The Perplexity Trap: When Patent Law Makes Human Writing Look Like AI

The European Patent Office (EPO) reported a record 200000 filings in 2025 and the 2026 EPO Guidelines made the applicant strictly responsible for any LLM-assisted content under Article 83 and Rule 42; together they create a hard operational pressure on patent prosecution and a near-term need for automated solution that flags suspected Artificial Intelligence (AI)-generated text for exam- iner review. However, two main challenges arise: first, the realistic infrastructure of patent offices, Intellectual Property (IP) law firms, and Small and Medium- sized Enterprise (SME) R&D departments is usually a single consumer-grade GPU, not A100-class servers which is required to run current AI-text-detectors like DetectGPT, Fast-DetectGPT, Binoculars and GPTZero. Second, the struc- tural constraint of Article 84 European Patent Convention (EPC) legally requires patent claims to be “clear and concise”, which forces human-authored claims onto the same low-perplexity, low-burstiness manifold an LLM produces by default. We benchmark three open-source detectors on a balanced corpus of 500 granted EPO H04 telecommunications patents and 500 LLM-generated counterparts (with five distinct prompting strategies), all evaluated under the consumer-hardware envelope. Every detector exceeds 60% False Positive Rate (FPR) at claim-level granularity (Binoculars 78.3%, Fast-DetectGPT 61.3%, DetectGPT 80.5%); the failure persists even when the AI side is regenerated with Qwen2.5-3B-Instruct, or when the scoring head is replaced with a Pythia-2.8B model LoRA-adapted to the EPO corpus. Cross-IPC-class extension to A61K pharmaceuticals, C07D heterocyclic chemistry and F03D wind motors and head-restored re-evaluation with the originally published Falcon-7B / GPT-J-6B scoring stacks on an NVIDIA H100 80 GB GPU confirm both: three vocabulary-orthogonal IPC classes show the same structural pattern (mean detector FPR 84.6%), decoupling the failure from any concern about substitute-head capacity. Finally, a logistic-regression classifier on seven likelihood-orthogonal linguistic-complexity features (type-token ratio, hapax- legomena ratio, dependency depth, Flesch–Kincaid grade, subordinate-clause ratio, noun-phrase density, sentence-length variance) reaches 74.0% accuracy at 28.1% FPR — a +13 pp absolute lift over the perplexity-only baseline at comparable FPR — without using any likelihood information at any stage and within the same hardware budget. The solution is the right one for patent texts and is, based on the current evidence, the only one that survives the joint constraint of Article 84 EPC and the deployment infrastructure.