There are a number of hurdles facing an applicant for a patent on an artificial intelligence invention. These hurdles include: (1) the prohibition against patenting pure software;¹ (2) the prohibition against patenting data;² (3) ownership and privacy issues regarding data used in an artificial intelligence application;³ (4) given the rather opaque nature of the precise data relationships within the artificial intelligence platform, satisfying the enablement⁴ and definiteness⁵ requirements of the Patent Statute may be problematic; and (5) for inventions invented using an artificial intelligence platform, it may be that no human is the actual inventor.⁶

From an enforcement standpoint, apart from validity challenges to an issued patent based on one or more of the foregoing grounds, an artificial intelligence patentee faces the difficult task of determining whether a system actually employs the invention as claimed. By way of example, unless an infringer has published the details of its artificial intelligence system, the owner of the afore-referenced ’700 Patent will likely be unable to prove that a putative infringer is using the categorizer DBN as claimed. This would not give the patentee the requisite good faith belief in the bona fides of a complaint for patent infringement.⁷

With regard to the aerospace antenna discussed above, assuming a claim were drafted covering the use of artificial intelligence to design the antenna, one looking at a potentially infringing antenna would be unable to tell whether its design was generated through the use of artificial intelligence. Of course, a claim could be drafted directed to the antenna itself without reference to the manner in which its design were generated. As noted above, however, an inventor must be an “individual.” The artificial intelligence platform, which is the true instrumentality responsible for generating the antenna’s design, cannot be an inventor. The person that instructs the platform to search for an optimal antenna design is no more an inventor than one who merely identifies a problem for others to solve.⁸

Other practical considerations may affect a patentee’s decision regarding whether to enforce an artificial intelligence patent, such as the potential ability of an infringer to design around the claimed invention. Looking again at the ’700 Patent, an infringer could conceivably replace the claimed DBN with a convolutional neural network (CNN), thereby avoiding infringement. More generally, with the advances in overall computer technology, the likelihood that a patented invention will remain state of the art for the entire 20-year patent term is remote. Thus, a patent’s commercial value from an enforcement standpoint will typically expire long before the patent itself.⁹

These enforcement issues may very well explain the metrics regarding what companies are actually doing with their patents on artificial intelligence inventions. Approximately 900 families of artificial intelligence patents have been the subject of litigation in the United States.¹⁰ Four percent of those families were owned by four companies, three¹¹ of which are, or have a relationship with, non-practicing entities.¹² At the same time, the two companies applying for the most patent families directed to artificial intelligence¹³ have applied for a collective total of approximately 14,200 patent families.¹⁴ Those same companies have litigated a combined total of 14 families, or 0.1% of the patent families for which they have collectively applied.¹⁵ The two companies having applied for the third and fourth most amounts of artificial intelligence patent families,¹⁶ despite having collectively filed patent applications for more than 10,000 artificial intelligence patent families¹⁷ have never filed a patent infringement suit involving any of those families.¹⁸ Given that the average cost for filing a software patent application is $16,000,¹⁹ a conservative estimate²⁰ of the total amount that these top four companies spent on prosecuting their collective artificial intelligence patents is $387.2 million.

Irrespective of the ultimate reasoning, patent application filings have not translated into assets amenable to enforcement (and, thus, commercialization) in any sort of meaningful fashion. Whether correlative or causative, these metrics indicate that there will likely not be a meaningful return on the investment companies have made in prosecuting their respective artificial intelligence patent applications. When this metric is considered along with the afore-discussed issues attendant to securing patent protection for artificial intelligence inventions in the first place, an alternate approach to securing patent protection for such inventions warrants consideration.

Considering again the three categories of artificial intelligence inventions, an improved artificial intelligence platform, new applications for artificial intelligence, and inventions made via an artificial intelligence platform, at least the first two categories appear amenable to trade secret protection. Beginning with the first category, in the event that a novel artificial intelligence platform is invented, such as a platform with an improved neural network or an algorithm and associated software for powering or combining such networks, that platform and the associated neural network and software would be entitled to trade secret protection as long as the platform is maintained as a secret. In a fashion analogous to computer source code, the internal workings of an improved artificial intelligence platform should not, and in the vast majority of instances, would not be disclosed to the public.

