Citation:
Timothy D. Robinson, A Normative Evaluation of
Algorithmic Law, 23 Auckland U. L. Rev. 293 (2017)
Content downloaded/printed from HeinOnline
Tue Nov 27 03:17:26 2018
-- Your use of this HeinOnline PDF indicates your
acceptance of HeinOnline's Terms and Conditions
of the license agreement available at
https://heinonline.org/HOL/License
-- The search text of this PDF is generated from
uncorrected OCR text.
-- To obtain permission to use this article beyond the scope
of your HeinOnline license, please use:
Copyright Information
Use QR Code reader to send PDF
to your smartphone or tablet device
A NormativeEvaluationofAlgorithmicLaw
A Normative Evaluation ofAlgorithmic Law
TIMOTHY D ROBINSON*
Advances in technology have enabledthe abilityfor law to
be prescribed to specific circumstances according to
personalattributesorfact-specific information. Law can be
promulgated by a predictive algorithm which is perfectly
calibrated to produce some balance of outcomes. This
algorithmiclaw has two characteristics.It is, first, law that
is highly tailored to individuals and circumstances
accordingto a predictive model. It is, secondly, law that is
dynamic and evolves over time to respond to new
information. By tying law to risk or a prediction of
behaviour, algorithmic law can be extremely efficient at
achieving desired outcomes, which might range from
reducing trafficfatalities to maximising wealth. This article
presentsa normative evaluation of algorithmiclaw through
analysing two legal values that algorithmic law might
undermine; the rule of law andfreedom. For rule of law, it
is shown that algorithmic law can undermine equality
before the law and affect legal certainty. From a freedom
perspective, algorithmic law can subjugate the freedom of
individualsfor the majority interest in new ways, and can
act harshly to restrict an individual'sfreedom based on
theirpast, possibly unrelated, choices. This article identifies
algorithmiclaw as a distinct andnovel category of law that
has both immense potentialfor beneficial outcomes, and
unique ability to subvert legal values in new ways.
I INTRODUCTION
For the rational study of the law the black-letter man may be the man of
the present, but the man of the future is the man of statistics and the
master ofeconomics.
- Oliver Wendell Holmes, Jr.'
Algorithmic law is law made by a predictive algorithm, "perfectly
calibrated" to achieve some balance of objectives.2
Such an algorithm
* BSc in Computer Science, LLB(Hons). The author would like to thank Dr Arie Rosen of the
University ofAuckland, Faculty of Law, for his significant guidance, patience and enthusiasm.
1 OW Holmes "The Path ofthe Law" (1897) 10 Harv L Rev 457 at 469.
293
Auckland UniversityLaw Review
predicts whether a given situation is likely to produce a desired outcome and
then prescribes law on the basis of this prediction. Two characteristics define
algorithmic law. It is, first, law that is highly tailoredto individuals and
circumstances according to the predictive model. It is, secondly, law that is
dynamic and will respond to new information to evolve the law over time.
To visualise this law, consider an example offered by Anthony
Casey and Anthony Niblett - the regulation of traffic speed.3
Algorithmic
law would prescribe an objective, such as keeping road fatalities below a
certain number, then issue personalised commands to individuals according
to relevant situational information. These commands, defined as microdirectives,
could adjust permitted speed according to variables such as
weather conditions, the time of day or driver history.
This article presents a preliminary normative evaluation of the
desirability of algorithmic law. It does not intend to conclude whether such
law is always desirable - the desirability of such law will depend on the
particular instantiation of algorithmic law proposed. Rather, this article aims
to highlight algorithmic law as a novel category of law with unique
characteristics and jurisprudence, which, by its own nature, will undermine
and uphold particular normative values.
Parts II and III describe different scholarship on algorithmic law and
define the concept of algorithmic law. Part IV presents the argument for
algorithmic law by describing the efficiency-enhancing potential of
mathematically calibrated laws. Part V critiques algorithmic law according
to the legal values of the rule oflaw andfreedom.
II THE AGE OF THE ALGORITHM
Overview
We live in the "age of the algorithm".4 Actuarial algorithms "that can predict
individuals' future behavior" shape important decisions in areas as
widespread as education, medicine, politics, crime and insurance.5
Predictive algorithms can be used to assess creditworthiness.6
They
can evaluate teachers - to the extent that some are dismissed. They can
2 Anthony J Casey and Anthony Niblett "The Death of Rules and Standards" (2017) 92 Ind LJ 1401
at 1402.
