Tuesday, February 23, 2021

SUI GENERIS PATENT REGIME FOR AI RELATED INVENTIONS

 

ABSTRACT

 

AI technology, one of the primary contributors of 4th Industrial Revolution, has witnessed an unprecedented growth and extraordinary progress in last decade. The frenetic pace with which the technology has evolved left different sections of society with mixed emotions. As the scientific community is rejoicing at marvels they have gifted society with, the ethics community is implicating traditional qualms over moral intuition & absolutism, while the law makers are both perplexed and exhilarated understanding legal complications accompanying it.

Amongst array of legal challenges, rubric of patent law protecting AI solutions is hugely confounding patent law makers. Some of intriguing questions pertain to inventorship/ownership rights, patenting process, disclosure requirements of inventions enabled by or borne out of AI machines. Present paper explores above legal confines for both human and/or non-human contributor, and categorically attempts to address the hotly contested issue of patent law adjustments via adaptation through a sui generis framework.


 

TABLE OF CONTENTS

 

 

 

   S. NO.

PARTICULARS

PAGE

 

1.

INTRODUCTION

3

 

2.

INVENTORSHIP AND OWNERSHIP PROSPECTS FOR AI RELATED INVENTIONS

 

 

 

 

 

A.    NATURE & DYNAMICS OF AI MACHINES

 

B.    WHAT MAKES PATENTING FOR AEIs DIFFERENT FROM CIIs

 

C.    DETERMINIING AND DEFINING AEIs AND ABIs

 

D.    INVENTORSHIP AND OWNERSHIP ISSUES OF AEIs & ABIs

 

4-6

 

 

6-7

 

 

7-9

 

 

10-11

 

 

 

3.

SUI GENERIS PATENTING FRAMEWORK

 

 

 

 

 

 

A.    LEGAL REQUIREMENTS & DISCLOSURE JUSTIFICATION

 

B.    WHAT CONSTITUTES SUFFICIENT DISCLOSURE FOR AEIs

 

C.    DISCLOSING FORESSEABLE ASPECTS OF AEIs

 

D.    PROCEDURAL FLOW FOR  ABIs COMPRISING UNFORESEEABLE ASPECTS

 

E.    DISCLOSURE REQUIREMENTS FOR ABIs

12-13

 

 

14-17

 

 

18-19

 

 

19-22

 

 

 

22-24

 

 

 

4.

CONCLUSION

24

 

 

INTRODUCTION

Technology of intelligent and autonomous machines, popularly captured in term Artificial Intelligence (AI), is a new buzzword that has power of changing world reality of today. Coined by Professor John McCarthy at a Dartmouth conference in 1956, AI is emerging as a key driver of the ‘fourth industrial revolution’[1]. Understandably, the term describes the capacity of a computer to perform tasks commonly associated with human beings[2]. As the machines get empowered in their ability to autonomously retrieve relevant information, identify subtle patterns and relationships between various data segments and make intelligent predictions or recommendations, relevant stakeholders are becoming increasingly skeptical regarding motley of legal challenges concomitant with advancing technology.

Amidst anticipated disruption of numerous legal frameworks, impact on patent law appears immensely pervasive and significant than any of previous technological changes[3].  Clearly, the swelling wave of innovations enabled by or borne out of AI machines is on a collision course with basic tenets of patent laws, primary debate being centered on a proposition whether or not to accord legal personhood status upon electronic entities. From here, ensues a series of contentious topics such as patentability standards, inventorship or ownership issues, accountability or liability issues, risk, infringement etc. as AI systems can neither own any property nor can partake in any employment relationship[4]. To study the impact of these issues, WIPO recently invited comments from interested communities to various pertinent questions[5] related to above listed topics.

Detailed analysis of the suggestions posted in response to discussion points by individuals and organizations from different member countries[6] posits a blatant truth that the existing patent regime is not well suited to competently manage AI related inventions. While many have advocated absolute negation of the concept of machine inventorship for such patent applications are likely to result in unnecessary deviation from the basic rationale underpinning the patent regime; many contended that these works shall fall in public domain[7], or conveniently clubbed with existing software patent regime[8]. Other experts have contested continuation of inapplicable laws and persuasively established dire need for a new statutory framework - a sui generis system specifically tailored for addressing doctrinal challenges related to AI technology.

The objective of present paper is to critically analyze different forms of inventions enabled by or borne out of a human or non-human agent or a combination thereof, and develop a framework as a probable legal solution to secure their patent interests. Part I identifies and classifies human/non-human related AI innovations into two broad categories: a) AI Enabled Inventions (AEIs) that either embody an advance in field of AI or apply AI to other field, and b) AI Borne Inventions (ABIs) that are produced by AI. Part II proposes a sui generis framework predicated upon fundamental justifications for patent rights to administer and protect AEIs and ABIs with minimal legislative overhaul.