The same is true for a unique data set used to generate output from an artificial intelligence platform. Turning again to patient data that would be subject to certain privacy restrictions, the treatment of such data as a trade secret would likely comport with the privacy requirements governing such data. The manner in which such data is mined and manipulated for optimal use of an artificial intelligence platform can also be maintained as a trade secret. This interface between the artificial intelligence platform, the data to be input to such platform, and any particularized handling of such data can provide a broad swath of trade secret protection.

Turning now to the second category of artificial intelligence inventions—novel applications for artificial intelligence—it is again easy to envision an environment in which artificial intelligence is used in a secret fashion to enhance information processing. A company may employ artificial intelligence to evaluate different combinations of battery materials to arrive at combinations representing an improvement over existing lithium-ion batteries. That company can maintain as secret the manner in which it is investigating novel battery material combinations.

Considering the afore-depicted artificial intelligence eyewear application, the commercialization of such a product will inform the public that the subject eyewear uses artificial intelligence, and such use cannot be a trade secret. On the other hand, the particular artificial intelligence platform, e.g., the type of model powering the platform, may well be protectable as a trade secret. Such is also the case for a particular database analyzed by the platform and the manipulation thereof.

Finally, the third category of artificial intelligence inventions—those made via an artificial intelligence platform—may be appropriate for trade secret protection. The lynchpin to this issue, at least under the DTSA, is whether the public can discern the invention from the commercial embodiment thereof. In the case of the afore-described battery material combination, it is likely that one could reveal the combination through reverse engineering. As such, the combination is “readily ascertainable,” and therefore cannot be a trade secret.²¹ Such reverse engineering does not constitute an improper means for acquiring the subject information under the DTSA.²²

If, however, the artificial intelligence platform is used to determine an optimal method of processing these same materials and that method may not be amenable to invention through reverse engineering, the method can be maintained as a trade secret. Similarly, if the Coca-Cola company could somehow employ artificial intelligence to improve upon its beverage formula, empirically, there is evidence that it is entirely possible to maintain the secrecy of that formula, even though it is used to manufacture a product sold around the world.

In light of this reverse engineering issue, the propriety of seeking patent, as opposed to trade secret, protection for this third category of artificial intelligence invention will likely need to be evaluated on a case-by-case basis. In general, however, as suggested above, the issue of proper inventorship may be a bar to patentability where an invention has been entirely invented by a machine. Thus, for this third category, the choice may very well be trade secret protection or nothing. In any event, trade secret protection does not engender incremental costs analogous to those associated with patent prosecution. Therefore, without incurring incremental costs, a company may initially treat this type of artificial intelligence invention as a trade secret, leaving to a later time the determination of whether to pursue a claim in the event of the misappropriation of such information.

For each of these three categories of artificial intelligence invention, the owner must, pursuant to the DTSA, take “reasonable measures” to maintain the invention as a secret.²³ The bar for such maintenance is likely not particularly high, with confidentiality agreements and password protected entry points in most cases being sufficient to satisfy this requirement.²⁴ As an aside, this bar would allow a trade secret owner to protect trade secret status during evaluation by a third party by simply putting in place an appropriate non-disclosure and use limitation agreement.

Those engaged in the effort of artificial intelligence research and development may be lured toward patent protection by well-publicized large damage awards for patent infringement. However, in the more than 700 patent infringement cases between 2000 and 2014 in which compensatory damages were awarded, the median award was $372,000.²⁵

However, trade secret misappropriation can also yield large awards, with its combination of actual loss, unjust enrichment, and exemplary damages. By way of example, in a state court case in Texas, a jury awarded a trade secret counter-claimant more than $700 million.²⁶ Perhaps surprisingly, the median damage award in federal court for trade secret misappropriation during the 2001-2015 period was $443,453, which is somewhat higher than the median for patent infringement damages.²⁷

Thus, the owner of an artificial intelligence invention does not sacrifice the potential to recoup damages in the event the invention is maintained as a trade secret, as opposed to the subject of a patent application. And when the potential for damages is coupled with the availability of injunctive relief, a trade secret owner has available a combination of remedies every bit as meaningful as those available to a patentee. It is also worth reiterating that these remedies come without the challenges attendant to seeking patent protection for computer-based inventions. Finally, trade secret status brings with it the threat of criminal sanctions for misappropriation, which can provide a powerful deterrent, thereby decreasing the likelihood of such misappropriation.

