=Paper=
{{Paper
|id=Vol-2381/xaila2018_paper_3
|storemode=property
|title=Trust by Discrimination: Technology Specific Regulation & Explainable AI
|pdfUrl=https://ceur-ws.org/Vol-2381/xaila2018_paper_3.pdf
|volume=Vol-2381
|authors=Jakub Harašta
|dblpUrl=https://dblp.org/rec/conf/jurix/Harasta18
}}
==Trust by Discrimination: Technology Specific Regulation & Explainable AI==
https://ceur-ws.org/Vol-2381/xaila2018_paper_3.pdf
December 2018
Trust by Discrimination: Technology
Specific Regulation & Explainable AI
Jakub HARAŠTA a,1 ,
a Institute of Law and Technology, Faculty of Law, Masaryk University, Czech Republic
Abstract. Regulation of emerging technologies such as AI is partially controver-
sial, because of the strive towards ’technology neutral’ regulation. This paper sum-
marizes the different approaches hidden behind the grand term of ’technology neu-
trality’ to unravel its competing meanings. One of those meanings is then used for
proposal of discriminatory approach towards deployment of AI in different ser-
vices where society requires more trust and hence explanation. Regulatory barriers
should be put forth to prohibit deployment of non-explainable AI into crucial ser-
vices, such as medical diagnostics and triage. Paper argues that trust can be build
by discrimination non-explainable machine learning models.
Keywords. explainable AI, technology neutrality, discrimination of technology,
technology specific regulation, critical services
1. Introduction
The term technology neutrality became widely used at the end of the 20th and the be-
ginning of the 21st century. It was meant to describe the desirable mode of regulation of
emerging technologies, mainly in ICT. However, as stated by Reed [15, p. 265] the term
was used without deep discussion about its meaning and about specific regulatory tools
related to its use. Only in 2006, Koops deconstructed this policy one-liner to describe
its hidden complexity of often mutually exclusive regulatory approaches and competing
values [9].
This paper presents certain issues related to technology neutrality in different con-
texts. I argue for a technology specific approach to technology neutrality ([9, p. 85] and
[7, p. 247]) as a way to stimulate the wide use of explainable AI in society. I approach
this issue from the perspective of regulating AI as social innovation, and not purely mar-
ket innovation [20]. Therefore, I consider it axiomatic that in some cases innovation does
not present a value per se and requires regulation to be valuable for society. However, to
prevent hampering innovation throughout the whole field of AI, I suggest distinguishing
different AI uses. The use in crucial services, such as diagnostic or medical triage tools,
or the use on vulnerable groups, such as minors or ethnic minorities, should be regulated
towards explainability.
1 E-mail: jakub.harasta@law.muni.cz.
December 2018
2. Technology Neutrality
2.1. Different Faces of Technology Neutrality
Koops dissected technology neutrality into four following distinct legislative goals [9].
The first goal aims to regulate actions of users and consequences of these actions. This
happens regardless of the technology used to mediate these actions. This form of tech-
nology neutrality is rather extreme, as it focuses on lowest common denominator and
does not distinguish between different technologies. As an example, the purpose of both
a hand-written and an electronic signature is the expression of will or identification of an
individual. Following this aim, technology neutrality does not focus on technology, but
on the act of the expression of will.
The second goal aims to achieve functional equivalence between different technolo-
gies. This type of technology neutral regulation contains plethora of technology-specific
norms. These norms are used to compensate for differences between technologies in
terms of their use, the effect of such use or related costs. To follow the example set above,
this type of technology neutrality would set up specific norms for hand-written and elec-
tronic signatures to compensate for reasonable differences between those two means of
expressing one’s will.
