Artificial Intelligence, Threats, Challenges and
Opportunities
Prometea, the First Predictive Artificial
Intelligence at the Service of Justice is Argentinian
by Juan GUSTAVO
CORVALÁN, Administrative and
Tax Litigation Judge of the
Autonomous City of Buenos
Aires. Currently serves as Deputy Attorney General in Contentious Administrative and Tax
before the Superior Court of Justice of the Autonomous City of Buenos Aires[1].
Towards the end
of this second decade of the XXI century, technological advances are
transforming science fiction into reality. Until a few
years ago, it was utopian for the following questions
to become a challenge for
the legal systems: Who’s responsible for
the consequences of intelligent machine functioning (autonomous
vehicles, amongst others)? How can we guarantee the human auto-determination in the artificial intelligence era? How is it
possible to “program” the artificial intelligence in order to be able to include a legal and ethical approach? Is it feasible to consider the access to artificial intelligence as a
right? How can we avoid
artificial intelligence from worsening the
inequalities between
people? The answers to these
questions demand a transcendent effort to
rethink and innovate about the
challenges of the new revolution
that new that we are going
through.
During
the last three centuries, three major industrial revolutions are usually mentioned. The first one related to the development
of railroads and the steam engine in order to mechanize production. The second one related to electrical energy and the assembly
line to develop mass production. The third revolution revolves around the emergence of electronics computers and the information technology to automate production[2].
We
are currently undergoing a new revolution that is linked to several phenomena (nanotechnology, biotechnology, robotics, internet of things, 3D printing)[3]. The most disruptive phenomenon is the development
of artificial intelligence (hereinafter also AI), that is presented as
an innovation connected
to the technological advances pertaining to
information and data
processing (also in this area there are other inventions from the last
century such as computers, internet, the world wide web – www- search engines,
and others that can be mentioned). The epicenter of the
“Fourth Industrial Revolution”[4]
is the exponential increase of two
factors; the storage capacity and processing speed of
the information and data.
To get an idea, it is now possible to measure the flow generated
from the use of information
and
communication technologies main tools, in real time and on a global scale. For example, on May 29th 2017, 458.090 thousand
tweets were sent,
63.980 thousand pictures were posted to Instagram, 3.629.947 million searches on Google
were
made, and the
web
processed 2,702.994
Gigabytes[5]. This massive volume of data and information cannot be efficiently processed by human beings. Therefore,
artificial intelligence is the revolution of revolutions. Its current and potential development is due to the fact that it
manages to equal or widely surpass certain cognitive capacities, by processing data and information more efficiently in increasingly more human activities.
Although we might not be aware of it, we are undergoing an unprecedented era in
human history. Among
many other reasons, this is
because we are witnessing the progressive elimination comprehension of language comprehension barriers almost instantaneously, from the exponential development of the system of
artificial intelligence used by the Google
translator uses. The
famous artificial intelligence translator is
one among several artificial
intelligence systems that process natural language. In essence, it uses a learning method associated to a vast number of related cases; that is to say, it is not based on learning or applying the grammatical rules of each language.
Simply put, it creates giant databases linked to common translations; which are supported by documents translated from one language to another, using documents translated by the United Nations Organization (UNO) into different languages. By the end of
2016, the Google translator almost matched up to 500 expert human translators. Let us see more closely the operation of this AI system.
The Google translator test consisted of using a scale from 0 to 6 to score translation fluency of 500 sentences taken from Wikipedia or the
news. For English to
Spanish translations,
Google’s new system obtained
an average score of 5.43, very close to the 5.55
obtained by the 500
human translators. If you have not tried the translator for a long time, it is
probably due to the fact that it presented
several mistakes before. In
Google’s testing, human beings graded the new system between a 64% to an 87% better
than the previous version.
That is to
say, instead of learning concepts or grammar, the
algorithms establish their own ways of breaking down texts into smaller fragments, which
normally seem to lack any
sense and generally don’t correspond to speaking phonemes. And not to mention the speed at which the sophisticated AI
system works. One of it
designers stated: “It might seem disconcerting, but we have tested it on several sites and it simply works”[6].
However, the
understanding of a phenomenon like this requires briefly addressing the concept
of human intelligence with which cognitive science experts works. Let’s see.
Among several definitions and conceptions
of the intelligence concept, the common element is the
capacity to process information in order
to solve problems so as to reach objectives[7]. The
notion of intelligence is
indissolubly connected to information
processing.
It is important to clarify that we speak of information
in a broad sense and as per the approach of the cognitive
sciences, which allude information processing or the information
flow of the environment that is codified, organized, selected, stored and retrieved through the sensory, perceptive systems,
among others. Human intelligence is related to a relatively autonomous set of capacities
or cognitive qualities that are often
classified in “intelligence profiles” or “multiples intelligences”. These
are: social intelligence, linguistics (or musical) intelligence, logical-mathematical
intelligence, interpersonal intelligence, and intra-personal or emotional,
fluid intelligence among others.
Basically, our brain controls the ability to process information from the environment
and
our own body[8]
that is used to evaluate and choose future courses of action. And here come the decision-making process and evaluation, which consists of
selecting, cutting and organizing the
available information.
Based on human
intelligence, multiple and diverse technological innovations have been
developed. We will be dealing here, with information
processing for problem solving and making decisions based on machines
operating with the referenced intelligent algorithms. AI is substantiated by intelligent algorithms or learning
algorithms that, among other aims, are used to identify economical tendencies, predict crimes, diagnose
disease, foresee our own digital
behavior and so on. An algorithm can be defined as a precise group of instructions or rules, or as a methodical series of steps that can be used to make
calculations,
solve problems and make decisions. The
algorithm is the formula used to make a calculation[9].
