Regulation of emerging technologies such as AI is partially controversial, because of the strive towards 'technology neutral' regulation. This paper summarizes the different approaches hidden behind the grand term of 'technology neutrality' to unravel its competing meanings. One of those meanings is then used for proposal of discriminatory approach towards deployment of AI in different services where society requires more trust and hence explanation. Regulatory barriers should be put forth to prohibit deployment of non-explainable AI into crucial services, such as medical diagnostics and triage. Paper argues that trust can be build by discrimination non-explainable machine learning models.
The term technology neutrality became widely used at the end of the 20th and the
beginning 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 [
This paper presents certain issues related to technology neutrality in different
contexts. 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
market innovation [
Koops dissected technology neutrality into four following distinct legislative goals [
The second goal aims to achieve functional equivalence between different technologies. 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 electronic 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 technology. The framework following this mode is forbidden from selecting technological winners at any given point in time and also over longer periods by subsidizing specific technologies. 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].
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
innovative 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
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 [
As is evident, technology neutrality contains different and often contradictory approaches. 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.
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 definition 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 towards 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.
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 reasoning 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
risks [
A transparent decision-making by technology is one of the precursors for
developing and maintaining trust ([
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
decisionmaking. As reported in [
The right to explanation (as a legal requirement for explainability) was heavily
debated over the course of drafting the new EU Data Protection Framework, mainly of the
General Data Protection Regulation. This attracted significant attention by legal
scholars debating the extent of obligation to provide the logic of automated decision-making
([
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 innovative 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
Laycock [
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
services, 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 [
Different areas of AI deployment described in [
Lipton [
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 hamper 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 infrastructure 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 typical for human decision-making. Regulating AI from perspective of technology neutrality 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.
Author gratefully acknowledges support by the Czech Science Foundation under grant no. GA17-20645S.