The third possible aim of technology neutrality is the non-discrimination of technol-
ogy. The framework following this mode is forbidden from selecting technological win-
ners at any given point in time and also over longer periods by subsidizing specific tech-
nologies. The regulation following this maxim has strong competition and innovation
aspects. It serves the purpose of opening any given field for further innovations and aims
to lower the barriers to entry. To follow the example set above, this aim of technology
neutrality prevents us from creating legal framework requiring hand-written signatures
and not recognizing legal validity of electronic signatures.
The fourth possible aim of technology neutrality aims to create a flexible framework
accounting for future changes and accommodating future innovations. This regulation is
formulated in general terms and is often accompanied by open textured formulations and
flexible tests.
Very broadly put, technology neutrality serves as an aspiration to enact laws that
could be sustainable over time and would not require frequent reviews. Similarly, it can
be interpreted as a duty not to make technological choices by creating a restrictive legal
framework, but contrarily to leave these choices to market actors. Framework legislation
or general soft-law guidelines should stay clear of technological concepts, relying on
functional or economic ones instead. On the other hand, a technology-specific reasoning
should be implemented at the lowest possible level, e.g. the level of individual decisions
of regulatory authorities [1, p. 75–76].
2.2. Regulating Innovation
The general approach towards technology seems to be to avoid regulation out of fear of
failing behind countries with less strict regulation or no regulation at all. Zarsky, among
others, recalled an anecdote claiming that data protection framework hampers EU in-
novative potential, as the connection between the strength of privacy laws and the level
of ICT innovation is often evident [23, p. 154-155]. However, Zarsky goes beyond this
December 2018
anecdote and brings forth Stewart’s distinction between market innovation and social
innovation [20, p. 1277-1279]. The market innovation allows firms to offer new and/or
improved products to customers [20, p. 1279], while the social innovation leads to social
gains beyond a pure market-oriented approach. Stewart invokes technology leading to
cleaner air [20, p. 1279], while Zarsky applies this to actors offering stronger protection
of privacy [23, p. 127]. This is, in my opinion, the line between different approaches to
technology neutrality as captured by Koops [9]. The first, the third and the fourth aim
as explained in subsection 2.1 seem to aim primarily towards market innovation. The
second, functional equivalence, seems to be concerned with social innovation. The claim
that any regulation will hamper the innovative potential of AI is, without any doubts,
correct. However, in the domain of regulating AI as a social innovation, it carries little
value for further discussion.
As is evident, technology neutrality contains different and often contradictory ap-
proaches. A specific regulatory approach towards technology neutrality is based on a
specific context. Arguing for technology neutrality without providing such context could
lead to confusion in the creation, application and interpretation of legal framework. It is
possible to stimulate the use of explainable AI by implementing the technology neutral
legislation, but only if such legislation follows the second aim - functional equivalence.
Human and algorithmic decision-making carries certain differences and these differences
should make it into legal framework of the future use of AI, through the requirement of
explainable AI.
3. Regulation of Artificial Intelligence
I will stay away from trying to define what Artificial Intelligence is. First of all I do not
feel competent enough to even try, and second of all, there is an apt pragmatic defini-
tion available, which will serve my purpose: ”[AI is] the science of making machines
capable of performing tasks that would require intelligence if done by [humans].” [11,
p. V]. Essentially, AI aims towards replacing humans in certain activities. Or at least to-
wards mimicking them and perform certain tasks either well or well enough for it to be
economically viable under specific circumstances.
If we are to put forth a legislation that would push AI towards explainability, there
are two main concerns. First, how to maximize the social innovation potential of AI
through legislation. Second, how to minimize the impact this will inevitably have on AI
as a market innovation.
3.1. Explainable AI for Achieving and Maintaining Trust
When it comes to decision-making and reasoning, the blackest box of all is human mind.
It is theoretically possible to apply technology neutrality stricto sensu (the first aim as
described in Subsection 2.1) and require no more from AI than we already require from
humans. We have settled for decision-making based partly on intuition and poor reason-
ing skills. However, in order to build trust in the AI decision-making, we require different
standards. We demand an explanation.