Now then, during
the last decades, different methods to
develop algorithms using large volumes of data and information have been employed (some of these methods are:
neutral networks, genetic algorithms, reinforcement learning, amongst others). In essence, it is sought, though
AI implementations, for technologies to allow the computing systems to
acquire: self-reliance, self-adaptive reconfiguration, intelligent negotiation,
cooperative behavior, survival with reduces human intervention[10], among other features.
All this implies the use of different
techniques that are based on the
recognition of patterns in order
to solve problems[11], maximize objectives and
optimize information processing. Let’s see another
example understand how the most
successful AI systems work.
In 2011, the artificial development developed by IBM named “Watson”, competed against human beings on a game of Jeopardy. This United States television contest based on
answering questions about numerous subjects such as history, languages, literature, among others, and lies in that each of the three contestants choose one of the game board panels that, when uncovered, reveals a
clue in the form of a response. The contestants have to give their answers in the form of a question. For example, some of the questions are formulated as follows: “It could mean a
gradual mental development or something that is carried during
pregnancy” or “A long and tedious speech written in the trivial cake dressing.”
Ken Jennins is the supreme game winner (74 times in one year). In February 2011,
Watson competed against him and another champion, Brad
Rutter. Watson won
the game, and it did so
because, in essence, it can process more data and faster than the human contestants. While AI considers millions of possible hypotheses at the same time,
it will take several centuries for a human
being to analyze all the deliberations that Watson in three seconds.
Let's see some figures that reflect the data and information magnitude that this AI processes.
Watson was able to process 200 million pages coming from documents, including entire
encyclopedias, which include Wikipedia and natural language full of ambiguities
and vagueness (we are talking
about one thousand books per
second). The system it uses also includes a
subsystem that helps to calculate the degree of reliance in the response that the AI will ultimately provide. Simply put, Watson has an assistant. Intelligent algorithms managed by another algorithm that functions as an expert consultant. We are witnessing the creation of another
character which is added to the classic investigating
pair: Sherlock (IBM), Watson, and its expert algorithm consultant.
As this artificial intelligence has read hundreds of a million informative pages that include
stories, it is able to follow the thread through
complicated sequences of events. Watson optimizes information from hierarchal statistical processes, learns
from its experiences and at a speed impossible to overcome by a biological
organism. But the most relevant is that of the acquired knowledge comes from
itself and from all the information it obtains, beyond the few other areas of
the data that were programmed to this AI directly by human beings[12][F1] .
The Google
translator and
Watson are two examples among many
others that exemplify the artificial intelligence tsunami that is being developed in multiple fields. These are systems that create music, paint
pictures, recognize faces, objects, predict successful companies
in the stock market, detect
diseases, and
help to protect the environment, among many others. We
are witnessing a real race to develop
AI to simplify and optimize various human activities. It is in this scenario that three main challenges arise within
the legal field. On the one hand, how do we protect ourselves from the
intelligent algorithms that are
replacing and surpassing us in
multiple activities. On the other hand, how do we make this new technology contribute to sustainable and inclusive development of human beings? Lastly, how will human rights
be protected and
transformed in a transition that seems to
be directed towards symbioses between the biological, digital and
artificial.
In this first approach,
all these issues could be redirected from exploring two main axes. The
first one, related to what can be referred as “the luminous side of artificial intelligence.” Here, this
technology is drastically disruptive to enforce certain rights and, at the
same time, represents a qualitative lead leap on the
ways organizations and their relationship with citizens shall be managed.
The second axis is related to what we refer to as “the dark side of AI”. From this
perspective, it is important
to highlight the risks that are generated from
the so-called existential risk associated to the
possibility that human beings may lose control over artificial intelligence. But leaving this extremely complex issue
that is projected in the medium and long term aside, other aspects related to the short term must be addressed. For example, issues related to the impact
produced on the fundamental rights of human beings by the development and use
of AI systems. A few brief proposals will be highlighted in the following
points.
Many times, we tend to refer to asymmetrical development which usually is inherent to fewer advantages
countries. In the technological sphere, the presence of this phenomenon accounts for various
asymmetries within its development. For example, in the Autonomous
City of Buenos Aires there is an entire
digital Public Administration (in addition to this, more
than 58e regulations which regulate
the Public Administration’s digitalization have been issued), while in other provinces this process has not even started. This situation also occurs on a much deeper level. Currently, 17% of the world’s population, 1.300 millions of
people, does not have access to the benefits from the second industrial revolution (electricity). And, more than half
of the world populations, 4,000
millions of people are still not connected to the internet[13].
In asymmetries within development are
normally accompanied by the need for protection
and rights effectiveness which are also
dissimilar. That is
to say that the challenges regarding first necessity issues must be faced (as
water access, essential services, and so on), but also it is important to make progress in protecting other
massive rights violations in the digital world. Even
so, we often consider that as it is difficult to solve certain urgent problems,
it would be illogical to try to address others. However, it is important to think and act conversely. Taking care of basic
issues, it is also important to undertake the new challenges; if not, the spectrum of right violation increases. For example, in
the criminal field there are several challenges that are
very complex to be addressed (such as drug trafficking, illegal sales, and armed
robbery, among others). But at the same time, the digital world entails other challenges and new rights
to be protected appear. We are referring to all the issued related to
cybercrime (such as
cyberbullying, child pornography, grooming, computer fraud, and others). A citizen must also be
protected from these violations to their rights.
In short,
asymmetrical development increases the complexity and demands for major efforts in order to make effective the rights of the people. In Latin America in general, and in Argentina in
particular, there are multiple differences among the people, the districts, and
vulnerable areas. On the one hand, the
digital divide[14]
that exists among citizens that are connected through the internet and those that are not, must be considered. On the other hand, the organization and public authorities’ development
are asymmetrical in terms of
infrastructure and development. However, this does not stop the State from improving on several matters at the same time and to accelerate the transition to adapt to this new space and
time revolution.