 

PART I: INVENTORSHIP AND OWNERSHIP PROSPECTS FOR AI RELATED INVENTIONS

 

A.    NATURE & DYNAMICS OF AI MACHINES

Intelligent machines of today do not exclusively rely on linear set of programming instructions or number-crunching but also “thinking” and capacity to reason for itself[9]. Recent technologies of neural networks, genetic programming or evolutionary engineering are some example of creative and self-replicating techniques for independent problem-solving. In absence of any uniform definition, AI can be understood as completely autonomous machines with cognitive features capable of learning from input data, experience and interaction, surpassing degree of intelligence once held to be characteristic exclusive of human mind[10]. These are highly distinguished from traditional human guided computer hardware programmed to perform a particular task[11].

In present context, no General human like intelligent AI machines[12] fully capable of independent judgment, reasoning, agency, creativity or decision making without any human intervention, has been objectively or evidently known. In turn, most popularly known AI machines belong to genre of ‘Narrow AI’ that provides solutions to a limited set of narrowly defined problems arrived at by varying degree of human input/interaction[13]. Examples include advancements made in fields like autonomous vehicles, predictive analytics, speech recognition, computer vision or image recognition, customer service bots, spam filters, recommendation systems and so on. Plainly, AI machines exhibiting such discernment as that of a human agent is yet an unfulfilled and unrealized technology[14]. Although computers are not yet capable of completely autonomous invention, it could still be on the horizon as AI undergoes fast-paced innovation enabled by increased availability of improved computational resources, high capacity storage, advances in Big Data and advent of special hardware with specialized chips capable of supplying enormous computational power[15].  

In face of rapid technological changes and accelerated innovation activity, focus on patenting trends and their societal effect becomes paramount. A convoluted gap between the racing AI technology and slow-chasing legal stature already exists and has grown big enough to necessitate radical changes in the patent system. Optimistically speaking, knowing the challenges of tomorrow very clearly today, is rather a generous relaxation for law makers to raise legal guards and adopt a well-defined and strictly enforceable framework to safeguard interests of next in order AI machines.

 

B.    WHAT MAKES PATENTING FOR AEIs DIFFERENT FROM CIIs

Necessity of a new patenting regime for AI related inventions appears no more an academic exercise, but an immediate, fundamental problem loudly banging on patent doors. Today, account needs to be taken of existing patent regime’s capacity in reasonably handling changed circumstances of boundless advancement in machine intelligence akin to existing computer implemented inventions (CIIs). Well, amongst all advances seen in realm of computer sciences, none has been so far capable of demonstrating intelligence that can challenge, limit or question extent of human involvement. AI machines, in wide contrast to computer programs have remarkable quality of extracting patterns, correlations within dataset to conclude a meaningful output, with or without any human supervision[16]. One famous example is Stephen Thaler’s ‘Creativity Machine’, which like a human brain, is capable of generating novel patterns of information rather than simply associating patterns, and it is capable of adapting to new scenarios without additional human input[17]. 

For computer-implemented inventions (CIIs), even the specialized computer hardware ‘configured for’ yielding novel and inventive claims simply implements programmer’s algorithmic instructions. A human agent has always been a moderator, and machine never assumed to    approximate mental capabilities of human as it is guided at each step to obtain a static and specific output defined by its human operator. However, groundbreaking innovations achieved using AI techniques have clearly established that machine can be no more seen as a tool subservient to human commands and following digital orders. If fed with suitable inputs, they can learn how to perform tasks, prove mathematical algorithms and find solutions to a task independent of direct human supervision. Further, machines have even surpass human blind spot in achieving increased productivity and efficiency at decreased cost of innovation[18], leading to increased complexity in dealing with patentability issues of inventions enabled by borne out of AI machines. Another critical aspect that marks a striking contrast between AEIs and CIIs is the nature of claims drawing the boundaries of these inventions-while static for CIIs, claim scope is dynamically varying for AEIs. Hence, it is convenient to decide terms of grant well in advance for CIIs; but for AEIs conclusively finalizing boundaries of invention is unsettling as there may be outputs which are foreseeable, but cannot be promised for reasons of uncertainty.

This discussion is necessary as the inventions enabled by or borne out of AI machines cannot be contained within legal brackets of conventional CIIs where humans are solely awarded as the true and first inventor, and applications filed with machine as inventors are outrightly rejected (e.g. Dabus)[19]. Speedy adoption of these technologies have the potential to impact patent system on a scale that it is not currently equipped to accommodate. A rethinking of traditional patent tools is definitely required[20]. Unless cured of its current impotence, patent law may slide towards a detrimental conflation of otherwise distinctive “human-dominant-machine” and “machine-dominant-human” continuum, thus failing advancement of purpose underlying innovation incentivization.

 

C.    DETERMINIING AND DEFINING AEIs AND ABIs

This section draws a clear distinction between AI enabled inventions (AEIs) and AI borne inventions (ABIs), aiming to provide perspective on why a completely independent patenting solution is necessitated for different types of learning for AI machines. Notably, machines have learned to recursively self-replicate beyond human comprehension, by way of fetching expansive volumes of datasets, performing algorithmic processes on its own, and even outputting smarter and more utilitarian results than previously known models[21]. Machines can demonstrate intelligence to the extent of improvising, autonomously, final output up to varying degree of sophistication.