Trust in technology is determined by multiple factors - human characteristics, such
as the user’s personality and ability to understand the technology and deal with related
December 2018
risks [19]; environment characteristics, such as culture (including socioeconomic status),
the task for which the technology is used, and institutional factors including existing
regulation and its enforcement [13]; and technology characteristics, such as performance
of the technology, transparency of its process and purpose of its use [17].
A transparent decision-making by technology is one of the precursors for develop-
ing and maintaining trust ([14],[4],[21],[16]). It addresses human characteristics, namely
the ability to understand technology. For general population, AI is currently too opaque
not only in its decision-making, but also in its general functioning. Requiring trans-
parency also directly promotes certain technology characteristics by pushing specific
(non-transparent) designs off the market. Also, a legal framework for AI or its use would
affect institutional factors of trust building, because it would send a message that legis-
lators understand the technology well enough to require certain changes in its design.
At this point in time, different algorithms are used for machine learning purposes.
Machine learning classifiers include the use of decision trees or decision lists, neural
networks, k-nearest neighbour algorithms, support vector machines, Bayesian networks
etc. Usually, we tend to think about artificial intelligence in terms of a trade-off between
accurate and black-box models, and inaccurate and white-box models. Lipton argues
against this belief by stating that linear models are per se no more interpretable than
deep neural networks [10, p. 7]. According to Lipton, the issue of explainable AI is
not predominantly concerned with the use of different machine learning models, but
rather with the issue of implementation of these models. Additionally, the discussion
about interpretability of specific models is made difficult by missing a clear definition of
interpretability [10, p. 7-8].
Without aiming for proper definition of interpretability, it is important to note one
of the basic motivations. The explainable AI allows us to verify the results of decision-
making. As reported in [2], due to bias in training data, AI can achieve incorrect results.
In this specific case, the model learned that patients with asthma and heart problems
have lower risk of dying of pneumonia. This result is directly opposite to what current
medicine knows about pneumonia. This particular bias leading to wrong classification
was easy to spot once the model was evaluated by medically trained personnel. However,
in a lot of areas, decision-making is currently not as linear as the relation between heart
disease, asthma, and pneumonia. Taking a lot of variables in account does not necessarily
lead to better results. We tend to be rather surprised when complex models get outper-
formed by simple linear predictors [3]. However, in my opinion, a lot of variables should
mean stronger pressure towards explainability regardless of model performance.
The right to explanation (as a legal requirement for explainability) was heavily de-
bated over the course of drafting the new EU Data Protection Framework, mainly of the
General Data Protection Regulation. This attracted significant attention by legal schol-
ars debating the extent of obligation to provide the logic of automated decision-making
([18],[22],[8]). Formulation of specific requirements for the explainable AI in legislation
specifically targeted to use of AI in society would without any doubts lead to a signifi-
cant opposition. However, at the same time, requiring specific design – in this case the
explainable design – of AI would arguably lead to greater trust in automated decision-
making.
December 2018
3.2. Minimizing Negative Effect on Innovation
Directly resulting from the conclusion to require explainable AI by law in Subsection 3.1
is the issue of potential AI winter.
We surely desire explainable AI. However, if all AI is to be explainable before it
could be implemented, we might be facing another AI winter, at least in countries with
strict (or any) regulation. Unregulated development and deployment means more innova-
tive potential, however at cost of social benefits, such as dissolution of trust. On the other
hand, strictly regulated development and deployment means less success in AI research
compared to other countries, e.g. China. But we cannot possibly have a pie and eat it too.
However, I believe there is a partial solution to this conundrum. Harwich and Lay-
cock [6] laid out certain issues in health industry where AI is beneficial or could be ben-
eficial in the future. These issues are wide-ranging. AI could be useful in boosting gen-
eral well-being of population through health promotion or targeted prevention. It could
be used to augment cognitive capacities of medical professionals and provide us with
improved and fast diagnostics and triage. Additionally, AI could make our health sys-
tem more efficient overall by taking away some of the administrative burden through au-
tomation, or by allowing for better management and self-care for chronically ill patients.
