DeepSeek’s rise reveals wider AI risks

The meteoric rise of Chinese AI company DeepSeek has sent shockwaves through the tech industry and financial markets. In just a week, DeepSeek's AI chatbot topped the App Store charts and created panic in the American tech sector, leading to the largest single-day market cap loss in history for Nvidia (over $580 billion).

This article examines how DeepSeek's development - from its reported £6 million development cost to its flexible deployment options - reveals broader challenges in enterprise AI adoption. These insights apply whether you're considering DeepSeek, established providers, or other emerging AI solutions.

The DeepSeek Phenomenon

DeepSeek has accomplished something previously thought very difficult in the AI world: creating models that rival industry leaders while claiming to spend just £6 million on development. Their V3 model claims to match the performance of GPT-4 and Claude 3.5 Sonnet, whilst their R1 model reportedly outperforms OpenAI's specialised reasoning tools. To put this in perspective, Meta reportedly spent over £1 billion developing their Llama models.

If we take a step back from the details, there are three main reasons why DeepSeek has caused such a great impact:

Performance and Cost

DeepSeek’s models are easily accessible for free, can be used by programmers at roughly 10 times less the price of other models, and were apparently made for a fraction of the cost as other frontier models.

Openess

DeepSeek claims its models are open, as users can download and deploy its models and also adjust the model weights. However, this should not be confused with true open access. The training data, algorithms, and methodologies remain proprietary and undisclosed.

Flexibility

Some versions of DeepSeek can run locally on modest hardware, eliminating the need for external server connections – a significant advantage for privacy-conscious users and organizations. However, it's worth noting that most users currently access DeepSeek through its cloud-connected app and website, similar to traditional services like ChatGPT and Gemini.

These aspects of DeepSeek - its performance claims, partial openness, and flexible deployment - have sparked important discussions in the AI industry. However, the implications stretch far beyond just one company.

Universal Challenges Across AI Models

DeepSeek's rapid rise forces us to confront fundamental questions about how we evaluate and implement AI tools. As Trilateral Research's AI ethics researcher Dr Zachary Goldberg explains:

 While much of the discussion around DeepSeek focuses on ethical concerns, these issues are not unique to one model. The real challenge isn't just how we use DeepSeek. It's how we, as a society, decide to use large language models as a whole. Businesses must recognise that transparency, bias, and security risks exist across all AI tools, and responsible AI adoption requires addressing these challenges industry-wide.

Let's examine these universal challenges in detail.

1. Transparency

When businesses use AI models, they're often working with what data scientists call a 'black box' - you can see what goes in and what comes out, but not how decisions are made. This creates real challenges that manifest differently across platforms:

DeepSeek has released their model weights (the mathematical rules the AI uses to make decisions), but like all AI models, the actual decision-making process remains hidden. This means whilst businesses can download and use the model, they can't understand why it makes specific choices.

ChatGPT takes a different approach. OpenAI provides extensive documentation about safety measures and publishes research papers, but keeps the system closed. For example, when ChatGPT generates a business proposal, there's no way to verify how it determined the recommended strategy or pricing.

Gemini offers regular transparency reports and technical papers but keeps its core training data and methods private. In practice, this means businesses can read about how the model should work in theory, but can't verify this in their specific use cases.

These limitations create real business risks. Consider a bank using AI for loan decisions. Without understanding how the AI weighs factors like credit history versus current income, they can't ensure their lending practices are fair or compliant with regulations.

2. Bias

Recent events have shown how each AI model comes with its own set of restrictions and biases that can significantly impact business use:

DeepSeek's content moderation reflects Chinese regulations. It has been observed to engage in self-censorship and exhibit biases on politically sensitive topics related to China. This means certain topics, like discussions about Taiwan or specific market regulations, are automatically restricted.

ChatGPT's content moderation system has its own challenges. While it uses human moderators to refine its filters, legitimate business queries sometimes get blocked. Cybersecurity firms have reported difficulties getting advice about network protection because terms like "firewall" and "penetration testing" trigger safety filters.

Gemini recently faced criticism when its image generation system produced historically inaccurate representations, revealing biases in its training data. This raised concerns about using Gemini for marketing materials or historical research.

These limitations affect everyday business operations in ways many don't expect:

• A market research firm might miss crucial insights if certain regulatory terms are filtered out
• An investment analyst might receive incomplete information about international markets
• A customer service AI might show subtle biases in how it responds to different types of complaints

3. Data Protection and Security Risks

Recent security incidents have highlighted how different AI models handle sensitive information, creating distinct challenges for businesses:

All cloud-based platforms have faced their own security challenges. In March 2024, ChatGPT temporarily shut down its service when 1.2% of users discovered they could see other people's conversations. This incident revealed how even major providers can experience significant data leaks. Similarly, DeepSeek's database exposure in early 2024 leaked over one million records, including user chat logs and API keys.

These security concerns have prompted strong regulatory responses:

• In March 2023, the Italian Data Protection authority temporarily blocked ChatGPT. They have since fined OpenAI $15.6 million for alleged GDPR Violations.
• On January 31 2025, Italy ordered an immediate ban on DeepSeek after the company allegedly refused to cooperate with an information request.
• The US Navy and Congress have banned DeepSeek use entirely, citing security risks.
• The EU AI Act includes several provisions regarding data handling for AI systems.

For businesses, these incidents create difficult choices. A healthcare provider handling patient data might prefer local deployment despite higher costs. Meanwhile, a retail business might accept the risks of cloud platforms for their less sensitive customer service operations.

Creating a Balanced AI Strategy

Successfully implementing AI requires a comprehensive approach that accounts for these universal challenges:

• Choose the Right Tool Use enterprise versions of AI tools when handling business data as free versions lack security controls and data retention policies.
• Know the Limitations: Understand that AI excels at pattern recognition and content generation but struggles with calculations and causation, requiring expert verification of outputs.
• Protect Your Data: Never input confidential business information into free AI tools and carefully analyse how seemingly harmless data could be combined to reveal sensitive details.
• Monitor for Bias: Regularly evaluate AI outputs across different demographics and maintain diverse human oversight to prevent stereotypical representations.
• Set Clear Rules: Create specific policies for approved AI tools, defining information sharing boundaries and establishing review processes for AI-generated content.

DeepSeek's rapid rise demonstrates both the potential and risks inherent in AI adoption. Whether choosing DeepSeek, ChatGPT, Claude, or Gemini, businesses must understand that each model comes with its own set of challenges around transparency, bias, and data protection.

Success lies not in finding a "perfect" AI model, but in implementing robust governance frameworks that account for these universal challenges. By understanding these shared risks and implementing appropriate safeguards, businesses can harness AI's benefits while protecting themselves from potential pitfalls.