I love this project that has the AI tech develop a detailed 20-30 page specification based upon a few claims and short description. https://information.patentepi.org/issue-2-2021/natural-language-generation-and-patents-be-ready.html
I expect that most folks won’t be satisfied, but in many cases it might be sufficient. And, if the technology is sufficiently mainstream, it suggests that the patent document actually needs less disclosure.
Artificial intelligence is to real intelligence what artificial leather is to real leather. I’d rather have a fake Gucci handbag than an AI-generated patent.
I’ve played around with GPT-3 for this purpose as well. The issue is consistency (using the same terms for the same functions) and accurate reference numbers. The effort to rewrite / double check the AI-geenrated specification exceeds the effort to properly write the specification in the first place. Training a dataset on patent specifications is unlikely to resolve these problems in the near future. You can experiment yourself here, try pasting claim limitations as the seed text:
link to deepai.org
Inputting the ‘famed’ Benson Claim 8, the AI spits back:
There is nothing new in this text.
There is nothing new in the data.
There is nothing new in the data.
It is a great text and the way the data is structured makes no sense to me (and then some further nonsense).
No thanks.
Cool idea. All you have to do is be willing to risk:
1. This company or their proxy(ies) quickly filing their own apps in multiple countries — especially if the invention is pioneering . . .
2. That the work they provide constitutes a public disclosure . . .
Other than those, sure thing, sign me up!
I would have thought, almost by definition, that if a machine learning model can produce a spec that is sufficient to satisfy 35 USC 112 (and corresponding disclosure requirements in other jurisdictions) from just claims and a short description, then the alleged invention is most likely obvious or anticipated. Unless the short description itself is sufficient, in which case the AI is superfluous.
AI is not magic. The only sources of content for the generated spec are the input claims and short description, and the training data (i.e. a subset of the prior art).
One big
for you,
“I would have thought, almost by definition,”
That is entirely unsurprising, Mark, GIVEN that your view on the lack of creativity of AI devices (vis a vis the DABUS case) is well known.
That being said, being well known – and STILL being incorrect, changes nothing.
IF you ever get to the point of appreciating that there may be an advance** that involves a non-human level of invention – that is, an invention is present to which no human may legally claim to BE the inventor, THEN you may be a bit more able to see why your own self-defined limitations are just that: self-defined limitations.
“ AI is not magic”
No one is suggesting this strawman.
“ sources of content” (with the faulty logical premise of “unless the short description itself is sufficient, in which case the AI is superfluous”)
Source – only — is not an issue – and never has been. Here, this is a clear logical flaw in your statements, given as you assume your own conclusion that AI cannot provide ANY inventive step from inputs to output. It is only by YOUR (flawed) definition that AI itself must be superfluous, as you have been more than clear that you think that AI cannot contribute anything inventive.
Now, I do grant that on this thread, “AI as inventive” is not being advanced. But I rebut your comment not so much for the assertion related to Qatent, but more so to the lack of vision on the larger topic that will necessarily blind you to thinking through even this scenario.
TFW a company in a foreign jurisdiction generates several patent applications using an AI tool, files PCTs, and when it comes to nationalizing in the US asks you to file them in the US, and you spend .6 hours staring at your monitor thinking to yourself “wtf am I going to do with this?”
That might make an interesting ethics question.
Once upon a time, this blog had its own dedicated (and active) ethics page.
Why do you love it, Dennis?
Three questions:
Who owns the work product?
Can an attorney submit this “work of another” when there is no oversight for the work itself?
Are the inputs to the process “not held in confidence prior to an actual filing?”
Comments are closed.