On the contrary, the more time it takes, the citizen is the one who “losses”. For example, the
multiple internet and social media problems (all issues related to cybercrime and privacy) have been affecting people that are connected for a long
time now, no matter if they live in different Argentinian provinces (Formosa, Tierra del Fuego, or others), in
Brazil, in France or in
Italy. Something similar occurs with the implementation of the electronic or digital case file. When Administrations save days with
physical transfers and in diverse paper related aspects, the time
“won” is the same for a citizen from an Argentinian southern province (as Neuquén) in regards to another from the north (as Salta). In conclusion,
we shall advance in innovations favoring inclusion
(innovative inclusion), beyond
the existence of an asymmetrical development. And here artificial intelligence comes to play as an instrument at the
service of justice and rights.
In a broad sense, this approach is related to a new technological that is called
intelligence at the interface. From this optic, the interface[15]
possesses a lot of information about the user, it understands it in context, it is proactive and it improved with
experience. To understand it better, let’s see how correlations develop in the
digital world. On the one hand,
there is a system by which the user chooses the path and technology that connects the dots, as it happens with hyperlinks. On the other hand, is the so-called “portal”, where the user chooses
one channel and the
technology transmits its content. Thirdly, we find one of the
most used, which refers to search engines (the most famous is Google). Here the
user establishes what he wants to search for and the
technology finds the relevant and
quality content in return.
Lastly, there is a
more efficient method: the intelligence at the interface. Here, the user simply interacts
(talking or chatting as if it was WhatsApp) and the
technology solves the problems through connections with different systems that can answer the user’s needs based on learning[16]. In the case of iPhone
cell phones, Apple has developed a voice assistant named Siri. When activated, it can be asked
several questions, it replies and can also schedule a meeting on your calendar if you just ask it to do so. This is also the case of
Prometea, the intelligence that we have developed
at
the Public Prosecutor's Office, which is located within this new paradigm. According to our
investigations, the intelligence at the interface – through
artificial intelligence systems- may have a decisive impact in the reassurance of certain
access rights, and even more when it is the case of
vulnerable or disabled individuals. Let’s see this specifically.
Since we undertook the Public Prosecutor’s Office of
the City of Buenos Aires management, and thanks to the support of
Luis Cevasco, Deputy Attorney General in charge of the Public
Prosecutor’s Office of the City of Buenos Aires, we have been strongly working on new
technologies. Particularly, focusing on information and communication
technologies (ICTs). During the course of 2017, we have developed the first predictive artificial
intelligence system[17] that also works with a voice assistant (as Apple’s Siri) and allows
the confection of a Legal Opinion in full.
The predictive model is amazing, unprecedented and recent (it has one week). The procedure is entirely managed by AI, as follows: a case file that has
not been analyzed by any human being
arrives in order to pass an opinion on this file. The case file number is recorded
on the artificial intelligence, Prometea,
and in a matter of seconds,
everything details below happens. The AI
system searched for the title of the
case in the Superior Court of Justice of the City of Buenos Aires’ website, it associates it
with another number (linked to the main proceedings) and then heads to the Judiciary website of the Autonomous City of Buenos Aires
(Juscaba). It
searches and reads First and Second Instance
ruling, and it then it
analyzes more than 1.400 Legal Opinions (issued
during 2016 and 2017) to finally issue a prediction. In short,
it detects a
specific model to solve the case file, and offers the
possibility of completing a few
facts so as to print or to send the
Legal Opinion to be revised based on that model (the same could be done in order to render a sentence).
During the month of October, we will be working on validating the predictive system (which implies measuring the time reduction
among other things) beyond these first samples, Prometea’s accuracy is
amazing. Even, a case
file that arrived at the Public
Prosecutor’s Office that had already been
pre-classified by a member of our team about one subject, in a few seconds, Prometea suggested the application
of another judgment model
related to another
different subject, and precisely the latter was the correct one.
But even before we were able to develop this predictive system which (according to our investigations)
is unprecedented, we had already implemented artificial
intelligence using a virtual assistant modality (application of intelligence at the interface). We have been working on a pilot
test for several months with different
case files,
where the person that opens it
(they are not digitalized yet), and once they are
ready to draft a model, they
activate Prometea by a
voice command on a
cell phone (mobile device) or
by
chat, as if you were to have a conversation through
WhatsApp). The entire process is
done by means of artificial
intelligence. From a ‘hello’, to
several questions and answers between Prometea and
the individual, which includes searching and “bringing” laws and decrees to the Legal Opinion, until we get to the point
(depending on the model we are working on) where Prometea tells us
that the ‘Legal Opinion is complete’. Then, we can order it to ‘print’, ‘download’ or to send the draft by email or to an internal
network, so as to be corrected.
If performed by voice command, the entire process is completed without touching
the keyboard or the mouse. When we turn on the computer and activate Prometea,
it asks us to inform the case file number, and then it obtains the case file
title from the
official website of
the Superior Court of Justice of the Autonomous City of Buenos Aires, and it offers a model
of the Public Prosecutor’s Office with a complete case file title and the subject according to what we express. For
example, a citation model, a housing model, and/or a self-sufficiency model.
It also notifies if the Legal Opinion model is not applicable, because the
deadlines have expired or due to the lack of formal requirements.
In order to avoid any failure, as we are in the experimental phase, the models executed by Prometea, before they are sent to be signed, are checked by the team that works at the
Public Prosecutor’s Office.
By the end of the year, we aspire to pass judgment on at least one
half
of the case files that arrive
to the Public Prosecutor’s Office[18]
by using Prometea.