Most AI machines are trained by labeling and categorization of underlying data, commonly known as ‘supervised learning’[22]. Such works are ‘enabled’ by AI techniques, which are employed in setting a desired output to a particular problem and then fitting a supervised learning component into a bigger system. Primarily, the steps of selecting features to represent data, transforming data, choosing an appropriate algorithm , tuning of parameters, and finally assessing quality of resulting model via a feedback mechanism is not completely deprived of human dominion as virtually all steps contain some modicum of human activity or creativity[23].

Human intervention is manifested in various forms as they invest higher-order cognitive skills such as reasoning, comprehension, meta-cognition, or contextual perception of abstract concepts in selection and curation of input data, configuration of training model, defining (technical) problem statement or improving target performance metrics[24]. Results are then examined by domain experts or practitioners to obtain desired behaviors qualifiable as commercially valuable technical output. Having fruitfully contributed to the inventive concept, not insignificant in quality when measured against the dimension of full invention, the human agent meaningfully proves playing of a measurable role as a ‘co-contributor’ or ‘joint inventor’[25] of derivable AI enabled invention (AEI).  

On the contrary, in an unsupervised learning mode, the machine learns patterns within input and does not require any human feedback or labeling for discovering structure of data or detecting outliers. Theoretically, an unsupervised system can achieve “artificial general intelligence”[26].  Here, the machine learns the way human learns-‘on its own’. In the process of uncovering patterns, the machine may exhibit inventive skills in performing exploratory analysis or dimensionality reduction in given data. Evidently, human has a very limited role in the inventive play of generating these better trained models. So, this output remains entirely ‘machine borne’, and final product discretely an AI borne invention (ABI).

 Certainly, such intelligent machines deserves due recognition as they significantly expanded the range of things that a human can discover. It will be against the moral fabric of patent system to acknowledge non-contributing human agent as a joint inventor, whose role has merely been managerial, administrative or financial. Consequentially, for ABIs, human contribution will always remain lowly visible as most of computing effort along with intellectual contribution is passed onto AI machine.

Concluding from above, it will be unfair to over-reward machines (for AEIs) or humans (for ABIs), for conceptions they never contributed to substantially, when examined in isolation[27]. Also, it will be in contravention to fundamental principle of attribution of inventorship to true inventor, which at least in case of AEIs and ABIs is rarely a product of human or non-human agent alone[28]. Thus, an optimal balance over impersonal realities of inventorship may only be struck by acknowledging de-facto contribution of a sensible combination of human intellectual effort and unsupervised machine’s effectiveness in improving or optimizing system performance relative to some objective function[29]. Undeniably, this rightful acknowledgment of true inventors creates an absurd situation with lurking issues of ownership, accountability, infringement risks, or liability, which are presently taken charge of by human agents all alone.     

 

D.    INVENTORSHIP AND OWNERSHIP ISSUES OF AEIs & ABIs

In light of ratio of human-to-machine contribution to inventive processes progressively shifting in favor of machine, a more rationale and justified conceptual model of patenting AEIs & ABIs is suggested[30].  However, for convenience sake, many have advocated complete eradication of concept of crediting machines as one of the inventors, or alternately clubbing it with CII patenting regime. In addition, other mystifying scenarios are also being considered when such AEIs and ABIs are chosen to be protected under trade secrets or through extensive nondisclosure agreements as a safer and independent course of action. Clearly, this is not a mandate for a well-functioning, robust patent framework, which has earnestly evolved over many years to uphold legitimate interests of inventors within their proper bounds. 

As discussed previously, by virtue of their inherent abilities, AI machines may autonomously replicate. During such replication, some forms of “not-so intelligent” machine-dominant-human systems may even replicate the bias, unfairness and discrimination in data on which they feed. Other limitations include overgeneralizations in pattern detection, reduced accuracy resulting from incomplete data sets, and inherent limitations surrounding the use of existing data to anticipate or predict future novel legal and ethical issues[31]. In these circumstances, intellectual and meaningful domination of human agent over such not-so-intelligent non-human agents becomes inevitable. Tying AI’s action to a human agent, remains as only viable solution to fill this accountability gap because- first, our legal system is built on a fundamental assumption that penalties and remedies can only be levelled against humans; and second, we cannot punish, imprison, or impose fines on AI machines whether it has legal personhood or not[32]. 

 So, how do we intend to address the most controversial inventorship/ownership issue for AEIs & ABIs[33]?  Who’s accountable – developer, manufacturer, operator, owner, user or machine itself. Can the co-inventors be co-owners as well? European Parliament resolution of 16 February 2017 with recommendations to the Commission on Civil Law Rules on Robotics declared that accountability and liability of AI machine per se for damage done to third party certainly makes no sense[34]. So far, it is deliberated and discussed over various forums that determining liability of a non-human agent seems to be an impracticable solution today[35]. Logically, a human agent who conceptualized the machine and had been a co-inventor in its predictable outcomes should be the one bearing responsibility of infringing or damaging acts alone, simply because machines cannot.