These implementations require different levels of duty of care as they carry with them
different consequences in case of a mistake. Some areas, such as diagnostics, should
be made more transparent, even if only for the issue reported in [2]. Diagnostics is in-
herently more complex than automation of repetitive administrative tasks or reminding
to patients with cognitive impairment to take their medicine. The crucial services sup-
ported by AI decision-making require more trust and this can be achieved by requiring
transparency. Other areas can be deemed as less important or less crucial.
This sort of distinction among various services for purpose of legal regulation has a
precedent. Critical infrastructure protection currently works along the same line. Some
sectors, such as energy, water management, food industry and agriculture, health ser-
vices, financial market etc., are deemed as critical and as a result are regulated differently.
Different countries approach these issues in multiple ways, labelling different sectors as
critical. However, labelling specific sectors as critical attracts an increased attention in
terms of specific applicable policy and legal framework [5].
Different areas of AI deployment described in [6] require different levels of duty
of care. Mistakes lead to different consequences, from administrative nuisance to life-
threatening situations. Any regulation should, in this regard, prescribe the use of explain-
able AI in some areas that we deem crucial for society – either within the scope of the
current critical infrastructure protection framework or outside of it. This approach allows
us to maximize the benefit of AI as social innovation (because we opt for regulation)
without stopping AI as market innovation dead in its tracks (because we opt for targeted
regulation following risk analysis).
4. Conclusion
Lipton [10] formulated different levels of explainable AI. He noted the transparency at
the level of entire model (simulatability), transparency at the level of individual com-
ponents (decomposability), and at the level of training algorithm (algorithmic trans-
parency).
December 2018
I argue that different levels of explainability should be required from AI deployed
into different areas, different services or different individual decision-making tasks. A
legal requirement for explainable AI design could help us build trust in automated
decision-making. However, requiring explainable AI everywhere could seriously ham-
per the innovative potential of AI. There has to be a regulation requiring explainable AI
models to ensure that we develop and maintain trust. However, the two-tier classification
of services ensures proportionality of such regulation and prevents unreasonable barriers
to block further development of AI. Similarly to the existing framework of critical infras-
tructure protection, we require a different level of protection once the asset is considered
a critical infrastructure. Discriminating between explainable and non-explainable AI in
access to those areas makes perfect sense.
Approaching AI in a technology neutral way – or more precisely in its first, third
and fourth aim as described in Subsection 2.1 – means regulating it as market innovation.
This approach is related to the perspective of law and economics, which aims to regulate
market failures. Innovation and welfare from constantly innovating services is of utmost
importance and considered critical for modern society.
However, AI mimics human intelligence and is increasingly deployed into areas typ-
ical for human decision-making. Regulating AI from perspective of technology neutral-
ity seeking functional equivalence adds reasonable regulatory burden to new technology.
This aims to protect the values and tame the market-driven innovation. This approach
targets AI as social innovation.
What I described in this paper should serve as a compromise between these two
forces and two regulatory approaches. We should strive to tame the flame of innovation,
so it would serve us. But extinguishing it completely would leave us in the dark.
Acknowledgment
Author gratefully acknowledges support by the Czech Science Foundation under grant
no. GA17-20645S.
References
[1] A. Butenko, P. Latouche. Regulation for Inovativeness or Regulation of Innovations. Law, Innovation
and Technology, 2015, vol. 7, no. 1, pp. 52–82.
[2] R. Caruana, Y. Lou, J. Gehrke, P. Koch, M. Sturm, N. Elhadad. Intelligible Models for HealthCare:
Predicting Pneumonia Risk and Hospital 30-day Readmission. Proceedings of the 21st ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, 2015, pp. 1721–1730.