However, this innovation that takes place in the public sphere implies a qualitative
development in regards to speed and precision in our daily work to provide a better justice
service. The tests we have performed on more than 40 case
files demonstrate that Prometea
is between a 200% and a 288% more
efficient, depending of the model being considered. Also, these figures will increase even more when
the predictive model is fully operational.
The most important aspect we noticed when developing Prometea,
is linked to its
extension into other areas.
Simplifying the interaction with a prosecutor, organize internal judicial processes, optimize
citizenry-State relationships and, above all, to focus its use in vulnerable areas and individuals with
disabilities. Here is where the luminous side of the AI becomes evident.
That is, that artificial intelligence can be a key tool in
the citizenry-State relationship. A system like the one we developed in thea system like
the one we developed in the Prosecutor’s Office could be applied to multiple procedures and
services within the Administration or as a bridge to radically simplify the
logic of many access rights. The
State’s procedures or services could be provided
through digital voice assistance or through chatbots[19]. In fact, if you
have
an iPhone and enable the Siri function, try calling 911 just by requesting it and the assistant will do so. Even with this technology, it is much simpler to
guarantee the centrality of the user through a unique or digital[20]
portal.
In turn, the use of
artificial intelligence could optimize
the data and information flow
available for public (and private) organizations
in order to solve issues that before used to
require multiple steps, procedures and
phases that could not even be resolved.
Although it exceeds this article to keep extending on all these issues, this type of
technology, at the service of rights, becomes a right itself. Even more so, if we consider
that the Argentinian Digital Law No. 27.078 (Sections 1
and 2) declares the development of information technologies and
communication (ICTs) of public interest; and speaks of ensuring the human rights
to communications, telecommunications and, also, the access to
ITCs
services.
All technological innovation produces benefits, risks and damages. Among other advantages, the Internet is vital to ensure the
right to freedom
of expression, but, for example,
it is also used for arms and organ-trafficking, as well as many other crimes that take place in
the digital world. Taking this aspect into account, in the two previous points
we have addressed the luminous die of artificial intelligence. Now, we will trace a few lines about the
risks, challenges and treats that this new technology has for us. The “dark side of AI.”
Earlier this year, more than two thousand five hundred experts (Stephen Hawking amongst them) established the following principle:
“Risks posed
by AI systems, especially catastrophic or existential risks, must be subject to
planning and mitigation efforts, commensurate with their expected impact. And,
to a greater extent it shall be subject to strict security and control measures[21]”.
Currently, there are multiple
challenges to ensure the compatibility of artificial intelligence development
with the State’s domestic law and with existing international law. Intelligent algorithms are used to capture
all our data, recommend us what to
search for, where to go, what to do, how to get to a determined place
faster, diagnose diseases, prevent them and so on. All these issues require
specifying certain aspects that have to be taken into account.
Firstly, it is
fundamental to know how this technology works. An adequate regulation
cannot be thought without knowing the dynamic of the object to be regulated. Something
similar occurs with legislation
regarding food, medication, amongst others. In this aspect, as the AI systems develop
exponentially, it is
essential to be constantly updated with the new
methods used.
Secondly, develop it is fundamental to analyze certain
areas and rights in a particularized manner. It is different
the assumption in which the AI systems recommend and manage our musical preferences in our
Spotify account (or which videos I would like on
YouTube), and another different
assumption is the way in which intelligent algorithms predict if I
will have an illness or if a restrained individual shall be granted
probation.
Thirdly, when it comes to
fundamental rights, it is essential to consider an outstanding aspect of all
the most sophisticated systems or artificial intelligence used nowadays (Watson
of IBM, Alex, Quid, Siri, among others). It is about authentic black boxes. In essence, this
means that algorithms cannot offer a detailed explanation on how they reach a certain result.
That is, it cannot be established how the AI system evaluates and analyzes the data and the information
that it processes. That is why it is referred to “black boxes”. Computational tools in which one understands the entered data and results,
but cannot comprehend the
subjacent procedure,
is called a black box system. Here, the code is inscrutable because the
program “evolves” and human beings cannot understand the process that
the programming followed
in order to
achieve a certain solution[22].
In fourth place, considering the above principles, it is indispensable to
assure the equity and no discrimination principles when facing artificial intelligence predictions
regarding fundamental rights. For example, certain predictive AI used in the United
States of America, are based on a source code that considers race, gender, among others. And this provokes an inadmissible case of structural
algorithm discrimination. In the case “State vs. Loomis”[23] the appellant sustained that
the intelligent algorithm used the
gender evaluations incorrectly[24].
Let’s consider this
issue in detail. The independent news agency ProPublica, held an
investigation on the reliability of the prediction of recidivism by using
intelligent algorithms within the criminal scope. Basically, it analyzed the
way COMPASS[25]
works. Risk scoring
was assigned to more than 7,000 individuals under arrest in Broward County, Florida,
that were assessed between 2013 and 2014, and it determined how many were subsequently charged
with
new offenses over the next two years. The scoring turned out to be
remarkably unreliable
in predicting violent crimes, as only twenty percent (20%) of the
individuals predicted to commit
violent crimes actually committed them[26].
In contrast, when the total number of offenses including minor offenses - such as
driving with an expired license - was taken into account, the prediction proved to be more
accurate. In this case, the success rate of those who actually was sixty-one percent
(61%). In addition,
they detected a higher recidivism prediction rate in
African American offenders and a greater number of false positives (that is, incorrect
recidivism predictions) in this group of offenders. This compared to
the results obtained and corroborated in the
predictions regarding individuals of a Caucasian ethnicity. In addition,
white defendants were
poorly labeled as low risk, more often than
African American defendants.