Along with benefits of inventorship, risks associated with its ownership unconditionally ensue. How to make human inventors fully accountable for collaborative endeavors without inadvertently impacting them of wrongs they never intended machine to perform? What about acknowledging machines and humans as co-inventors while vesting ownership entirely upon the human agents, simply endorsing high level principles of patenting regime[36]. Apparently, concepts of inventorship and ownership may not be completely entwined; for it seems explicable to adopt a unique approach that is theoretically sound and practically workable in addressing inventorship/ownership issues of AEIs and ABIs. Next section of this article presents a sui-generis framework as this unique approach.

 

PART II. SUI GENERIS PATENTING FRAMEWORK

  First, this part will briefly review the conceptual background of patent laws as applicable to AI related disclosures, and then examine the proposed framework particularly in context of disclosures required for AEIs and ABIs, which have been a cornerstone of patent policy[37]. Some reforms in patenting process and other administrative procedures are suggested for quick adoption and conformance. Once established, the proposed framework after being run through a number of simulations may be further examined for its faults and revised on its workability.  

 

A. LEGAL REQUIREMENTS & DISCLOSURE JUSTIFICATION

Patents are awarded as a quid pro quo for disclosing the invention all across the globe[38]. Disclosure theory centrally focuses on inventor receiving exclusive patent rights in exchange for fully disclosing the invention to society, rather than keeping the invention secretive. Recent America Invents Act reads:

“The specification shall contain a written description of the invention, and of the manner and process of making and using it, in such full, clear, concise, and exact terms as to enable any person skilled in the art to which it pertains, or with which it is most nearly connected, to make and use the same, and shall set forth the best mode contemplated by the inventor or joint inventor of carrying out the invention[39]”.

As explained, detailed submission of ‘useful technical information’ in complete patent specification is quintessential for receiving substantive patent rights as “the test for sufficiency is whether the disclosure of the application relied upon reasonably conveys to those skilled in the art that the inventor had possession of the claimed subject matter as of the filing date.”[40] For AEIs, it is important to verify how humans are involved in different aspects of its conceptualization and constructive reduction to practice[41].  

Lately, some aberrations are observed in making true admissions for AI related patent specifications. Though, 35 U.S.C. Section 103 states: “Patentability shall not be negated by the manner in which the invention was made”, AI machines may be sometimes deployed to invent en masse thousands of alternative patent applications or defensive publications merely by linguistic manipulation[42]. This form of non-inventive claiming can rattle the current patent landscape especially when it comes to identifying true and onerous machine inventing[43].

Only relief comes from the fact that such claim language to serve as a new, inventive and useful disclosure or to play as an analogous prior art may have to be in a form of printed publication, be publicly accessible and most importantly satisfactorily enabling to render a disclosure patent eligible or other following invention invalid[44]. Evidently, these mechanically generated claims will not be adequately supported by an appropriate written detailed description or any other background information, and hence the burden will always remain on the patent office to determine ex post facto whether the disclosure qualifies for an eligible patent grant or if such claims floated as a prior art disclosure is disclosed in sufficient detail to be invalidating. Importantly, seldom are the chances that these machine generated random claims obtained by manipulating phrases overcome obviousness rejections. These factors should remind us that while admitting AI applications, the patent offices must examine them through disclosure and explainability lens to assure that unwieldy thicket of technical information is transformable to a full inventive repository[45].

B.    WHAT CONSTITUTES SUFFICIENT DISCLOSURE FOR AEIs

One of the major hassles towards accepting AEIs as patent worthy is based on a presumption that these patent applications are incapable of properly and fully disclosing technical constructs, for larger part of invention building happens within deep layers of intelligent machine, not exactly known to any human. Therefore, whatever is submitted in the name of complete disclosure will always be deficit of pertinent information that makes the invention reproducible.

However, there still exists an invaluable portion of disclosure that has merits owing to human contribution of intellectual nature that goes beyond the provision of a mere abstract idea, as discussed earlier. Important highlight of human contribution begins right from providing insights on training data used for building training models (AKA “pre-trained model,” “learned model,” etc.) to defining weighted parameters and determining implementation detailed of algorithm to obtain a trained model.