[3] J. Dressel, H. Farid. The Accuracy, Fairness, and Limits of Predicting Recidivism. Science Advances,
2018, vol. 4, no. 1, eaao5580.
[4] A. Glass, D.L. McGuiness, M. Wolverton. Toward Establishing Trust in Adaptive Agents. Proceedings
of the 13th International Conference on Intelligent User Interfaces IUI, 2008, pp. 227–236.
[5] J. Harasta. Legally Critical: Defining Critical Infrastructure in an Interconnected World. International
Journal of Critical Infrastructure Protection, 2018, vol. 21, s. 47–56.
[6] E. Harwich, K. Laycock. Thinking on its own: AI in the NHS. Reform, 2018.
[7] M. Hildebrandt. Legal Protection by Design: Objections and Refutations. Legisprudence, 2011, vol. 5,
no. 2, pp. 223–248.
[8] M. Kaminski. The Right to Explanation, Explained. 2018. SSRNID 3196985.
December 2018
[9] B.-J. Koops. Should ICT Regulation be Technology Neutral? In: B.-J. Koops, M. Lips, C. Prins, M.
Schellekens (eds.). Starting Points for ICT Regulation. Deconstructing Prevalent Policy One-Liners. The
Hague: T.M.C. Asser Press, 2006. Pp. 77–108.
[10] Z.C. Lipton. The Mythos of Model Intepretability. 2017. arXiv:1606.03490.
[11] M. Minsky. Semantic Information Processing, Cambridge: MIT, 1968.
[12] P. Ohm. The Argument against Technology-Neutral Surveillance Laws. Texas Law Review, 2010, vol.
88, no. 7, pp. 1685–1714.
[13] K. Oleson, D. Bilings, V. Kocsis, J.Y. Chen, P.A. Hancock. Antecedents of Trust in Human-Robot Col-
laborations. Proceedings of IEEE International Multi-Disciplinary Conference on Cognitive Methods in
Situation Awareness and Decision Support CogSIMA, 2011, pp. 175–178.
[14] P. Pu, L. Chen. Trust building with explanation interfaces. Proceedings of the 11th International Con-
ference on Intelligent User Interfaces IUI, 2006, 93–100.
[15] C. Reed. Taking Sides on Technology Neutrality. SCRIPTed: A Journal of Law, Technology and Society,
2007, vol. 4, no. 3, pp. 263–284.
[16] M.T. Ribeiro, S. Singh, C. Guestrin. ”Why Should I Trust You?” Explaining the Predictions of Any
Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, 2016, pp. 1135-1144.
[17] K.E. Schaefer, J.Y. Chen, J.L. Szalma, P.A. Hancock. A Meta-Analysis of Factors Influencing the Devel-
opment of Trust in Automation: Implications for Understanding Autonomy in Future Systems. Human
Factors, 2016, vol. 58, no. 4, pp. 377–400.
[18] A. Selbst, J. Powles. Meaningful Information and the Right to Explanation. International Data Privacy
Law, 2017, vol. 7, no. 4, pp. 233–242.
[19] K. Siau, Z. Shen. Building Customer Trust in Mobile Commerce. Communications of the ACM, 2003,
vol. 46, no. 4, pp. 91–94.
[20] R. Stewart. Regulation, Innovation and Administrative Law: A Conceptual Framework. California Law
Review, 1981, vol. 69, no. 5, pp. 1256–1377.
[21] N. Tintarev, J. Masthoff. Designing and Evaluating Explanations for Recommender Systems. Recom-
mender Systems Handbook, 2010, pp. 479–510.
[22] S. Wachter, B. Mittelstadt, L. Floridi. Why a Right to Explanation of Automated Decision-Making Does
Not Exist in the General Data Protection Regulation. International Data Privacy Law, 2017, vol. 7, no.
2, pp. 76–99.
[23] T. Zarsky. The Privacy-Innovation Conundrum. Lewis & Clark Law Review, 2015, vol. 19, no. 1, pp.
115- -168.