ProPublica researchers wondered whether this disparity could be explained by the previous crimes of the accused individuals or the type of crimes for which
they were arrested. The answer was negative. Thus,
through a statistical test that isolated the ethnicity effect of criminal
history and recidivism, as well as the age range and gender of the accused, it was
shown that African American defendants still had
seventy-seven percent
(77%)
more
possibilities of being linked to a greater risk of committing a future violent
crime and forty-five percent
(45%) more possibilities to be expected to commit a future crime of any kind. Even, as in the best style of the Greek oracle Calcante,
it was concluded that the developers of COMPAS (Northpointe
company) do not publicly reveal which calculations are used to reach the
defendant’s risk scores, so it is not possible for the accused – nor to the public – to see what could
be causing the disparity[27].
These “artificial oracles” also do not seem to offer an
intelligible explanation – in human language
– regarding how the
factors are weighted or analyzed to
reach the percentages. Imagine the reader who is
discussing with an individual whether or not he owes
the sum of twenty thousand pesos ($20.000) and goes to court. Suppose the judge rules that he
owes nothing, because he
analyzed several factors. But he does not express
how he evaluated them.
Undoubtedly, we would be faced with a typical case or arbitrary judgment. So
if intelligent algorithms are used to help
criminal judges decide on conditional release/parole, our constitutional and
conventional system impedes using such a system. In fact, it is not only a question of ensuring that the system is not based on distinctions of race, gender, or
others, but also that the AI must be
able to explain in a language understandable to humans, which factors it uses
and how it analyzes the elements that sustain them[28]. For these reasons, artificial
intelligence cannot be applied in these fields today. Hence, the importance of the movement called “Explainable Artificial Intelligence”. Otherwise, it is difficult that in the face of such opaque systems, the massive development of autonomous
vehicles, “artificial oracles” that predict health,
safety, and everything related to war weapons by the State matters, be
admitted.
In fifth place, those who
develop AI often rely on trade secret protection and in patent
rights. And while this is a
reasonable at first, when it comes to AI
systems linked to commercial matters (online
sales, advertising, marketing,
or others), it cannot be opposite when it
comes to issues related to an individual’s dignity. In this aspect, section 13.2 subsection a) of
the American Convention on Human
Rights, becomes particularly important when quoting
that the right to freedom of expression “shall not be subject to prior
censorship but shall be subject to subsequent imposition liability, which shall
be expressly fixed by law to the extent necessary to ensure: a. respect for the
rights or reputation of other […][29]”.
As we cannot continue to detail each of the issues involved in the development of AI,
these brief samples highlight the complexity and difficulty
of addressing this innovation. That
is why it is
crucial to put the issue on the agenda, to think about an
international cooperation scheme and, at the same time, to create favorable regulatory. This also implies
incorporating guiding principles applicable for artificial intelligence systems
that are compatible with our human rights model; that is too say, we strive for AI development that is compatible
with
the constitutional State and
with the international law of Human Rights.
The “human rights model” is crystallized from a protective paradigm
that emerges
from international covenants and which
is essentially based on human dignity. In this way, the epicenter of the
system is based on equality, peace, minority protection, of the most
vulnerable or the weakest[30]. It is a scheme that
obliges States and the international community to guarantee the effectiveness
of rights, principles, and rules that are embodied in constitutions,
international covenants and domestic laws[31]. But a human rights approach
linked to new
technologies presupposes accepting a starting point: inclusive innovation for
sustainable development[32].
On this platform, for AI development to be compatible with a ‘‘human rights model’’,
it is
important to promote a regulation that incorporates a series of principles that to a great extent are linked
to three incipient categories that,
by a matter of extension,
will be expanded
in future publications: algorithmic dignity, algorithmic
identity and algorithmic vulnerability. All of them are presented as a derivative of digital dignity (which in turn is integrated into the digital identity) of human-being in
the digital world. In essence, it is a question of making the
protection system more robust by incorporating
a series of general
principles into the legality block tending to regulate it. Let us see.
Prevention/precautions. These principles
constitute two different
functions, with a common denominator the need to act before
any damage production. Radically, they intervene on different types of risks. For potential risks, precautions. For
verified risks, prevention[33].
The precautionary principle within AI (analogous to what happens
in environmental law) is linked with a total lack or absolute scientific
certainty about the absence of risks. When artificial intelligence is intended to impact
on people’s fundamental (health, freedom,
equality, and non-discrimination, security), AI systems could not be used if the
following circumstances: i) a closed source code or existence of a system in which one
understands the data entered and the
results, but the underlying procedure cannot be inferred
(‘black box’); ii) absence of algorithmic traceability; iii) the inability to assure an ‘off button’ or a fail-safe mechanism for AI containment; iv) when at any stage - design, development or application - it is noted that the system is based on distinctions that violate the principle of
equality and non – discrimination. Here this would operate as a kind of
suspicious algorithmic category.
Algorithmic
self-determination. Self-determination
is a fundamental right derived from the dignity of the human being[34]. It aims to ensure the ’’free development
of human personality’’
by recognizing informational self-determination oriented at guaranteeing the
right to choose – associates with freedom of
information, the ‘right to know’, ‘knowledge and ‘information self-regulation’[35]. On this basis, the States and the international community must responsibly
invest and make
every effort to ensure human
self-determination against the use
of intelligent algorithms. As
AI increasingly intervenes between data/information and individual’s decisions, it is
essential to protect their rights by promoting respect for
the necessity, purpose,
proportionality and personal data ownership principles.
Algorithmic
transparency and the impartial validator principle.
When AI systems are intended to be used in the field of health, freedom,
security or other fundamental rights, the design,
development and
use of artificial intelligence must ensure
that no ‘black boxes’ are configured, or that architecture
failures are checked, when they may cause damage or injuries. That is to say, artificial intelligence must be transparent in
its decisions, which means that an ‘understandable explanation’ about the criteria applied to arrive to a certain conclusion, suggestion or result can be inferred or
deduced. This issue has two crucial sides.
