ChatGPT enhances patent drafting efficiency, yet legal risks remain

A newly published peer-reviewed study has found that ChatGPT’s latest version, GPT-4o, demonstrates substantial promise in generating technical and legal content for patent applications in the building sector, but it still requires expert supervision to meet professional standards.

The research, titled “How ChatGPT Is Shaping Next-Generation Patent Solutions” and published in Buildings (2025, 15, 2273), analyzed how GPT-4o performs in drafting patents for prefabricated integral buildings (PIBs), an area known for its technical complexity and high legal precision requirements.

How well can GPT-4o write patent claims?

The study explores whether GPT-4o could reliably draft high-quality patent claims and technical descriptions for volumetric modular construction technologies, also known as prefabricated integral buildings (PIBs). These patents are structurally dense, often describing factory-produced three-dimensional modules with integrated mechanical and structural systems.

Researchers conducted an experimental evaluation using 230 real-world patents from China and Russia. GPT-4o was tasked with generating patent claims based on input elements such as titles, abstracts, and figures from each case. Initial drafts were reviewed using a five-level rubric, and those rated “Good” or higher were passed to human experts for further evaluation. If not, the input was iteratively enriched with component lists, technical specifications, and connection methods until the model’s output improved.

Statistical analysis revealed that GPT-4o significantly outperformed competing large language models (LLMs) like Claude Sonnet 4 and Gemini 2.5 Pro in error detection tasks. It achieved a precision of 0.90, a recall of 0.58, and an overall F1 score of 0.70 in identifying semantic and logical inconsistencies. Expert feedback confirmed that GPT-4o-generated claims generally demonstrated high linguistic quality and structural coherence. Nonetheless, recurring issues included redundancy, ambiguity in descriptions, and vague technical boundaries, which diminished the claims' legal robustness.

What are the strengths and limitations in real drafting scenarios?

The model’s real-world effectiveness was tested through a hybrid methodology combining expert reviews and statistical modeling. GPT-4o was guided through a structured prompt system and refined iteratively. Texts were evaluated across five core dimensions: quality of claims, quality of description text, linkage of features to disclosure, linkage of text to figures, and auxiliary sections such as field of application.

Regression analysis revealed that claim and description quality were the strongest predictors of overall expert approval. In both the Russian and Chinese datasets, these two variables had statistically significant impacts (p < 0.001), suggesting that GPT-4o is especially capable in generating legally meaningful and technically sound sections of a patent. In China, the linkage of text to figures also emerged as significant, indicating an emphasis on visual-verbal consistency in that jurisdiction.

Despite these strengths, the study emphasized key weaknesses. GPT-4o frequently introduced repetitive language and failed to provide detailed implementation steps. In legal contexts, where claim clarity and disclosure sufficiency are critical, such shortcomings could undermine a patent’s enforceability. Moreover, the model often generated multiple superficially different technical solutions within the same claim, creating confusion about the invention’s novelty and scope.

Concerns were also raised about the reliability of self-evaluation practices. The model was not allowed to assess its own output; instead, evaluations were conducted manually by researchers and validated by domain experts. This approach helped offset potential bias but underscored the necessity of independent human review in AI-assisted patent workflows.

Can GPT-4o replace human patent drafters?

The authors argue that while GPT-4o can streamline the early stages of patent preparation, it cannot yet function autonomously in high-stakes legal environments. Instead, it serves best as an augmentation tool, accelerating drafting and translation tasks but requiring human oversight to ensure legal compliance, technical adequacy, and textual precision.

Key recommendations from the study include:

• Integration of Expert Review Mechanisms: Human experts must be incorporated into the patent drafting pipeline to validate and refine AI-generated content.
• Expansion of Training Datasets: Future improvements to LLM performance should involve broader, more diverse patent samples, especially across different jurisdictions and technical fields.
• Implementation of Rigorous Review Workflows: Redundancy filters, ambiguity checks, and legal adequacy audits should become standard practice in AI-assisted drafting.
• Enhanced Evaluation Frameworks: Combining automated scoring systems with third-party expert panels can increase transparency and reduce subjectivity.

In terms of cross-national applicability, the study noted cultural differences in patent assessment. For example, Russian evaluators placed greater emphasis on the completeness and practicality of “Other Text” sections like objectives and application fields, whereas Chinese reviewers prioritized technical descriptions and text-figure coherence.

Overall, the results signal a turning point in AI’s role in intellectual property services. With further refinement, ChatGPT and similar models could help democratize patent services, reduce drafting costs, and accelerate innovation cycles, provided that robust guardrails are in place.