user consent. It merely says a user “may have a right to access, change, oppose … withdraw [consent], or limit” how DeepSeek collects their data.
For some personal and business users, DeepSeek’s vague consent policy may feel inadequate.
From a data security perspective, many are having second thoughts about trusting DeepSeek. Their servers exist in China, which means the Chinese government — or the people it employs — has the power to easily access them. Given China’s questionable record regarding human rights and privacy, some have expressed concern.
For instance, suppose you’re a developer for a fintech company, and you’re coding software designed to automatically manage sensitive user information. Instead of using ChatGPT or Gemini, you hop on DeepSeek to get help writing your code. The code DeepSeek produces now lives on a server in China. If the Chinese government wanted to hack the software you created, they could simply hand the code to a software engineer and say, “Please find the vulnerabilities in this code.”
On a larger scale, a company may choose to use DeepSeek to build a code generation tool, similar to solutions provided by companies such as Replit and Qodo. Every user who turns to the solution to build code would have their programming stored on a server in China, which may introduce more risk than a programmer is willing to accept.
Censorship and Political Bias Issues
DeepSeek’s model has built-in censorship mechanisms, such as those that eliminate content around controversial topics like the Tiananmen Square killings or the Chinese treatment of Uyghurs. It also puts a pro-China and anti-U.S. spin on the content it produces with “Sorry, that’s beyond my current scope. Let’s talk about something else,” a common refrain for politically sensitive topics.
The model can change or suppress information that doesn’t align with its backers’ political leanings.
Beyond censorship, Chinese manufacturers have a long history of copying U.S. designs, producing similar products, and then selling them for far less money. DeepSeek not only puts China in a leadership position in the AI market but also enables AI developers with shallower pockets to create products at a fraction of the cost they’d pay OpenAI or other U.S. companies.
While they may seem similar in terms of capabilities, a few key differences can help leaders better evaluate DeepSeek.
What’s the Big Difference Between DeepSeek and Other Models?
DeepSeek claims its AI model is as powerful as OpenAI’s models, yet they reportedly built it for only $5.6 million. While this is being refuted, the announcement had a ripple effect that felt like a tidal wave. NVIDIA’s stock and other tech heavyweights dipped as news of a cheaper way of producing AI splashed across the ticker.
Let’s compare the models.
Deepseek’s Training Method
The training method is one of the biggest differences between DeepSeek and OpenAI. While OpenAI’s o1 models use large-scale Supervised Fine-Tuning (SFT) combined with reinforcement learning, DeepSeek started R1-Zero using only reinforcement learning — a first for open-source models.
On the other end of the AI-training spectrum is a dense or monolithic model. With this approach, a huge stack of computational resources joins forces to solve each problem in the training process. If you think of each computational unit as a node in a network, imagine every single node firing up simultaneously and all of them working together to train a single facet of an AI model. That process consumes much more energy.
Mixture of Experts Model
Another differentiator is that DeepSeek used the mixture of experts (MoE) model to build its solution, which may have resulted in lower computational costs. Unlike traditional AI models, where every part processes all tasks, MoE divides the model into specialized sub-networks, each focusing on specific problems. A routing system then directs incoming tasks to the most appropriate expert, ensuring efficient processing.
In other words, a model specifically designed to handle a certain type of AI training gets assigned to do what it does best. Only this single “expert” must expend energy toward building the element of the model. That saves money.
To illustrate, imagine you have a team of specialists ready to build 10 houses. Some are electricians, others specialize in flooring, others are plumbers, some are framers, and so on. You tell all of them to start working on the framing.
Even though they all try their best, the electricians and plumbers expend far more energy — and make far more mistakes — while framing walls than those who specialize in the role. Similarly, the plumbers, even if they try their best, may struggle with installing three-way switches, choosing the right size breakers, and other electrician-specific tasks.
On the other hand, suppose you assign your framers, plumbers, electricians, flooring specialists, and others to teams. You then tell them to tackle only the tasks they specialize in. They would expend far less energy and complete the job much faster. As the general contractor, you would save considerable cash along the way.
By using an MoE approach, DeepSeek did the same thing. Thus, they produced a high-quality product more efficiently.
The Cost of Data
There’s a good chance DeepSeek obtained high-quality data at a fraction of the cost of other AI leaders, and OpenAI says it has proof DeepSeek used OpenAI data to train its model. Granted, no one knows exactly how much OpenAI and others have paid for the rights to use data from encyclopedias, newspapers, medical journals, and more, but the price tag may have been significant.
DeepSeek may have gained access to equally robust datasets for very little money. For example, since the company has the Chinese government’s support, it could’ve had a government agency foot the bill for costly data. Then, it could run its MoE models on data that cost little to nothing.
Government Support
DeepSeek may be getting significant financial support from the Chinese government. For a private sector organization, such as OpenAI and other leading tech companies, it can be tough to secure government funding because of the fear of taxpayer backlash.
Reinforcement Learning
DeepSeek used reinforcement learning, also called group relative policy optimization (GRPO), which may be more efficient than traditional AI-training methods. The model is instructed to achieve a specific goal with goal optimization reinforcement. Once it has hit its milestone, it succeeds. On the other hand, many AI models use a complicated reward system to gradually guide AI models toward success. In some cases, this may be less efficient.
It can take a talented team of data scientists significant time to engineer the right reward system for motivating an AI model, and it’s easy to get it wrong. But by using goal optimization reinforcement, a team could simply chunk a larger task into several smaller, easier-to-achieve goals. This can save time and energy for both the team and the computers working on the model.
The potential risks and differences between the AI models lead us to our overall recommendation for leaders weighing their options with DeepSeek.
Recommendations for Businesses
Due to security and privacy concerns, we advise businesses to avoid DeepSeek’s hosted version now, with one caveat: DeepSeek R1 can be deployed through Microsoft Azure AI Foundry.
This Azure-hosted version offers enterprise-grade security controls and allows organizations to control their data while still using DeepSeek’s capabilities. While the model’s built-in censorship controls remain, this version gives organizations control over their data location and security.
When evaluating DeepSeek and other AI tools:
- Conduct thorough bias explorations
- Build a team to identify and mitigate biases in the model’s results
- Evaluate privacy, censorship, and other risks to see if potential cost savings justify using the model
For better or worse, DeepSeek represents a shift in how AI solutions can be developed and deployed. Traditional enterprise approaches have required upfront investment and infrastructure, but alternatives like DeepSeek may become more prevalent as the industry evolves. This opens the door for broader market participation, fundamentally reshaping the AI development landscape.
We help companies maintain their competitive edge by partnering with them on their AI journey. Our recent work with a wholesale food delivery company shows how.
Our experts guide leaders through the disruption AI may cause to help you go from uncertain to excited about the potential of using AI. Ready to get started? Contact Us
