In recent years, the application of AI large models in the intellectual property sector has become increasingly widespread, with their potential in patent search and analysis drawing significant attention. However, the legal status and evidentiary admissibility of AI-Generated Content (AIGC) in patent examination have emerged as urgent issues to be resolved.
Recently, the China National Intellectual Property Administration (CNIPA) published a typical case where the core dispute centered on whether AI-generated content could be used as a direct basis for inventiveness determination.
The case involved a re-examination of a patent application titled “A Metallographic Preparation Method for Tungsten-Zirconium Alloys.” While responding to a re-examination notice, the petitioner attempted to use a response generated by an AI large language model as evidence to prove the absence of technical teachings in the prior art, thereby asserting the inventiveness of the invention. Through Re-examination Decision No. 1927206, the CNIPA rejected the re-examination request.
The decision explicitly stated that responses generated by AI large language models can only serve as a reference and cannot be directly used as a basis for evaluating inventiveness. The primary reasons for this decision include:
- Lack of Objectivity: Constrained by data sources, algorithm models, training precision, and prompting methods, AI cannot objectively reconstruct prior art in the sense of patent law, nor can it accurately reflect the cognitive level of a person skilled in the art.
- Questionable Reliability: Tests have shown that different AI models may yield different conclusions for substantially identical questions, indicating that the reliability of their generated content needs improvement.
- Inability to Replace Legal Judgment: Patent examination strictly relies on legal provisions and requires adopting the perspective of a person skilled in the art. AI-generated content lacks professional legal expertise and targeted analysis, and thus cannot replace the legal judgment of professionals.
The reasons outlined in the decision can be categorized into two dimensions:
1. Technical Dimensions
Limitations in Data Sources and Algorithms: The training data for AI large models is extensive and uncertain, potentially containing erroneous, outdated, or one-sided information, which causes the output to deviate from objective facts. Meanwhile, current algorithms struggle to perform deep modeling of technical details in specific fields, failing to comprehensively and accurately reconstruct prior art as defined by patent law.
Impact of Training Precision and Prompting Methods: Models may fail to adequately cover knowledge in certain niche fields and lack deep comprehension of complex technical issues. Furthermore, AI outputs are highly dependent on the phrasing and format of the input prompts; subtle changes in wording can lead to entirely different responses. This instability renders them unreliable for inventiveness determinations.
2. Legal Dimensions
Non-compliance with Patent Law Provisions: Patent law imposes strict standards for recognizing prior art. AI-generated content typically lacks clear sources and authoritative endorsement, making it difficult to qualify as legally compliant evidence. Its accuracy and consistency cannot be guaranteed, falling short of the rigorous requirements of patent examination. Additionally, AI is currently unable to accurately simulate the legal fiction of a “person skilled in the art.”
Inability to Replace Legal Judgment: Patent examination is a highly specialized legal activity involving comprehensive evaluations of novelty and non-obviousness, as well as complex legal logic and value judgments. AI-generated content is data-driven and lacks legal professionalism, rendering it incompetent for this task. Examination also requires considering policy orientations and the balance of social interests, which fall beyond the capabilities of AI models.
To recap, this typical case provides the following insights:
1. Exercise Caution When Using AI-Generated Content as Argumentative Evidence
According to this decision, content generated by AI large language models does not yet possess evidentiary admissibility under patent law and cannot be used as the sole core argument. In patent examination, traditional forms of evidence such as patent documents and journal articles should be adhered to, ensuring the objectivity and verifiability of the state of the prior art. Applicants and patent attorneys should avoid using AI-generated content as the sole or primary basis for argumentation.
2. Rationally Leverage AI as an Auxiliary Reference
Although it cannot be directly used for inventiveness determination, the auxiliary reference value of AI cannot be ignored. AI tools can be utilized for brainstorming, assisting in searches, and conducting preliminary analyses to improve work efficiency. For instance, they can quickly screen massive amounts of literature to help locate key information. However, their use must be guided by professionals and combined with solid technical analysis and legal expertise to complete the final argumentation. Patent practitioners should recognize the limitations of AI, verify and validate the information provided by AI, and rationally leverage AI based on professional judgment to enhance work efficiency.
By Meilian Jin