So, let’s explore in which all ways sufficiently detailed disclosure allows a reproduction of the intended technical solution[46]. It will also help in gaining an ancillary understanding- if the disclosure around machine contribution can stay a bit compromised, and yet fulfil foundational requirements of patent obtainment process. So, contributions and disclosures by relevant stakeholders, especially data scientist and programmers, in invention building process requires consideration at granular level before awarding inventorship or other moral rights in a patent application[47].

a)    Data Scientist: Since data is a primary feed or raw material for an AI algorithm to function and produce an actionable output, role of data scientist becomes eminent. Their valuable contribution, therefore, needs a critical evaluation.  In case of intelligent data mining, a data scientist is primarily tasked with formalizing of technical problem, curation of structured/unstructured data that eventually assists AI scientist in selecting the fundamental blocks- methods, algorithms, architecture, NN topology, etc. to be used[48]. The real-world data is messy and often needs to be normalized, transformed, have outliers removed, or otherwise processed so that the AI model can produce useful, concrete, and tangible results. In order to do so, the data scientist can either use known techniques from a library or software tool or develop proprietary algorithms that may be adapted to the context of technical problem, such as designing specific classification algorithms. Right from input data preparation and its quality ranking[49] – how is data gathered, pre-processed, handled, or parsed upon use by the AI model constitutes measurable parameters for generating a useful invention.

Under such situations, where the data scientist employs inventive techniques to prepare quality data of particular relevance, provides guidance to AI machine to uniquely contribute towards finally commercially valuable output in a non-obvious way, then he shall share titlehood of such invention. On the contrary, if a data scientist merely collaborates with an expert and performs an obvious step of creating training data set under directions of such expert, then it is a mere administrative activity or workshop variation of what was already known. Hence, as is prescribed in patent law, contribution by way of non-technical factors will not confer any inventorship/ownership in patent rights.

b)    Programmer/Developer: Next, the developer or programmer selects a set of mathematical models or writes an initial algorithm to process curated data and build the training model. A trained model is an algorithm based upon a mathematical function that generates optimal output based on the learned patterns in the training data. Determination of optimal architecture before the training process relies much on heuristic* methods and human know-how[50]. During the training process, training data is fed into the model, based on which the training algorithm optimizes trainable parameters to minimize loss function. Here too, choice of particular training methods requires technical know-how for which it relies on certain heuristic methods*- an approach to problem-solving relying on experience and intuition rather than a pure scientific methodology. Heuristic methods are often used due to the lack of sufficient computing power or the absence of exact methods for the solving of certain problems. Role and contribution of programmer therefore inarguably remains noteworthy, and qualifiable for patent inventorship.

Post creation of trained model, the machine can make predictions and recommendations, and also continue improving its end results with self-training and learning. These details- everything from mapping of input data to the model, set of mathematical constructs, training process to obtain the training model and validation methodologies are important inclusions for disclosure of an AEI[51]. How the training data is collected, data mapped into “features” (the actual inputs of the model), input data pre-processed for feature extraction (if any), or model being trained, type of data or features provided to the trained model, or model output post-processed or interpreted are a set of important questions, the answers to which the examiner and those interested in field will be tempted to look for in such patent applications.[52]

A marked distinction should be established between the direct output of a model and the potential practical application(s) achievable post processing of intermediary output, if any claim lists so. For example, in some cases the raw output of a model has to be transformed, normalized, or run through another algorithm to provide useful output data. In others, the output of one model may be used (with or without intermediate processing) as input to another model, say for example, a particular layer of a neural network encoding a semantic meaning of the input. Such modifications and possible end results obtainable from these modifications need to be vividly and sufficiently disclosed.

Similarly, if a deep neural network is generated as trained model structure that has artificial neurons organized in multiple layers to process input data with multiple levels of abstraction[53], then the model structure along with specific non-generic features (e.g., a neural network with non-conventional number of nodes at given layers, multiple hidden layers, etc.) and mapping of input data to categorical, commercially valuable output to generate labels on future inputs are crucial details that will be expected from such disclosures. Further, other significant details such as mode of implementing training parameters, training algorithm (e.g. regularizers, tree size, learning rates), hyper parameters, input variables, optimization variables, training data sets, validation data sets or number of layers utilized to derive potentially meaningful and useful output, and other such details requires a detailed discussion in patent draft[54]. Network’s detection of fine features in input data, working of multiple neural networks in parallel or in tandem, application of weighting function –all of these are fundamental aspects for a neural network tool, and hence all details related to even setting of weighting parameters, teasing out subtle proxies or patterns within data or finding differences in input data are other examples indicative of extent of disclosure warranted[55].

Other seemingly important disclosure includes type of algorithm, type of training methods used to develop algorithms, type of training data, period of training, optimization of outputs, and other such extensive implementation details etc. Even if models appear ‘intelligent’, they generate output by merely relying on probability calculations. They are not autonomous (i.e., they do not ‘reason’ on their own) and need to be fine-tuned by machine learning experts. While it can be challenging to explain why an AI algorithm made a particular decision or took a specific action (due to the black box nature of such algorithms once they are fully trained), it is generally not difficult to describe the structure of algorithm or how a system embodying it works.