In the first place, it is relevant to consider so-called
trade secrets, which protects the confidential business information giving
companies competitive advantages. This encompasses industrial or manufacturing
secrets and trade secrets.
The unauthorized use
of
such information by individuals other than the owner is considered an unfair practice and a trade
secret
violation. Depending on the legal system and the country, the
protection of trade secrets is
part of the general concept of protection
against unfair competition
or is based on
specific provisions or court decisions on the protection of confidential information[36].
In the second place, even if the developer
is willing to ‘open’ the system (the source code), in more advanced AI systems,
there is no technical way to determine step-by-step (traceability) about how algorithms arrive at the result, decision or prediction. This frequently occurs in one of the most used methods: artificial neural networks[37].
Now, so as to address the black box phenomenon, it is important to insist on the fact
that AI systems are designed to maximize results and to optimize
the information and data
processing. However, when individual's fundamental rights (health, life, freedom, privacy,
freedom of speech, among others) are at stake, it is fundamental for the systems’ intermediate
results be validated. This implies that the
reasoning or reasoning structures that are
followed to arrive at decisions or
predictions must undergo
three-phase process: 1) verification, 2) validation and 3)
evaluation[38]. Moreover, here quality and transparent of
algorithmic process assurance comes
into play.
The basic idea is to
achieve that the information and data
processing systems carried
out by IA systems comply with certain quality processes so that the results are the expected ones
and are not obtained at
any cost. The first stage we referred
to is linked to the
architecture of the IA
(verification). It is about ensuring certain standards or principles such as
consistency, completeness, correctness and non-redundancy. Among other methods, it is a question of allowing human experts to stimulate, to the extent possible, the process in order to detect discrepancies.
Here a central factor appears before certain AI that affects or will have a strong
impact on fundamental rights of individuals. Those who design, train or develop
intelligent algorithms cannot participate in the validation process. We will refer to this
as the impartial validator principle. Moreover, it
is essential that the public authorities intervene in the process and be legally obliged to do so. This does not mean that
all AI systems are subject to
this verification and validation process. However, it is important when
developing algorithms that affect life, safety, freedom and health of individuals.
Artificial
Intelligence traceability. Traceability
is the “ability to trace the history, application or location of any entity
through recorded indications”[39]. An AI based on a human rights approach must be able to explain, step by step,
the technical operations it performs from the beginning to the end of a given process. As a rule, the intelligibility of the
intelligent algorithm decision-making process must be guaranteed.
Maximum access. The right of access to algorithmic information. When the State and the public
non-state individuals, by
themselves or through
third parties,
design, develop or use
information or communication technologies based on AI or intelligent algorithms (which
involves any type of machine or intelligent robot), they must guarantee the
maximum access to the
processing system that those technologies perform[40].
The algorithmic non-discrimination principle. The
design and/or implementation of intelligent algorithms should respect the non-discrimination principle, which consists of preventing AI
systems from processing information or data under bias or human
distinctions,
based on race, colour, gender, language,
religion, political or other opinions, national or social
origin, property, birth or other status (Article 2, subsection 2; Covenant on Economic, Social and Cultural Rights).
Promoting the ‘luminous
side of artificial intelligence’, and protecting human rights
from the “dark side of
the AI”, are
presented as the two
transcendental challenges of the Fourth Industrial Revolution. The work that awaits us is monumental, since we must do on a
par with an
asymmetrical development, which places us in front of other problems that have
not been resolved for decades.
However, it is essential not to waste time, because otherwise it would be much more
dramatic than what happened with the emergence of the internet and the development
of Information and Communication Technologies (ICTs) within the United
Nations (UN)[41].
Thus,
while we expanded our possibilities and
simplified the environments
through digitalization, there were
multiple violations of classic and new rights
that the legal system e could
not attend. It is logical for this to
happen, since Law, by
rule, works in a reactive
way. Nevertheless, this technology is different from all and therefore we must be proactive, in trying, to
approach the issue from a multidisciplinary, integral, multipolar, flexible and
dynamic perspective.
In relation to the protection of
individual human rights, it is essential to consider two interrelated aspects. On the one hand, how to guarantee human intervention against the decisions or predictions of intelligent algorithms, by trying to create systems that value the previously developed principles.
On the other hand, if data protection experts are reflecting on human intervention in regards to algorithms[42], then we must work on the following question: how much human
intervention is necessary so that information
processing and
AI data system results are
legit, respectful and promote the effectiveness of human rights.
Prospectively, we believe that
certain challenges presented
by AI have to do with our
identity as species. If human beings are
characterized by diversity, randomness and imperfection, we are entering an era of automation that could put those features in
crisis. Although it sounds improbable, in
a not too distant future, it will be essential to think seriously about guaranteeing a fundamental right, which could be the
foundation of the artificial intelligence era: the right to the inherent and
random and imperfect diversity of the human being.
[1] Lawyer (UBA) Ph.D. in Legal Sciences (University of
Salvador). Visiting Professor of the Master in Digital Law at the University
of Paris 1 Pantheon-Sorbonne and two year Post-Doctorate by the same University. Professor of
Administrative Law at
the
University of Buenos Aires. Administrative and
Tax Litigation Judge of the
Autonomous City of Buenos
Aires. Currently serves as Deputy Attorney General in Contentious Administrative and Tax
before the Superior Court of Justice of the Autonomous City of Buenos Aires.
[2] See Kl. Schwab,,
The Fourth Industrial Revolution pp.
20-12, Debate, Barcelona, 2016, pp. 20-12.
[3] Regarding the influence of some of these
technologies in society, see K.
Michio, The
Physics of the Future,
Debate, Buenos Aires, 2012.