¹   See, e.g., Microsoft Corp. v. AT&T Corp., 550 U.S. 437, 449 (2007).

²  Digitech Image Techs. v. Electronics for Imaging, 758 F.3d 1344, 1350 (Fed. Cir. 2014).

³   For example, the Health Insurance Portability and Accountability Act of 1996 (“HIPAA”) and accompanying regulations limit the use that regulated entities can make of patient data. See, e.g., 45 C.F.R. § 164.502(a)(5) (listing prohibited uses of patient data). Where artificial intelligence is used, for example, to improve diagnostic predictability, the use of patient data may well have restrictions on its use pursuant to HIPAA.

⁴   35 U.S.C. § 112(a) (“A written description of the invention . . . in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains . . . to make and use the same”).

⁵  35 U.S.C. § 112(b) (“The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention”).

⁶   35 U.S.C. § 100(f) (“The term ‘inventor’ means the individual or, if a joint invention, the individuals collectively who invented or invented the subject matter of the invention” (emphasis added)).

⁷   Fed. R. Civ. P. 11(b)(2) (“By presenting to the court a pleading, written motion, or other paper . . . an attorney or unrepresented party certifies that to the best of the person’s knowledge, information, and belief, formed after an inquiry reasonable under the circumstances . . . the claims, defenses, and other legal contentions are warranted by existing law or by a nonfrivolous argument for extending, modifying, or reversing existing law or for establishing new law”).

⁸  As noted above, an inventor invents an “invention.” 35 U.S.C. § 100(f). The Patent Statute provides that a patent may be obtained for “any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof.” 35 U.S.C. § 101. An identified “problem” would not fall within any of these categories.

⁹  Irrespective of any remaining commercial value of a patent, a patentee will be required to pay a total of $12,600 in fees to maintain a patent for its entire 20 year term. See 37 C.F.R. §§ 1.20(e)-(g).

¹⁰  World Intellectual Property Organization supra note 38, at 110.

¹¹  American Vehicular Service, Automotive Technologies Int’l, and Excel Innovations.

¹²  World Intellectual Property Organization supra note 38, at 110.

¹³   IBM and Microsoft.

¹⁴   World Intellectual Property Organization supra note 38, at 60.

¹⁵   See id. at 110.

¹⁶   Toshiba and Samsung.

¹⁷   World Intellectual Property Organization supra note 38, at 110.

¹⁸   Id. at 60.

¹⁹  Gene Quinn, The Cost of Obtaining a Patent in the US, IP Watchdog (Apr. 4, 2015), https://www.ipwatchdog.com/2015/04/04/the-cost-of-obtaining-a-patent-in-the-us/id=56485/.

²⁰   This metric assumes one patent application per family.

²¹   18 U.S.C. § 1839(3)(B). In contrast, the California UTSA does not include this “readily ascertainable” language in the definition of “trade secret,” and the theoretical ability to reverse engineer does not affect trade secret status. ABBA Rubber Co. v. Seaquist, 235 Cal. App. 3d 1, 7 (1991) (“under California law, information can be a trade secret even though it is readily ascertainable, so long as it has not yet been ascertained by others in the industry”).

²²  18 U.S.C. § 1839(6)(B) (excluding reverse engineering from the definition of “improper means”).

²³  18 U.S.C. § 1839(3)(A).

²⁴  Protection Technologies, Inc. v. Ribler, No. 17-cv-00144, Order (D. Nev. March 8, 2017) at 4. Conversely, where a putative trade secret owner discloses the subject information without restriction to a third party, the owner has not taken the requisite reasonable measures to maintain the secrecy of that information. M.C. Dean, Inc. v. City of Miami Beach, Florida, No. 16-civ-21731, Order (S.D. Fla. Aug. 8, 2018) at 12-13.

²⁵  Brian Howard, The Truth About Patent Damage Awards, Law360 (Oct. 16, 2014, 10:21 AM), https://www.law360.com/articles/557734/the-truth-about-patent-damage-awards.

²⁶  Title Source Inc. v. HouseCanary Inc., No. 2016CI06300 (Tex. 73^rd Dist. March 14, 2018).

²⁷   John E. Elmore, A Quantitative Analysis of Damages in Trade Secrets Litigation, Insights (Spring 2016) 79, 91.