IEEE Technology and Society Magazine - June 2021 - 63

* Assessment list and use cases: Based on the ethical
purpose and the implementation methods,
the framework sets out a preliminary and nonexhaustive
assessment list for AI developers, deployers,
and users to operationalize trustworthy AI.
Given the application specificity of AI, the assessment
list will need to be tailored to specific applications,
contexts, or sectors. We selected number
of use cases to provide an example.
To this moment, only a few examples of regulation
of algorithmic justice can be mentioned. One
is the Basic Law on Robots (South Korea) of 2019,
and the proposed 2019 Algorithmic Accountability
Act (U.S.) which seeks to regulate bias in automatic
decision-making systems. It does not specify a definition
of fairness, allowing for flexibility. It would
require companies to audit their ML for bias and discrimination
in an " impact assessment. "
Independent authority and auditing
infrastructure to enforce
The White Paper mentions that the regulatory framework
should only be applied to high-risk AI applications.
We contend this position, since regulation could
affect all applications-all applications should be safe,
protect privacy, and be transparent to a reasonable
degree-but high-risk applications should be subject
to greater and more specific and frequent (or continuous)
controls like inspections and audits, while no-risk
applications should still follow regulation, but without
third-party auditing or inspection.
The approach that seems to inspire the White
Paper is a process auditing approach, where companies
are required to prove that they have followed
appropriate processes in selecting the training data,
keeping records, provision of information, adequate
measurement of accuracy, and design robustness, in
line with common auditing practices (see [27]). Such
a process would very seldom prove specific safety,
accuracy, and robustness of the system, but would
only prove that the company followed a required
design and implementation practice. To provide real
" trustworthiness " to final consumers and stakeholders
for a " high-risk " application, an independent,
third-party inspection and test should be required.
Since training of the algorithms can be concurrent
with production usage of the AI models, a continuous
testing and benchmark infrastructure should be
used for these types of high-risk applications. Several
June 2021
such oversight mechanisms have been proposed by
international agencies, like, for instance, the French
CNIL, Commission Nationale de l'Informatique et
des Liberte (French data protection authority), in
[28].
In our view, a specific AI certification agency
and associated technical infrastructure need to be
developed to ensure " trustworthiness. " This would
also require a detailed technical framework for
AI and the development of a standard set of data
exchange format requirements, so the continuous
auditing can be performed in an efficient way and
also providing the necessary security for companies
and individuals.
This arTicle explOres the approach to ethical
and societal implications of AI present in the
European Commission White Paper on Artificial
Intelligence. Without this analysis, AI would not
contribute to SDGs, as the social impact is crucial in
the 2030 Agenda. While the European Commission
article focuses mainly on risk and some individual
rights (such as privacy), the collective dimension
of society as a whole is overlooked. This is the first
contribution of this article. A human-rights-based
approach should be followed in the regulatory
efforts as it is necessary to establish a universal
governance model and a general normative framework
for AI. Human rights should replace ethics as
the dominant framework for debate. A description
of the main principles of the rights approach is
offered. The authors add some principles to the list
in the White Paper: contestability and auditability
are the more important ones. There is a need to
develop a sound technical framework within the
regulation. Any regulation on a technical matter
should encompass an architectural model on how
the overall system functions and interacts. This is
a key point not sufficiently stressed in the literature.
An auditing system through an independent
authority is also required to allow accountability in
the algorithmic process. Finally, another important
issue should be to get more transparency in the
legislative process. Only big lobbies are closed to
the legislative projects in the EU, but the general
public needs access to them to study and propose
new perspectives which allow a more just and fair
AI. The human-rights-based approach requires participation
of citizens in the process of regulation as
one of its main principles.

63
IEEE Technology and Society Magazine - June 2021

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - June 2021
