The original title is "The Perplexity Trap: When Patent Law Makes Human Writing Look Like AI"

Short summary

Researchers benchmark three open-source AI-text detectors on 500 granted EPO patents versus 500 LLM-generated counterparts, finding false-positive rates exceeding 60% at claim level because patent drafting style naturally occupies the same low-perplexity, low-burstiness manifold as LLM output. A seven-feature linguistic-complexity logistic regression achieves 74% accuracy at 28.1% FPR, outperforming perplexity-only baselines without requiring likelihood scores at inference. The findings suggest AI detection in patent contexts is structurally difficult, not merely a hardware or model-capacity problem.

• •AI text detectors show 60-80% false-positive rates on real EPO patent claims
• •Patent drafting style inherently mimics LLM text characteristics (low perplexity, low burstiness)
• •A linguistic-complexity logistic regression improves accuracy to 74% at 28.1% FPR without likelihood-based scoring

Generated with AI, which can make mistakes.