AI detectors flag 78% of human patent claims as AI-generated

A paper titled "The Perplexity Trap: When Patent Law Makes Human Writing Look Like AI" (arXiv, June 2026) reveals a critical flaw in current AI-generated text detectors when applied to patent documents. The author, Anubhab Banerjee, tested three open-source zero-shot detectors (Binoculars, Fast-DetectGPT, DetectGPT) on 500 real EPO H04 telecom patents versus 500 LLM-generated counterparts using five prompting strategies — all running on consumer GPUs (8 GB VRAM). At the claim level, all detectors exceeded 60% false-positive rates: Binoculars at 78.3%, Fast-DetectGPT at 61.3%, and DetectGPT at 80.5%. The failure persisted across model regenerations (Qwen2.5-3B-Instruct), LoRA-adapted scoring heads (Pythia-2.8B), cross-IPC replication on A61K, C07D, and F03D patents (mean FPR 84.6%), and even H100 re-evaluation with published Falcon-7B and GPT-J-6B heads — indicating the problem is structural, not just a capacity issue.

The root cause lies in the European Patent Office's Article 84 EPC, which requires claims to be "clear and concise." This pushes human drafters toward low-perplexity, low-burstiness language — exactly the statistical signature that LLMs produce. In response, Banerjee developed a seven-feature linguistic-complexity logistic regression model that achieves 74.0% accuracy at 28.1% FPR — a 13 percentage-point gain over a perplexity-only baseline at a comparable operating point, without using likelihood at inference and within the same hardware budget. The findings have immediate implications for patent offices and law firms: current AI detection tools are unreliable for legal documents, and regulators like the EPO (which now holds applicants strictly responsible for LLM-assisted content under Article 83 and Rule 42) risk misclassifying legitimate human work as AI-generated.