Patent Law Traps Human Text in AI Detection Filters

The European Patent Office (EPO) is facing a significant challenge in identifying AI-generated content within patent applications, especially with new guidelines holding applicants responsible for LLM-assisted text. A recent study highlights a critical flaw: the very nature of patent law, which demands clear and concise claims, inadvertently makes human-written text resemble content produced by large language models (LLMs). This stylistic convergence, characterized by low perplexity and burstiness, creates a "perplexity trap." Researchers benchmarked three open-source zero-shot detectors against 500 granted human-written EPO telecom patents and 500 LLM-generated counterparts. Operating within typical consumer hardware constraints, the detectors showed alarmingly high false-positive rates for human-written claims, exceeding 60% for all tested tools. This issue persisted across different LLM regeneration strategies, model adaptations, and technical domains, suggesting a fundamental structural problem rather than a capacity limitation of the detection models. A linguistic complexity regression model improved accuracy but still faced a substantial false-positive rate.