AI-'Agents': to be, or not to be, in the legal domain

Author:Francesco Cavinato - Federica Casano
Position:Alma Mater Studiorum, University of Bologna School of Law (Italy), JD LL.M Candidate - Alma Mater Studiorum, University of Bologna School of Law (Italy), JD LL.M Candidate; Tilburg University School of Law (The Netherlands), LL.M. Candidate
Dundee Student Law Review, Vol. 5(1+2), No.2
AI-“Agents”: to be, or not to be, in the legal domain
Francesco Cavinatoa Federica Casano,b
AIE and Law
Recent technological developments have led to an algorithmic society where artificial
intelligence entities1 (hereinafter, AIEs) have moved into different fields of the human society2 such
as medicine, engineering and economy. Legal disciplines have been involved as well, especially in
private law, where they have obtained such a special interest in autonomously executing the
bargaining, formation and the performance of contracts. However, human users often have no
knowledge of the exact terms of the contract, or even that a contract is being made.3
† An earlier version of this paper was awarded the third place at the “Artificial Intelligence Legal Issues International
Paper Competition” promoted and organized by the William and Mary Law School - Center of Legal and Court
Technology (VA, USA) and publ ished on its website at -AI-
Agents_to-be-or-not-to-be-in-legal-domain.pdf. It was also presented at the last Critical Law Society Conference
“Metamorphosis and Law” (University of Kent, Canterbury, UK). An update has been due to recent developments in the
EU system and legal scholarships.
a Alma Mater Studiorum University of Bologna School of Law (Italy), JD LL.M Candidate
b Alma Mater Studiorum University of Bologna School of Law (Italy), JD LL.M Candidate; Tilburg University School
of Law (The Netherlands), LL.M. Candidate
1 For a general perspective on AI, see Ryan Calo, ‘Artificial In telligence Policy: A Primer an d Roadmap’ (2018) 3(2)
University of Bologna Law Review 180. At the first stage of our paper, we will use the terms artificial intelligence and
autonomous systems interchangeably, although they have different scopes and meanings. Indeed, for this current purpose,
the technical nuances are largely irrelevant. The alternative sy nonym “autonomous technologies” deals with classical
Artificial Intelligence, Machine Learning algorithms, Deep Learning and connectionist networks, generative adversarial
networks, mechatronics and robotics. However, we want to suggest for a deeper insight on autonomous systems- Thomas
Burri, ‘The Politics of Robot Autonomy’, (2016) 7(2) European Journal of Risk Reg ulation, 341.
2 Even though “Some leading technologists and futurists in Silicon Valley recently named artificial intelligence an
existential threat to humanity and called for answers to the ethical and legal questions it raises”- Thomas Burri, ‘Free
Movement of Algorithms: Artificially In telligent Persons Conquer the European Union’s Internal Market’ , in Woodrow
Barfield and Ugo Pagallo (eds), Research Handbook on the Law of Artificial Intelligence, (Edward Elgar, 2018), 538.
3 Emad Abdel Rahim Dahiyat, ‘Intelligent Agents and Contracts: Is a Conceptual Rethink Imperative?’ (2007) 15 (4)
Artificial Intelligence and Law, 375.
AIEs are both hardware and software entities that perform tasks in ways that are “intelligent”:4
indeed, they are not just programmed for a single and repetitive motion but they can adapt to do more
(and in a better way) by adapting to different situations and contexts. Nevertheless, they are able to
understand languages, recognize pictures, solve complex problems by themselves and learn5 as they
go along6 without constant supervision (e.g. machine learning7). Their decision-making process is
usually based on symbolic reasoning, analyses of the user’s behaviour, experience, data acquisition
and it is characterized by a heuristic8 approach.9 Differently from the human mind which depends on
the Johnson-Laird’s mental model theory10 and often falls into biases, AIEs can not only remove the
deductive fallacies in reasoning but they can increase the decision-making process through the
conditional probability (Bayes’ theorem). Furthermore, they are able to set up an inductive reasoning
reducing both decisional conflicts and cognitive dissonances.
4 Besides th e different attempts to define intelligence across psychological branches (cognitive, behaviourist, dynamic,
Piagetian), as well as its intuitive conceptualization by the common sense ,“… if we attempt to dig deeper and define it
in precise terms we find the concept to be very difficult to nail down… Intelligence involves a perplexing mixture of
concepts, many of which are equally difficult to define.” Shane Legg and Marcus Hutter, ‘Universal Intelligence: A
Definition of Machine Intelligence’ (2007) 17 Minds and Machines 391. An interesting approach to intelligen ce which
allow to get closer to AI structure (especially neural networks) is the “symbol system” approach, that is “…the ability of
human beings to use various symbolic vehicles in expressing and communicating meanings distinguishing human beings
sharply from other organisms.”- see Howard Gardner, F rames of Mind: The Theory of Multiple Inte lligences (3rd edn,
Basic Books 2011), 26.
5 Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (3rd edn, Pearson, 2016), 1 -5, 36-40, 64-
69, 693-850.
6 Namely, they “…will collect info rmation without an express instruction to do so, select information from the universe
of available data without direction, make calculations without being told to do so, make recommendations without being
asked and implement decisions without f urther authorization… [they] will truly execute their decisions with real data in
a complex networked environment, and will af fect real world events”- Curtis E.A. Karnow, Liability for Distributed
Artificial Intelligence (1996) 11(1) Berkeley Technology and Law Journal 147, 152-153.
7 Especially this type of AIEs are dissimilar from traditional analytics: indeed, they modify the underlying constitutive
algorithm according to data which they have previously processed. As output, they learn new schem es of information.
8 A complex and innovative mix of “strategies using readily accessible information to control problem-solving processes
in man and machine” by using approximate algorithms- Judea Pearl, Heuristics: Intelligent Search Strateg ies for
Computer P roblem Solving (Addison-Wesley, 1984) 3. Regarding the AI context, heuristic is a function that (based on
trade-off criteria such as optimality, completeness, accuracy and time) ranks alternatives in search algorithms at each
branching step based on available information to decide which branch to follow. In these terms, the “heuristic search
remains as a core area of artificial intelligence. The use of a good search algorithm is often a critical factor in the
performance of an intelligent system. As with most areas of AI, there has been steady progress in heuristic search research
over the years. This progress can be measured by several different yardsticks, including nding optimal solutions to larger
problems, making higher quality decisions in fixed size p roblems, handling more complex domain s including dynamic
environments with incomplete and uncertain information, being able to analyze and predict the p erformance of heuristic
search alg orithms, and the increasing deployment of real-world applications of search algorithms”- Weixiong Zhang,
Rina Dechter and Richard E. Korf, ‘Heuristic Search in Artificial Intelligence’ (20 01) 129 Artificial Intelligence 1.
9 Fabio Bravo, Contra ttazione Telematica e Contra ttazione Cibernetica (Giuffrè Editore , 2007), 196-209.
10 Philip Johnson-Laird, Mental models and human reasoning.(2010) 107(43) P roceedings of the National Academy of
Sciences 18243.

To continue reading

Request your trial