[4] The
World Economic Forum and the International Labour
Organization (ILO),
highlight that the world is going through the Fourth Industrial Revolution. See
‹‹The Future of
Work Centenary Incentive. Informative Note.
International Labour Office, p.2, 2015 and, The Future of Jobs.
Employment Skills and Workforce Strategy for the Fourth Industrial Revolution. ››, World Economic Forum, Global Challenge Insight Report, January 2016, p. 1, [https://www.weforum.org/reports/the-future-of-jobs]; See
Kl. Schwab, The Fourth Industrial Revolution pp. 20-12, Debate, Barcelona,
2016, pp. 20-12.
[5] See ‹‹Internet Live Stats›› [http://www.internetlivestats.com/one-second/#instragram-band] consulted on: May 5th 2017.
[6] See T. Simonite, ,
‹‹Google Translates English Into Spanish Almost as well as a Human Expert››., MIT,
Technology Review, October 4th 2016,
[7] About all these issues: H. Gardner, ,
Intelligence Reframed, Paidos, Madrid, 2010, pp. 52 and subsequent sections (especially p. 115); H. Gardner,., Five Minds for the Future,
Buenos Aires, 2013, p. 17. Also see: FONDO DE CULTURA ECONOMICA, The
Theory of Multiple Intelligences, México, Dictionary
of Cognitive Sciences, 1987, p.
XXIII; MANES, FACUNDO, et al, The Argentinian
Brain, Planeta
Buenos Aires, 2016, pp. 269-270, 274, 275 and 301; F. Manes, , et al, Using
the Brain, pp. 115 and 130.
[8] As per a biological point of view, the DNA
is the essential carrier of genetic information. See M.
Gerard, ., et al, The Biology Book, Ilus Books, Madrid,
2015, p. 354.
[9] See
P. Domingos, The Master Algorithm: Hot the Quest for the Ultimate Learning Machine
Will Remake Our Worlds, Basic Books, New York, 2015, ps.
XVI, 1 and subsequent
sections.
[10] See, .J. Barrat, Our Final Invention,
2013, pp. 205-206.
[11] See more; G., A. Serrano, Artificial Intelligence, RC, Madrid,
2016, pp. 5 and 9.
[12] About this subject; J. Barrat, .Our
Final Invention, 2013, pp. 244-251. It is important to denote that Watson uses a
system called UIMA (Unstructured Information Management
Architecture) that acts as an expert agent that intelligently combines the results given by
independent systems.
[13] See
Kl. Schwab, The Fourth Industrial Revolution pp. 20-12, Debate, Barcelona,
2016, p. 21.
[14] ‘Digital
divide’ is the “separation that exists among people (communities, States,
countries) that use ICTs as a routine on their daily lives and those that lack
of access to the same and even if they did have, they do not know how to use
them.” See A. Serrano, et al, The
Digital Divide, Myths and Facts, UABC, Baja California, 2003, p. 8. Available at
labrechadigital.org. In this sense, the Inter-America Commission on
Human Rights acknowledges that, considering universal access principle
“increase the access and close the digital divide is related to the need that
the State make sure that the private actors do not impose disproportionate or
arbitrary barriers to access internet or use the main services.” See: OAS, Freedom of expression and the Internet,
Annual Report 2013 - Report of the Office of the
Special Rapporteur for Freedom of
Expression, chapter IV, OEA/Ser.L/V.II.149,
Doc.
50, December 31st 2013, paragraph 17.
[15] In terms of websites, the interface is the group of elements on the screen that allows the user to perform actions on the website that is being
visited. That is why, the identification,
navigation, content and action elements are considered part of the interface.
[16] Regarding interface
intelligence, see; T. Gruber,
‹‹Intelligence at the
Interface: Semantic Technology and the Consumer Internet Experience››, May
2008, [http://tomgruber.org/writing/semthech08.pdf];
T. Gruber, ‹‹Collaborative Knowledge Management - Intraspect››,
May 2008,
[17] It is important to mention the support
of Daniel Pastor and the Instituto de Neurociencias y Derecho
(INEDE) have provided to us. The
Institute and its work team are innovators on the neuroscience sphere applied to
Criminal Law. In fact, they have created a software
that estimates delivery time on criminal matters prescriptions.
[18] During 2016, without considering the case
files associated with criminal matters, the Public Prosecutor’s Office on
Contentious Administrative and Tax Matters passed 912 Legal Opinions, as per Recurso de Queja (type
of appeal filed when a lower court judge improperly refuses to permit or delays
the filing of an appeal), Recurso de Inconstitucionalidad (appeal filled to request the
reversal of a judgment that violated the constitution) and ordinary appeals
entered to the Superior Court of Justice of the City of Buenos Aires.
[19] A chatbot is a conversation agent that interacts with users, at a determined domain or about a determined subject using natural language. See J.
Huang, et al, Extracting Chatbot
Knowledge
from Online Discussion Forums, International Joint Conference on Artificial
Intelligence, California, 2007, vol. 7,
pp.
423-428.
[20] The
digital platform is formed by the internet portals and mobile applications; process
guide; the citizen digital profile; call center services; public office for
in-person service; simple text
message services (SMS) and services offered through social media.
See Argentina National
Executive Power, National Public Digital Area Platform, Decree 87/2017,
section 1, February 2nd 2017; Argentina National Executive Power,
State Modernization Plan, Decree 434/2016,
section 1, March 1nd 2016.
[21] See
‹‹Asilomar AI Principles››, Future of
Life Institute, [https://futureoflife.org/ai-principles/].
[22] See J. Barrat, , Our Final
Invention, 2013, p. 92.
[23] See
Wisconsin Supreme Court, July 13th
2016, Wisconsin State vs. Eric L. Loomis, [https://www.wicourts.gov/sc/opinion/DisplayDocument.pdf?content=pdf&seqNo=171690].
