Limitations, Controversies, and Challenges

DeepSeek has made significant strides in the AI landscape. Since its global launch in December 2024, DeepSeek has amassed over 10 million downloads and boasts 1.8 million daily active users. DeepSeek excels in areas like Chinese language processing and offers cost-effective AI solutions, but it also faces challenges when compared to OpenAI's GPT-4 Turbo and other models, particularly in advanced reasoning and real-world applications.

Limitations and biases of DeepSeek models

Despite its rapid growth and impressive user base, DeepSeek’s models exhibit certain limitations that are important to consider.

• Weaknesses in advanced reasoning and logical consistency: DeepSeek models struggle with complex logical reasoning tasks, such as multi-step deduction, mathematical proofs, and problem-solving that requires long-term tracking of variables. OpenAI o3-mini is significantly stronger in this domain due to extensive fine-tuning on reasoning-based datasets.

• Weak mathematical and scientific deduction: DeepSeek is inferior to OpenAI models in mathematics and scientific reasoning, especially when dealing with symbolic manipulation, proofs, and high-level mathematical derivations. As seen in our lesson, the DeepSeek model struggled to apply integration by parts correctly and kept outputting ”Server is busy” before finally providing an answer.

Note: One could argue that the “server busy” response doesn’t necessarily indicate weak mathematical deduction. However, during our tests, we ran multiple simpler tasks in parallel, all generating responses without issue. The “server busy” message only appeared specifically when attempting the integration example, suggesting that the challenge may not have been server load alone but rather some other limitation(s) in handling complex mathematical reasoning.

• Limited multimodal capabilities: DeepSeek’s models are primarily text-based and cannot interpret visual data. This limitation restricts their applicability in tasks requiring the analysis of images, diagrams, or other visual elements. In contrast, while OpenAI's o1 and o3 models are text-based, they can extract basic shape information, offering slightly enhanced versatility.

• DeepSeek’s identity confusion: A peculiar issue with DeepSeek models is that they sometimes “believe” they are GPT models. Users have reported instances where DeepSeek-generated responses refer to themselves as ”GPT-4” or “GPT-3.5,” likely due to residual training data or an internal alignment issue.

• Political censorship and restricted information access: DeepSeek enforces content restrictions on politically sensitive topics, particularly those related to China, leading to inconsistencies in information availability. While OpenAI models also apply content moderation, DeepSeek’s approach appears more restrictive in certain areas.

Challenges and controversies of DeepSeek models

• Bias in training data and cultural representation: DeepSeek models exhibit biases in political, sex, and cultural topics, similar to OpenAI’s models. However, DeepSeek aligns more with Chinese-based training data, whereas OpenAI models lean toward Western perspectives.

• Misinformation and hallucinations: Like OpenAI and other models, DeepSeek hallucinates information, creating fake citations, incorrect historical events, and imaginary research papers.

• Security and privacy concerns: DeepSeek’s data handling practices have been scrutinized, particularly regarding potential data transfers to third parties without transparency.

• Jailbreaking vulnerabilities: Users have discovered methods to bypass DeepSeek’s content moderation, leading to potential misuse. OpenAI models also have moderation weaknesses but are harder to bypass than DeepSeek.

• Intellectual property concerns: In late January 2025, OpenAI accused Chinese AI startup DeepSeek of using its proprietary models to train an open-source competitor through distillation.