 

C.    DISCLOSING FORESSEABLE ASPECTS OF AEIs

For AEIs, one characteristic feature is the ability of machines to graduate itself after being trained once. Under such circumstances who should be held accountable for unpredicted results, which may manifest tomorrow?  Should we be interested in capturing these hidden details; or is it absolutely fine to continue with the traditional paradigm of disclosure? One compelling reason for considering such foreseeable output as part of present disclosure stems from fact that these futuristic, anticipated results are borne out of extreme human endeavor and diligence, and are peculiarly not entirely machine generated[56]. Selection of data and training of the algorithm to produce foreseeable results are outcome of individual’s intellectual labor as final predictive outcome is originally ideated, implemented and intimated by human mind. Therefore, his rights over such foreseeable and predictable variations of invention cannot be outrightly denied, as exercised in KSR Int’l Co. v. Teleflex.

Along with submitting technical details of present technical output, the applicant shall also disclose in sufficient detail his insights on foreseeable results[57] that may be exhibited if machine continues to operate on a similar data set, execute algorithmic instructions in a linear fashion or improvise to an extent previously established by human co-inventor. In order to demonstrate that certain end results are foreseeable, predictable and succinctly replicable, a very detailed account of obtaining them shall be submitted as a conclusive proof. Preferably, detailed algorithmic instructions used for obtaining technical result must be characterized in written documentation or included as software codes meaningfully explained in English language in patent specification[58].

Disclosures related to training phase including how a model is trained, what weights are used with respect to what variables and substantive features contributing to corresponding advantages resulting from execution of training model will be crucial for determining spectrum of human intervention requisite in claiming foreseeable results. Whether or not the disclosure of sensitive training data sets is required and to what extent may be dependent entirely on the criticality of such training data sets for carrying out the invention. Primary reason being data is usually a subject matter of other forms or types of IPRs, and when exclusively claimed, it is essentially disclosed as a part of submission. However, if the training data set is not critical for reproducing/explanatory purposes, disclosure may not be necessary. Initial algorithm will be sufficient as they are relatively constant, and merely adapts to varying data over time.

Agreeably, there may not be exact mode of implementation for achieving foreseeable results as machine is continually upgrading its internal state in response to training data, improving its performance, and adapting to changes in database contents[59]. Nonetheless, the inventor shall submit with enough specificity the seed information, specific input configuration including newly invented methodologies or approaches that can unambiguously explain the predictable real-valued output[60]. Purpose is to explain the machine learning output, i.e. to understand the factors driving the given model to a concrete output.

At the same time, one has to be mindful of not letting this submission stand in contravention to long-established axiom of “acknowledging inventorship only when there is an actual participation in creation of invention beyond identifying of a goal or foreseeable result, rule embodied in Oasis Research, LLC v. Carbonite, Inc. Real participation will be established only when the specification states the possibly predictable variations or improvisations of invention explainable from submitted content.

 

D.    PROCEDURAL FLOW FOR  ABIs COMPRISING UNFORESEEABLE ASPECTS

As detailed above, foreseeable aspects of AEIs may be captured in a detailed disclosure to claim ownership over end results that a human agent presumes machines may output in due course of time. On the contrary, if it is discovered that machine has intelligently ingested new data and evolved itself to a magnitude inexplicable by previously patented technology, then logically a new patent eligible subject matter is borne, now referred as ABIs. Say for example, distinctiveness of previously designed machine and its particular results are now irreproducible or uninterpretable with regard to expansive functionality of machine, topology of machine or type of data manipulated in course of achieving new and inexplicable results. Simply put, AI independently creates a patentable invention and role of human agent is no more that of a non-inventing onlooker. No exclusivity can be declared over product that is efficiently generated, simulated, reflected[61] upon and evaluated amidst large number of potential solutions by sophisticated machine without usual limitations imposed by human biases or time constraints[62]. How to accredit machines with sole inventorship and make its human counterpart responsible? One probable solution to overcome this overhanging problem has been proposed here with some procedural adjustments suggested for present patenting system.

a)    Filing of a Technical Note aka Provisional Specification : To begin with, when a machine that was previously acknowledged as a co-inventor with human agent for a patentable subject matter along with its probable foreseeable results, develops an invention absolutely autonomously, human co-inventor may notify the patent office upon encountering AI Borne invention (ABI). Human co-inventor may have to establish in a technical note how ABI is not similar to parent patent application previously submitted for a similar subject matter. Once such an intimation along with a preliminary technical note is received by patent office, it may permit the applicant to treat this technical note similar to a provisional application for purposes of obtaining a priority date before competitors could appropriate the invention.

Following the usual course, human-co-inventor may now begin building upon the disclosed technical note aka provisional application to deduce necessary information that lead to the newly innovated product of machine. In order to explain it fully, he may have to reverse-engineer the final product to figure out technical approach that lead machine to build a new product. Similar to fair-use doctrine in copyright law that permits reverse-engineering of copyrighted software for at least some purposes[63], reverse-engineerability of a new found product can make successful integration of technical output with a practical application. This will also serve primary utilitarian purpose of patent law aimed at incentivizing and rewarding innovative activities, diffusing knowledge for proper use of benefit of society or progress of science and useful arts[64]. As Professor Jane Ginsburg has observed, “[e]ven the most sophisticated generative machines – those that employ adversarial neural networks to generate outputs – are no more than complex sets of algorithmic instructions whose abilities are entirely attributable to how programmers train them with input data, and how programmers instruct them to analyze that input data.

