A study into the layers of automated decision-making: emergent normative and legal aspects of deep learning

Abstract

The paper dissects the intricacies of Automated Decision Making (ADM) and urges for refining the current legal definition of AI when pinpointing the role of algorithms in the advent of ubiquitous computing, data analytics and deep learning. ADM relies upon a plethora of algorithmic approaches and has already found a wide range of applications in marketing automation, social networks, computational neuroscience, robotics, and other fields. Whilst coming up with a toolkit to measure algorithmic determination in automated/semi-automated tasks might be proven to be a tedious task for the legislator, our main aim here is to explain how a thorough understanding of the layers of ADM could be a first good step towards this direction: AI operates on a formula based on several degrees of automation employed in the interaction between the programmer, the user, and the algorithm; this can take various shapes and thus yield different answers to key issues regarding agency. The paper offers a fresh look at the concept of "Machine Intelligence", which exposes certain vulnerabilities in its current legal interpretation. To highlight this argument, analysis proceeds in two parts: Part 1 strives to provide a taxonomy of the various levels of automation that reflects distinct degrees of Human -Machine interaction and can thus serve as a point of reference for outlining distinct rights and obligations of the programmer and the consumer: driverless cars are used as a case study to explore the several layers of human and machine interaction. These different degrees of automation reflect various levels of complexities in the underlying algorithms, and pose very interesting questions in terms of regulating the algorithms that undertake dynamic driving tasks. Part 2 further discusses the intricate nature of the underlying algorithms and artificial neural networks (ANN) that implement them and considers how one can interpret and utilize observed patterns in acquired data. Finally, the paper explores the scope for user empowerment and data transparency and discusses attendant legal challenges posed by these recent technological developments.

De Sophia à inteligência artificial geral: Antropomorfismo e consciência no estado da arte atual da tecnologia

“…Technological advances in robotics and cognitive science have opened the way for even more sophisticated AI systems, capable of acting completely autonomously, to the point of mimicking human behavior in unprecedented ways, and making interactions between algorithms and humans very difficult to predict -racist comments and decisions from AI systems are good examples of this, as their early programmers did not intend to give them prejudiced biases, which were acquired after machine learning (Karanasiou;Pinotsis, 2017). This forces Law to inevitably develop a new concept of personhood to address the behavior of human-like agents.…”

Ethics-Based Auditing of Automated Decision-Making Systems: Nature, Scope, and Limitations

“…autonomous self-learning systems that gather and process data to make qualitative judgements with little or no human intervention, increasingly permeate all aspects of society (AlgorithmWatch, 2019). This means that many decisions with significant implications for people and their environments-which were previously made by human experts-are now made by ADMS (Karanasiou & Pinotsis, 2017;Krafft et al, 2020;Zarsky, 2016). Examples of the use of ADMS by both governments and private entities include potentially sensitive areas like medical diagnostics (Grote & Berens, 2020), recruitment (Sánchez-Monedero et al, 2020), driving autonomous vehicles (Evans et al, 2020), and the issuing of loans and credit cards (Aggarwal et al, 2019;Lee et al, 2020).…”

“…Yet others focus on the context in which ADMS operate. For example, so-called Human-in-the-Loop protocols imply that human operators can either intervene to prevent or be held responsible for harmful system outputs (Jotterand & Bosco, 2020 ; Rahwan, 2018 ).…”

Towards Transparency by Design for Artificial Intelligence

“…In this article, we are interested in the delegation of a decision-making process to an algorithm, i.e., automated decision-making (AlgorithmWatch 2019 ). We understand automated decision-making (ADM) as a subpart of AI, an automated process with no human involvement to reach a decision (Karanasiou and Pinotsis 2017 ; ICO 2020 ). We start with the premise that automated decision-making algorithms “make generally reliable (but subjective and not necessarily correct) decisions based upon complex rules that challenge or confound human capacities for action and comprehension” (Mittelstadt et al 2016 , p. 3).…”