[24] See Wisconsin Supreme
Court, July 13th
2016, Wisconsin State vs. Eric L. Loomis, whereas 17, 28, 34, 51, 93 and 94. [https://www.wicourts.gov/sc/opinion/Display Document.pdf?content=pdf&seqNo=171690].
[25] In the United States of America, the most widely used artificial
intelligence system that sets a risk score is COMPAS. This AI is based of risk
assessment in the field of insurance to determine the accident risk of a
particular person and what premium will correspond to pay. COMPAS provides a risk
scoring response on
a scale of 1 (low risk) to 10 (high
risk). That is, the score that artificial intelligence provides
is a comparison of how risky the individual is in relation to a segmented population. For example, if a score of 4
is obtained, then 60% of the population is seen as riskier than the subject analyzed, while 30% appear
less risky.
The scoring that corresponds to assign to an individual in
particular, evaluates and assigns weight to a series of criminogenic
factors (causes or concauses of criminality) that are present in
the subject in relation to a population with similar characteristics. For
example, if we use the COMPAS system to gain scoring, it would assign a
lot of weight to the age at which our subject committed its first offense, education level
and previous
history of noncompliance, among others. Thus, if the
person seeking parole was 25
years old and the first offense was committed at the age of 16, the risk score for reoccurrence would be of high risk; that is, located on the scale at a score higher than
or equal to 8 points. On the
contrary, the older the offender, the lower
the
score, even if he committed a felony. See more; Northpointe, Practitioners Guide to COMPAS.
[26] See J. Angwin,
et al. ‹‹Machine Bias: There’s Software Used Across the Country to
Predict Future Criminals. And it´s Biased Against Blacks.››,
ProPublica, May 2016 [https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing].
[27] The
developing company publicly replied that the ProPublica report presented technical mistakes. See
NORTHPOINTE, ‹‹Response to ProPublica: demonstrating
accuracy equality and predictive parity››,
[http://www.equivant.com/blog/response-to-propublica-demonstrating-accuracy-equity-and-%20predictive-parity].
[28] This is not leveled in all legal systems, for example, in the
Anglo Saxon system judgments passed by the guilt or non-guilt juries are taken
as innovators and their judgments are only appealable when a guilt statement
leads to a conviction.
[29] Also see: Section 52 of the Civil and Commercial Code
of Argentina “The human person injured in his or her personal or
family privacy, honor or reputation, image or identity, or that in any way impaired his personal dignity,
can claim the prevention and compensation for
the damages suffered (…)”.
[30] See L.
Ferrajoli, “About Fundamental Rights,
Theory of Neo-constitutionalism”, Trotta,
Madrid, 2007, pp.73-75.
[31] However,
as stated by Charles Beitz, it is not plausible to find a unique foundation or to create a list of rights.
See C.
R. Beitz,The Idea of Human Rights, Marcial Pons, Ediciones Juridicas y Sociales, Madrid, 2012, pp. 141-142 and 244.
[32] From the United
Nations Organization point of view: “(…) everyone shall take advantage from the benefits of
new
technologies , especially from information and
communication technologies, as per the suggestion made in the Economic and Social Council’s Ministerial Declaration 2000’’. See:
United Nations, 55/2 Millennium Declaration, section III, point 20, item five
[http://www.un.org/millennium/declaration/ares552e.htm].
[33] See L. H. Allende Rubino, The Preventive Action in the Civil and Commercial Code. The Relation with the
Preventive Measure Principle in the Environmental Right, MicroJuris online, 2016, quote
MJ-DOC-9989-AR and MJD9989.
[34] See S. Rodota,
The Right to Have Rights, Trotta,
Madrid, 2014, p. 182.
[35] See R. Pitchas,
Administrative Law of the
Information, Administrative Law Innovation and Reform, Global Law Press, 2ª ed., Sevilla, 2012, pp. 226-227 and 236. This author speaks of a paradigm shift in
administrative law of information, where each individual must be able to decide
under his own responsibility and autonomy between the possibilities and risks
generated by freedom of communication; R.
Pitchas, Administrative Law of the
Information, Administrative Law Innovation and Reform, Global Law Press, 2ª ed., Sevilla, 2012, p. 236.
[36] See ‹‹How to protect the commercial secrets of your PYME?›› World Intellectual Property Organization, WIPO
[http://www.wipo.int/sme/en/ip_business/trade_secrets/trade_secrets.htm].
[37] See J. Barrat, ,
Our Final Invention, 2013, pp.
240-241.
[38] See J. T. Palma Mendez,
et al, Artificial Intelligence,
McGraw-Hill Interamericana de España S.L., pp. 891-935.
[39] Definition
according to ISO 8402, complement of ISO 9000.
[40] The same access must be guaranteed in respect of any human or
legal person, public or private, linked to public
purposes or
public funds received, using such
technologies, provided that
the
design or use of artificial intelligence is related to
public purposes or to public funds received.
[41] For example, the UNESCO
encourages the creation of an enabling legislative environment in the field of
ICTs. In the same line, the UN argues that the development
of technologies, research and national innovation must be supported, guaranteeing a normative environment propitious to the
industrial diversification and the addition of value to
basic products.
In addition, States should
refrain from using information and communications
technology in contravention of
international law. See: United
Nations, General Assembly, Resolution
No. A/71/307, August 5th 2016, whereas 8,
p. 4; United Nations, General Assembly, Resolution
No. A/RES/70/1, October 21st
2015, whereas 9.b, p. 23; United Nations, General
Assembly,
Resolution No. A/RES/71/101 A-B, December 23rd
2016, whereas 4,
p. 4.4
[42] See S. Rodota, , The Right to Have Rights, Trotta, Madrid, 2014, p. 302.