In absence of human supervision, these smart machines may continue endlessly upgrading themselves, their valuable technological finding meeting a dead end without any tangible application or profitable end use. If no recognition is meant for these inventions, then why would there still be any such invention. Thus, human who tooled ABI in a particular way to generate the inventive output, irrespective of the fact that the "heavy lifting" has been done by the AI system itself, must be entitled with patent rights[65].

b)    Filing of a Complete Specification: Once the technical note is admitted, a complete patent specification demonstrating in fullest detail technical solution to a technical problem, and having utilitarian impact shall be submitted by human co-inventor within a time period of one year from filing a provisional application, in a manner very similar to conventional patent process. Within this period of 1 year, the human-co-inventor may draw up a way of manipulating black box operations towards a tangible application and create a patentable solution. One may argue that conceptualizing an altogether a new, useful and inventive product takes a considerable time, and period of one year may not justify development of a quick patent worthy solution. True it is, but the rationale is- here the product has already been invented by machine, and a human agent has to simply reverse engineer it and decode the technical means followed (discussed in next section).

For this enhancement, if the human agent is acknowledged as co-inventor in partnership with machine, the desire for driving ABIs to patented products is uplifted for human co-inventor. But now a next logical question follows- why would a human-agent even declare that he has reverse engineered the product? Why can’t he simply claim to have devised the machine and let the application proceed similar to AEIs? Next section attempts to find a way out.

 

E.    DISCLOSURE REQUIREMENTS FOR ABIs

Answering to a thorny question raised in previous paragraph, we need to first understand reverse engineering from machine learning context[66]. In principal, it is the possibility to extract or deduce certain elements of the machine learning process through access to other elements, which usually is controversial. Straightforwardly, it is unrealistic even for experts to predict what the algorithmic engine is capable of doing after it has rewritten itself several times over using machine learning without human intervention[67].

In nutshell, these black engines mostly remains inscrutable, and it is extremely challenging to reconstruct its internal workings or even recreate private data on which the machine has trained itself. Extracting exact parameters, decoding opaque algorithms or reverse engineering well trained and complex model characteristics[68] are discoveries virtually impossible. Once trained, ML algorithms are not aptly indicative why it gives a particular response to a set of data inputs[69]. Amidst these apprehensions, it is largely understood that written description expected of a wholly disclosed patent application may be bereft of significant technical implementation or executional “how’s” of disclosure, as evident in Ex Parte Lyren[70].  Besides, the claims may be only directed towards systems architecture achieving the final output, and not exactly detail the steps or process flow of claimed output. Analyzing few patents (US Patents 5659666, 7454388, 10423875) of Stephen Thaler’s Creativity Machine, it was observed that all of these applications embodied only system or device claims, ignoring the process/method claims. Though such patent applications have to be looked into greater detail, and are part of my future work, but it became quite evident that much legal uncertainty exists in drafting ABI related process claims, where methodologies and associated details may be subject to different interpretations by various courts.

Under these constraints, an equitable adjustment restricting the requirements of submission only to final utilitarian output achievable by machine along with enough explanation fastening the final output of machine with target use, seems a workable proposition. It implies that the applicant may need not have to disclose exact details of process flow by which the system arrived at final outcome. Candidly speaking, true disclosure relating to process/method patent claims for ABIs apparently subjects them to patentable subject matter exceptions (for being abstract) or denied for not disclosing enough. Preferable would be omitting such method claims, and reinstituting faith in product patent regime for ABIs[71].      

In consideration of such relaxation, product patent applications may be arranged for a quicker prosecution and shorter patent term as for petty patents[72]. Fast-paced grants may be key motivational factor for human (co-inventor) for disclosing submission under category of ABIs. Besides, it will also offer a psychological advantage of social recognition for his useful discovery.

Simultaneously, for the patent office, it will be less burdening to assess product patents as they may not have to delve deeper in complex performance details of ABIs. Further, product patent regime will promote scientific advancements demonstrated by ABIs instead of leaving them as plethora of meaningless references having limited practical application without human supervision[73]. Most importantly, above explained sui generis patenting system can be seamlessly integrated in current patenting doctrine without requiring major overhauls. Radical though they may be, the changes that this framework will bring shall, if properly managed, reinforce the societal and economic benefits that the patent system was always meant to deliver. The solution-sui generis legal framework for AI enabled inventions- does not solve the one-size-fits-all problem inherent to the patent system, but caters to the challenges of building a coherent AI subject matter doctrine and correcting deficiencies of patent law currently dealing with it.   

