Three Fundamental Limitations Facing LLMs

Large Language Models, which form the backbone of today’s generative AI, have demonstrated exceptional performance across a wide range of language generation tasks and have played a key role in driving the growth of the AI industry. However, beneath this success lie fundamental structural limitations that raise serious concerns about the social impact and long-term sustainability of AI technologies.

In particular, the following three issues are increasingly recognized as core obstacles preventing LLMs from evolving into truly human-centric and sustainable AI systems.

---

Risk of Personal Information Leakage

Most Large Language Models are trained and operated on centralized, cloud-based servers. In this architecture, all user input, including conversations, search history, and even sensitive personal data, is transmitted to and processed on external servers. This setup presents several significant risks:

• Potential leakage of sensitive information: High-risk data such as internal corporate documents, medical records, and financial information can be collected without proper safeguards, leaving them vulnerable to exposure.

• Legal risk: There is a high likelihood of violating data protection laws such as the European Union’s GDPR or California’s CCPA. In fact, some AI services have already faced legal scrutiny due to the lack of transparency in how they collect and handle personal information.

• Loss of user control: Users typically have little to no visibility or control over how their data is stored, how it is used, or with whom it is shared.

---

High Resource Use

Large Language Models must process billions of parameters in real time, requiring enormous computational power and server infrastructure. This leads to several critical challenges:

• Soaring operational costs: To run LLMs reliably, massive GPU clusters, high-speed networks, and continuous power supply are essential. These requirements result in a heavy financial burden even for large cloud service providers.

• Unsustainable energy consumption: The energy used during inference and retraining of AI models has been widely criticized for contributing to carbon emissions, raising serious concerns about the environmental sustainability of these systems.

• The cost of personalized AI: Providing tailored AI models for individual users would require dedicated computational and training resources for each person, causing operational costs to grow exponentially.

---

Limited Personalization

LLMs are fundamentally designed as general-purpose models, which leads to several limitations in their ability to personalize:

• Lack of contextual memory: Most LLMs are constrained by a fixed token limit (for example, 4,096 or 8,192 tokens), making it difficult to continuously incorporate long-term conversation history or user context. As a result, they often produce one-off responses disconnected from prior interactions.

• Inability to reflect individual emotions and habits: It is difficult for these models to learn and adapt over time to a user’s unique language style, preferences, or emotional state. Most services only offer generic responses that are made to appear personalized.

• Repetitive and formulaic answers: LLMs tend to generate the most statistically probable responses that apply broadly across users. This often prevents them from delivering truly tailored advice or action recommendations suited to individual users.

In the end, while LLM-based services may appear to possess sophisticated intelligence, they remain structurally limited in their ability to reflect each user's specific context and needs.