Rookie mistake: thinking the specification is the important thing.
The claims are the only thing that matters. I'm not saying it's impossible to generate those with AI, but claim language would be a much deeper project than anything you can knock off while you're bored. And "what does this claim actually cover? What did the PTO say recently about means-plus-function claims?" are not questions readily amenable to answering in an automated fashion.
At Google, I had the task of examining our 20,000 patent portfolio for a certain purpose which I won't mention. We had an in house tool which did a pretty good job of categorizing them, so I didn't have to read each one. There are a lot of legal technology firms already applying modern technology to the legal system.
As for searching, we sent a certain task to around 10 leading patent search firms, and compared their answers. The winner would cost around $5,000, if memory serves. If you're imagining it's worth millions of dollars: it isn't.
The patent system seems like it is a ripe subject for the kind of automation and disruption that we have seen lately in the legal sphere. Where you can fight your parking ticket and work within the law to defend your rights based on advice from an AI or similar system.
Has anyone done that for patents? You could have a section for proper writing of the patent, an accurate search function that runs in parallel and searches existing patents, an easy submission system that integrates with the world's patent database.
There have been quite a few patent drafting softwares over the years. They seem to do a reasonable job guiding someone through the process of writing the patent application. But search and submission are not automated.
Here are some softwares that are still currently on the market:
AI-based patent search seems to look at citations (which is somewhat worthless as I'll look at citations anyway) and have some measure of text-based similarity. In my experience it does a poor job when the terminology varies. There was no substitute for making a search string containing a lot of synonyms.
I think AI-based patent search could see a lot of improvement in the future if it simply started looking at the drawings. As a "mechanical" examiner, many of the applications I worked on were most effectively searched by flipping through patent drawings. Sometimes the prior art I'd use to reject an application would have drawings that are remarkably similar to those of the application I'm examining. But other times (particularly involving flow or electrical circuits), the drawing is equivalent in some sense but arranged differently. A more advanced AI/ML approach is needed for those.
Thank you for the great response. Yes, I can definitely see how mechanical or physics based patents would be beyond the AI's remit at the moment. As someone who is considering going for a software patent sometime in the near future I was pretty disheartened by the patent process - both submitting the patent (the expense can be up to $500k I was told) and also defending the patent from larger companies seemed impossible. Perhaps an AI can start with easily understandable patents like software, or primarily word- or keyword based patents, and go from there to an ML approach for physics and such.
I also have been discussing this idea with others for some time and wonder how it can be exploited to break or improve the patent system. While some now try to patent things build by an AI, I think the other would be interesting: creating random claims using patent language and timestamp those. Combine this with some kind of similarity search to link the generated stuff as prior knowledge to any newly created patent application. Reinforce the generator by its ability to 'predict' patents. Use the argument that patent offices use to fend of ai patents against them: if a computer can generate the creative height cannot be sufficient...
I understand the impulse however the cost for submitting a patent is around the half-million US dollar mark (or so i have been told by patent lawyers) so I think improving the process and helping genuine innovations get some protection would be better
Maybe through prior art. If AI can reliably produce patents, it can also reliably produce prior art by the trillions, all of which can invalidate new patents.
The claims are the only thing that matters. I'm not saying it's impossible to generate those with AI, but claim language would be a much deeper project than anything you can knock off while you're bored. And "what does this claim actually cover? What did the PTO say recently about means-plus-function claims?" are not questions readily amenable to answering in an automated fashion.
At Google, I had the task of examining our 20,000 patent portfolio for a certain purpose which I won't mention. We had an in house tool which did a pretty good job of categorizing them, so I didn't have to read each one. There are a lot of legal technology firms already applying modern technology to the legal system.
As for searching, we sent a certain task to around 10 leading patent search firms, and compared their answers. The winner would cost around $5,000, if memory serves. If you're imagining it's worth millions of dollars: it isn't.