Procedurally Generated Prior Art: A Closer Look
When defending a patent-infringement case, attorneys primarily focus on two avenues of defense. First, they argue that the accused product does not infringe the patent's claims. For example, the claim may require a hybrid car that has a "continuously variable transmission" (a feature available on more cars every year), but the accused hybrid car has a manual transmission (a truly fun feature that is, sadly, slowly disappearing). Second, attorneys argue that the patent is invalid because someone disclosed it to the world before the alleged inventor filed for their patent. Staying with the previous example, someone already thought to combine a continuously variable transmission with a hybrid car, so you cannot patent it.
To support an invalidity defense, attorneys go hunting for "prior art" that discloses the alleged invention. Attorneys themselves, as well as outside vendors, search the world over for some public disclosure of the alleged invention before the patent at issue was filed. It is a gold rush. If attorneys can find something, then the patent may be invalidated, there is nothing left to infringe, and the plaintiff's claim becomes no longer viable.
The running example above is a combination of two elements (a hybrid car and a continuously variable transmission), and the U.S. Supreme Court has recognized that many inventions are combinations of previous technological advancements. KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007) ("Inventions usually rely upon building blocks long since uncovered, and claimed discoveries almost necessarily will be combinations of what, in some sense, is already known"). What if software and natural language processing were used to make those combinations before humans were able to make the connection?
That is the goal of the longstanding All Prior Art project. It uses "procedural generation" (i.e., using a computer to generate some data) to create short disclosures of "inventions." These "inventions" are just combinations of technical terms the software found in existing patents or patent publications. Using natural language processing, the software generates public disclosures such as the following:
A wearable electric device includes a main body with a circuit module inside and at least a detachable battery strap with a battery module inside, and the main body and the detachable battery strap are detachably fastened together. The test device includes an addressable memory. Responsive to one of the plurality of processes attempting to access a late binding object by its identifier, a determination is made as to which late binding object is associated with the process. The composition may be applied to soil to control a population of a deleterious organism. Each of the strips is radially offset from one another. In the sealing step, long side edges of the battery case are crimped by a forming surface having a rounded cross section, and arc-shaped edges connecting both long side edges are crimped by a flat forming surface.
1461187725-f00e0aa8-68b8-4055-92c0-b74346ad03c3 (this seemingly nonsensical title identifies when the text was published to ensure it predates a target patent). As of this writing, the All Prior Art project has published more than 570 million such disclosures.
Questions and Considerations
The text is a jumble of combined elements, but could it serve as prior art in a patent litigation? It meets all requirements to serve as such. The Patent Act states in part that "[a] person shall be entitled to a patent unless  the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention[.]" 35 U.S.C. § 102(a)(1) (post-AIA). There is no reason the procedurally generated "inventions" from All Prior Art and All the Claims would not count as descriptions "available to the public" at a certain point in time. While there is a requirement that humans (not computers) invent the subject matter claimed in patents, nothing in the Patent Act requires that prior art disclosures originate with humans. In effect, All Prior Art and All the Claims are using algorithms to strip away inventive subject matter and designating that subject matter to the public.
Both projects have been in operation since 2016, but neither has been cited in any federal opinions relating to prior art. It remains to be seen if these projects will have any practical effect on patent litigation in the years to come.
Since 2016, natural language processing has advanced. A similar project that leverages the GPT-3 engine could likely produce even better results for procedurally generated prior art disclosures.