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]
[2]
[3]
[4]
[5]
[6]
[7] (Vertinsky and Rice, 2002)
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[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]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[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]
[42]
[43]
[44]
[45]
[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]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57] (Balganesh, 2009)
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]