AI Can Now Draft Your Patent and Grade Its Own Work

Filing a patent normally means paying a specialist thousands of dollars to write a dense legal document describing your invention. Researchers at Mitsubishi Electric built a testbed they call "Vibe Patenting," where one AI agent drafts the patent and a second AI acts as judge — scoring the draft and handing back structured notes for revision. The draft then gets rewritten, re-scored, and rewritten again.

The results are striking. When the AI got judge feedback, quality kept climbing with each round. When it just revised on its own, improvements quickly flatlined. Even better, the feedback loop let a cheap, low-power AI model nearly match a far more expensive high-reasoning model — a real cost saving. Stronger models helped, and specialized patent-drafting workflows helped even more.

Here's the catch. The team asked a professional patent attorney to independently grade the same drafts. The AI judge and the human agreed meaningfully — but only on some measures, not others, and the AI showed consistent biases, scoring things differently than a human expert would, round after round. In plain terms: the AI grader isn't wrong often, but it is wrong in predictable directions, which is exactly the kind of error that's hard to notice.

So what does this mean for you? If you're an inventor, a small company, or a startup founder, patent drafting is one of the biggest upfront costs of protecting an idea. This research suggests AI could soon produce solid first drafts fast and cheap, with a human attorney polishing rather than starting from scratch. The risk is the opposite too: a flood of cheap, AI-graded patents that look fine to a machine but wouldn't survive a real legal challenge.