 

CONCLUSION

In light of legal uncertainty in the context of rapidly advancing AI technology, it is important for policy makers to give serious consideration to the issue of inventorship/ownership to AEIs and ABIs. For purposes of issuing guidance in this area, it is imperative to reconsider the boundaries of patentability, patenting process, decide how solution of today can help us prepare for super intelligent machines of tomorrow, and what adaptations may be necessary to ensure that the patent system’s fundamental objectives are high held. While a probable sui-generis model of patenting AEIs and ABIs is suggested, it falls to policy makers and eminent thinkers to examine the fundamental rationale and justifications proposed framework may fulfill. Whatever may be the outcome, the fact remains that it is extremely urgent to address the patenting issues associated with AI machines in a proactive manner, before the courts begin setting unsettling precedents for this technology domain. 



[1] (Kemp, 2018)

[2] (Burke, Todd J.; Trazo, Scarlett, 2019)

[3] (Artificial Intelligence Collides with Patent Law, 2018)

[4] (Shemtov, 2019)

[5] (WIPO, 2019)

[6] (Public Consultation on AI and IP Policy – Submissions, 2020)

[7] (Vertinsky and Rice, 2002)

[8] (Shemtov, 2019)

[9] (Kohlhepp, 2008)

[10] (McLaughlin, COMPUTER-GENERATED INVENTIONS, 2018)

[11] (SUMMERFIELD, The Impact of Machine Learning on Patent Law, Part 1: Can a Computer ‘Invent’?, 2018)

[12] (Müller & Bostrom)

[13] (Davidson, 2019)

[14] (Joshi, 2019)

[15] (Josef Drexl, 2019); See also (Fraser, 2016)

[16] (McLaughlin, 2018)

[17] See Artificial Neural Networks, IMAGINATION ENGINES INC., http://imagination-engines. com/iei_ann.php [https://perma.cc/BB8K-G3FH] (last visited Jan. 25, 2016)

[18] (Walsh, 2018)

[19] (Morgan, 2020)

[20] (RAVID & LIU, 2017)

[21] (Shalev-Shwartz & Ben-David, 2014)

[22] (Chui, Manyika, & Miremadi, 2018)

[23] (Amershi, Cakmak, Knox, & Kulesza1, 2014)

[24] (Surden, 2014)

[25] (McLaughlin, Computer-Generated Inventions, 2018)

[26] (Heller, 2019)

[27] (Samuelson, 1985)

[28] (Abott, 2015)

[29] (SUMMERFIELD, The Impact of Machine Learning on Patent Law, Part 2: ‘Machine-Assisted Inventing’, 2018)

[30] (Fraser, 2016)

[31] (Surden, 2014)

[32] (Bartlett, Solving the AI Accountability Gap, 2019)

[33] (Tull & Miller, 2018)

[34] (Firth-Butterfield & Chae, 2018)

[35] (Samuelson, 1985)

[36] (Schuster, 2019)

[37] (Nicholas, 2013)

[38] (Fruh, 2019)

[39] 35 U.S.C. § 112 (emphasis added); America Invents Act sec. 4(c) (striking “of carrying out his invention” and inserting “or joint inventor of carrying out the invention”

[40] Ariad Pharm., Inc. v. Eli Lilly & Co., 598 F.3d 1336, 1351 (Fed. Cir. 2010) (en banc).

[41] (Ouellette, 2012)

[42] (Surden, 2014)

[43] (Cloem.com, n.d.)

[44] (Bjorkman, Voortmans, & Block, 2013)

[45] (Fruh, 2019)

[46] 35 U.S.C. §§ 101–103, 112 (2006), amended by Leahy-Smith America Invents Act, Pub. L. No. 112-29, §§ 3(b)–(c), 4(c), 125 Stat. 284, 285, 296 (2011).

[47] (Tremaine, 2019)

[48] (DeCosta & Carrano, 2017)

[49] (Clifford, 1997)

[50] (Josef Drexl, 2019)

[51] (Borella, 2018)

[52] (Hulbert & Berghoff, 2019)

[53] (Thaler, 2013)

[54] (Borella, Gin, & Herndon, 2019)

[55] (Russell & Norvig, 2009)

[56] (Abbott, 2016)

[57] (Balganesh, 2009)

[58] (Bhandari, 2009)

[59] (SUMMERFIELD, The Impact of Machine Learning on Patent Law, Part 1: Can a Computer ‘Invent’?, 2018)

[60] (Schuster, 2019)

[61] (Malenfant, Jacques, & Demers, 1996)

[62] (Plotkin, 2009)

[63] (Cohen & Lemley, 2001)

[64] (Motion Picture Patents Co. v. Universal Film Mfg. Co., 1917)

[65] (Block, 2017)

[66] (Josef Drexl, 2019)

[67] (Pai, 2017)

[68] (Tramer, Zhang, Juels, Reiter, & Ristenpart, 2016)

[69] (Lewis & Monett, 2017)

[70] (Higgins, 2019)

[71] (Rathi, 2019)

[72] (Richards, 1998)

[73] (Hattenbach & Glucoft, 2015)