THE BRIEFS: In the Era of Deepseek, PRC Law Firms Look To Address AI Anxiety Rationally

Since the beginning of 2025, Hangzhou-based AI startup DeepSeek has garnered significant global attention with its powerful AI product offerings and broad application potential, prompting various industries to embrace this emerging platform actively.

From Baidu search to WeChat and Huawei Cloud, numerous tech giants have begun integrating their products with DeepSeek’s capabilities. This integration has extended to government services as well: In February, Beijing’s Fengtai District Administration of Government Affairs and Data completed the on-premises deployment of DeepSeek’s large language model environment on its government cloud infrastructure. The district became the first to apply this technology to government services by launching the “Feng Xiao Zheng” digital assistant, accelerating the intelligent transformation of public services.

In the legal services arena, DeepSeek’s influence has become increasingly apparent. Law firms and legal tech companies have keenly identified this trend and are actively exploring deep integration possibilities between DeepSeek and the legal industry. The Hangzhou-headquartered AI company has begun gradually deploying specialised legal solutions designed to meet the dual demands of efficiency and precision within the legal sector.

At the judicial level, courts and procuratorates across various regions and jurisdictions have initiated DeepSeek training programs and even implemented deployments, aiming to leverage AI tools to enhance judicial efficiency and further advance intelligent judicial development. In March, China’s Judicial Convenience Platform successfully integrated DeepSeek to provide online legal consultation services.

DIFFERENT APPROACHES

Kevin Wang, COO of legal tech company L-Expert, acknowledges that DeepSeek’s open-source availability and commercialisation have enabled capabilities comparable to top-tier large language models at significantly lower costs, attracting widespread adoption among Chinese technology companies. Numerous law firms and legal tech providers have begun integrating DeepSeek to enhance comprehensive legal database searches, document generation, and contract review functionalities.

“Chinese law firms are showing tremendous enthusiasm for DeepSeek applications,” Wang observes. “Many lawyers are proactively experimenting and researching to find products and implementation strategies that genuinely boost their productivity.”

At the firm level, “managing partners are prioritising AI-enabled products when selecting new systems. Many firms are collaborating with us to develop and test relevant applications,” Wang adds.

Actually, L-Expert has already completed local deployment of AI large language models and is leveraging DeepSeek to enhance its product capabilities across system AI assistants, cross-database AI document searches with automatic organisation, and automated document generation and management.

In early March, Yingke Law Firm also announced comprehensive integration with DeepSeek, becoming one of the first Chinese law firms to formally connect DeepSeek with legal services. Yingke has reportedly implemented the full version of the DeepSeek-R1 inference model, aiming to introduce intelligent solutions to the legal sector.

“To optimise DeepSeek-R1 for legal applications, Yingke has established specialised teams across various practice areas to analyse our extensive industry experience,” the firm tells ALB. “Leveraging DeepSeek-R1’s technology, we have further enhanced our proprietary legal data resources, including vast repositories of lawyer profiles, case libraries, regulatory databases, and contract templates, creating a comprehensive legal knowledge framework. Through specialised training on legal terminology, principles, and reasoning, DeepSeek-R1 can more accurately understand and apply legal knowledge, providing robust support for Yingke’s legal services.”

Following its integration with the full version of DeepSeek-R1, Yingke plans to deploy the technology across seven key work scenarios to assist with corresponding legal tasks.

First, Yingke aims to construct a more multidimensional legal knowledge system through DeepSeek-R1. Previously fragmented legal resources will be consolidated into an integrated database, enabling lawyers to access and utilise professional resources more efficiently and improve knowledge management.

Second, for regulatory research, Yingke will utilise DeepSeek-R1 to implement real-time updates and precise maintenance of its regulatory database, ensuring authority and accuracy when citing legal provisions.

Notably, Yingke’s previously launched “YingFaBao AI Legal Space Station” will undergo comprehensive upgrades powered by DeepSeek-R1’s reasoning capabilities and knowledge distillation technology. The enhanced system will more precisely understand users’ legal requirements, addressing general, routine, and knowledge-based legal inquiries while significantly reducing resource consumption and operational costs.

For client communications, DeepSeek-R1 will assist in rapidly organising client inquiries, extracting keywords, and supplementing relevant information, enabling intelligent matching based on case types, geographical considerations, and lawyer expertise, thereby improving client satisfaction while reducing communication costs.

Additionally, Yingke will leverage AI technology to precisely categorise and efficiently retrieve historical cases, providing lawyers with rich reference materials to quickly understand similar case judgments and judicial tendencies, offering data-driven support for litigation strategy development.

In contract services, DeepSeek-R1’s deployment will enhance both efficiency and quality through intelligent generation and review functions, including risk identification, clause generation, and version comparison capabilities, potentially significantly improving the firm’s non-litigation service capabilities.

Finally, Yingke will utilise AI technology to implement intelligent management of multidimensional information regarding lawyers’ professional backgrounds, areas of expertise, and successful cases, optimising internal management and business allocation processes while promoting collaborative work among attorneys.

THE CHALLENGE OF HALLUCINATIONS

Similar to many general-purpose large language models, DeepSeek’s deeper application in the legal domain has triggered a series of challenges. Issues such as data security, intellectual property protection, algorithmic bias, and legal liability definition urgently need resolution, while also imposing new requirements on the regulated operation of the entire legal services market.

A typical scenario involves lawyers discovering fabricated data or even non-existent legal provisions when using DeepSeek to generate content. This raises the question: can legal professionals eliminate such issues by independently feeding DeepSeek training data to create reliable vertical domain-specific models?

Wang points out that this phenomenon, known as hallucination, occurs when models generate information that appears reasonable but is actually inaccurate or non-existent.

“This happens because models learn from massive datasets during training, but their generation mechanism is based on probability prediction rather than factual retrieval. Eliminating ‘hallucinations’ by feeding data is unlikely, while making AI rely on specified databases to answer questions could theoretically work but has extremely low operational feasibility at this stage,” he states.

The fundamental issue lies in the training methodology and architecture of large language models. Wang explains, “Responses from large models like DeepSeek are generated through recombination of their pre-training knowledge base and user-provided contextual information via complex deep learning architectures such as Transformers. Therefore, even when fed specific data, they can only improve accuracy to a certain extent without completely blocking the influence of their original knowledge base. Additionally, while deeply modifying the model’s core architecture is technically possible, the investment costs—including funding, data resources, and engineering development—are extremely high, with enormous implementation challenges.”

Yingke believes ensuring the accuracy of legal large language models is a complex process requiring approaches from algorithmic modelling, risk assessment, and data monitoring perspectives, supported by authoritative legal databases and substantial professional legal academic literature.

“DeepSeek’s hallucinations in serious contexts are issues that Yingke takes very seriously and must resolve,” the firm adds. “We ensure data quality by cleaning our proprietary data to remove errors and noise, while accurately annotating and categorising data— such as marking legal provisions with their applicable scope, and labelling cases with case types, dispute focuses, and applicable legal provisions — to facilitate model learning and understanding. Furthermore, we incorporate legal domain logical rules into DeepSeek, enabling rule-based reasoning and judgment to enhance the model’s accuracy and logical consistency when handling legal issues.”