AI detectors flag human patent writing as machine-generated at high rates, study finds

Benchmarking three zero-shot AI detectors on 500 granted European Patent Office telecommunications patents against 500 LLM-generated counterparts, researchers found all detectors exceeded a 60 percent false-positive rate at the claim level. Binoculars produced a 78.3 percent false-positive rate, Fast-DetectGPT 61.3 percent, and DetectGPT 80.5 percent. The failure persisted across model regeneration, LoRA-adapted scoring heads, cross-IPC replication on three additional patent classes, and re-evaluation on H100 hardware with published Falcon-7B and GPT-J-6B heads, suggesting the issue is structural rather than a capacity limitation. A seven-feature linguistic-complexity logistic regression achieved 74.0 percent accuracy at a 28.1 percent false-positive rate, a 13 percentage-point improvement over a perplexity-only baseline at a comparable operating point, without using likelihood at inference and within the same consumer-GPU hardware budget.