This presentation discusses the hard legal questions raised by the research, development and implementation of Articificial Intelligence technologies. It discusses ethics, ex post v ex ante regulation, data biases, and the legal framework for various categories social harms.

1. HARD QUESTIONS OF LAW AND REGULATION OF
ARTIFICIAL INTELLIGENCE (AI)
Future Law – Budapest, 27 September 2018
Prof Nicolas Petit, ULiege (BE) and UniSA (AU)

2. www.lcii.eu
 EU Commission High Level
Expert Group on AI
 Communication on AI
 HLG: 2 subgroups
 One tasked with drafting of
ethical guidelines on AI
 The other tasked with
recommendations for policy
an investment in AI
 Through mid 2019 + EU AI
Alliance:
https://ec.europa.eu/futuriu
m/en/eu-ai-alliance
Appetite for regulation?
 Parliament resolution, 16 Feb
2017, Civil Law Rules on
Robotics
 “Shortcomings” of legal framework
 “questions whether the ordinary rules
on liability are sufficient or whether
it calls for new principles and rules
to provide clarity on the legal
liability”
 Supports the “designation of a
European Agency for Robotics and
Artificial Intelligence in order to
provide the technical, ethical and
regulatory expertise needed to support
the relevant public actors”

3. www.lcii.eu
Hard questions of law & AI
1. AI in a nutshell
2. Ex ante v ex post regulation?
3. Normative issues: illustration from ALE
4. Ethics (before law): selected issues
5. Proposed framework: who, when and how to
regulate?

4. www.lcii.eu
1. AI in a nutshell
Core ideas
 Any physical process
including the mind process
can be modelized as a
computable algorithm
(Church Turing thesis)
 Machines can learn:
“Learning is any process by
which a system improves
performance from
experience” (Herbert
Simon)
Today
 Brute force computational
power now available (due to
Moore’s law)
 Zettabytes (or Yottabytes) of
data now available, and
distributed, and pre labelled
(cloud)
 « End of theory »
 Deep learning, neural networks, etc.
 Use cases: autonomous
vehicles, predictive justice,
predictive medicine,
automated law enforcement

5. www.lcii.eu
 Reactive
 Mostly common law process
 By courts
 Standards (content of law
unspecified ex ante)
 Preemptive
 Mostly civil law process
 By technocratic or
democratic
 By expert or non expert
organs
 Rules (content of law
specified ex ante)
2. Ex ante v ex post regulation of AI?
 « Collingridge paradox »
 Too early too late
quandary?

6. www.lcii.eu
Paradoxes of ex ante regulation
Disciplinary approach
 Lawyers need frictional case studies
to apply deductive reasoning
 One problem:
 End up looking at sci-fi to generate
case studies (EP, 2017)
 Or end-up looking at the past to
generate case studies
 Techno-speculation: driverless cars
not the dominant hypo in 1970s to
2000s Sci-Fi, but flying cars
 Human-speculation:
teleconferencing v conferencing
 Risk of irrelevant law
Technological approach
 Emergent behavior hypothesis of
AIs
 Inevitable assumption that there
are gaps in the legal system. New
law is needed. Law of the horse.
 Two problems:
 Specificity of AI and robots
emergent conduct is to emulate
human level intelligence (Minsky,
1984) + emergent behavior
relates to human conduct =>
existing law OK
 Tendency “to attribute programs far
more intelligence than they actually
possess” (Yazdani, 1986)
 Risk of redundant or extreme law

7. www.lcii.eu
 Can the law ever be agile?
 5 to 10 years to regulate
tech that’s already 5 to 10
years old
Paradoxes of ex post regulation
 « Treacherous turn »
(Bostrom)
 Dominant design problem

8. www.lcii.eu
3. Normative issues
 AI generates hard normative problems
 « Should » questions

9. www.lcii.eu
Case study: Automated Law Enforcement
 “any computer-based system that
uses input from unattended sensors
to algorithmically determine that a
crime has been, or is about to be,
committed and then takes some
responsive action, such as to
prevent the crime to inform the
appropriate law-enforcement
agency or to impose some form of
punishment” (Shay et al.)
 Surveillance, analysis, aggregation
and action
 Police (telemetric policing) but
other applications
 UK: Connect (tax); US: Loomis
case

10. www.lcii.eu
Should « full enforcement » be the goal?
 «Full enforcement » means perfect surveillance and punishment
of infringements + perfect pre-emption
 Legal: undercover criminality by law enforcement; civil
disobedience; whistleblowing
 Logical: paradox of disappearance of crime, and relaxation of
social consensus against crime
 Economics: imperfect enforcement is useful (Stigler, 1970)
 Sociological: acceptability of technology is driven by context. Full
enforcement more desirable in countries with endemic crime and
corruption (Honduras v Greece)
 Psychological: offender wishes to “to 'have his say' on the matter”

11. www.lcii.eu
 Human dignity
 Transparency and
informed consent
 Democracy
 Autonomy and self-
determination
 Safety and Cybersecurity
 Transparency, legibility
and explainability
 Non bias
 Responsibility and
accountability
 Compliance with HRs
 Red lines? But red rights?
4. Ethics
 Artificial ethics:
hardwired by design in
robots
 Natural ethics: rules
that apply to
researchers

12. www.lcii.eu
Ongoing developments
 Private ordering
 Standardization: data practices certification
 Unilateral products:
 IBM just launched a software service that scans AI systems
as they work in order to detect bias and provide explanations
 Works with Watson tech, as well as Tensorflow, SparkML,
AWS SageMaker, and AzureML
 Public ordering: soft law and guidelines

13. www.lcii.eu
5. Proposed framework
 Discrete externality (micro; personal)
 Negative: harm
 Positive: benefits
 For harms, ex post resolution, abstract standards, impact assessment
 Systemic externality (macro; societal)
 Negative
 Substitution effect in jobs
 Privacy
 Positive
 Complementarity effect
 Generative or General Purpose technologies
 For harms, more ex ante, prescriptive, impact assessment
 « Existernality »
 Negative: existential risk (Terminator)
 Positive: pure human enhancement
 For harms, more ex ante, proscriptive, deontological (no impact assessment or fat
tail risks)
More at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2931339

14. Liege Competition and Innovation Institute (LCII)
University of Liege (ULg)
Quartier Agora | Place des Orateurs, 1, Bât. B 33, 4000 Liege, BELGIUM
Thank you
Nicolas.petit@uliege.be
Twitter: @CompetitionProf