3 At 1404.
4 Cathy O'Neil Weapons of Math Destruction:How Big Data IncreasesInequality and Threatens
Democracy (Crown, New York, 2016), back cover. O'Neil also suggests we live in the "age of
data". At 218.
5 Ariel Porat and Lior Jacob Strahilevitz "Personalizing Default Rules and Disclosure with Big
Data" (2014) 112 Mich L Rev 1417 at 1434-1435; and O'Neil, above n 4, at chs 3 and 5.
6 See Ashlyn Aiko Nelson "Credit Scores, Race, and Residential Sorting" (2010) 29 JPAM 39.
7 Bill Turque "206 low-performing D.C. teachers fired" The Washington Post (online ed,
Washington, DC, 15 July 2011). See also the 2016-2017 IMPACT programme guidebooks, for
example, District of Columbia Public Schools GeneralEducationTeachers with Individual ValueAdded
StudentAchievement Data (18 August 2016) at 6-8.
294 Vol 23 (2017)
A NormativeEvaluationofAlgorithmic Law
screen job applicants.8
They can predict criminal reoffending9
and identify
patterns of crime.10 They can determine government funding for
pharmaceuticals" or schools. 12
And they can identify children at a
heightened risk of maltreatment. 3
The technology that enables these predictions is artificial
intelligence - more specifically, machine learning - which can examine
"gigantic databases" of information and identify patterns that predict
individuals' future behaviour.14
As "[c]omputational power [grows] at
exponential rates"'5
and increasing volumesl of data are collected from a
greater number of Internet-connected sensors, these predictions will only
become more accurate over time. 17
Machine learning techniques allow seemingly unrelated information
to be tied together in unanticipated ways. For example, it has been found that
purchasing birdseed, or felt pads for furniture legs, indicates low credit
risk.18
Risky driving has been correlated with risky financial decisions.19
Tesco, the retail store, has identified that customers who purchase nappies
tend to have an increased demand for beer.20
An organisation that can predict outcomes will have better decision
making and resource allocation. For this reason, annual spending on "big
data" is forecasted to reach USD 58.9 billion in 2020 (with a compound
annual growth rate of 22.6 per cent).2 1
For comparison, this figure is greater
than the Gross Domestic Product of Uruguay,22
and is eleven times the
regular budget ofthe United Nations. 2 3
8 Gideon Mann and Cathy O'Neil "Hiring Algorithms Are Not Neutral" HarvardBusiness Review
(online ed, Massachusetts, 9 December 2016).
9 Grant T Harris and others "A Multisite Comparison ofActuarial Risk Instruments for Sex
Offenders" (2003) 15 Psychological Assessment 413.
10 Michelle Dickinson "Science & Tech: Big data on crime" The New ZealandHerald (online ed,
Auckland, 30 January 2016).
11 PHARMAC Pharmaceutical Management Agency: Cost-Utility Analysis (CUA) Explained
(August 2015) <www.pharmac.govt.nz>.
12 Nicholas Jones "How new funding model could affect schools" The New ZealandHerald (online
ed, New Zealand, 12 April 2017).
13 See "Vulnerable Children Predictive Modelling" Ministry of Social Development
<www.msd.govt.nz>. See also Richard MacManus "Holding algorithms accountable" (10 July
2017) Newsroom <www.newsroom.co.nz>.
14 Porat and Strahilevitz, above n 5, at 1434-1435.
15 Casey and Niblett, above n 2, at 1410.
16 Andrew McAfee and Erik Brynjolfsson "Big Data: The Management Revolution" Harvard
Business Review (online ed, Massachusetts, I October 2012) at 62.
17 Michael Chui, Markus LOffler and Roger Roberts "The Internet of Things" McKinsey Quarterly
(online ed, New York, March 2010).
18 Charles Duhigg "What Does Your Credit-Card Company Know About You?" The New York Times
Magazine (online ed, New York, 12 May 2009). See also Porat and Strahilevitz, above n 5, at
1454-1455.
19 Edward R Morrison and others "Health and Financial Fragility: Evidence from Car Crashes and
Consumer Bankruptcy" (Coase-Sandor Institute for Law and Economics Working Paper No 655,
17 October 2013) Social Science Research Network <www.ssm.com> at 10-12.
20 "Crunching the numbers" The Economist(online ed, London, 19 May 2012).
21 Ashish Nadkarni and Dan Vesset "Worldwide Big Data Technology and Services Forecast, 2016-
2020" (December 2016) International Data Corporation <www.idc.com>.
22 The World Bank "Uruguay" <http://data.worldbank.org>.
23 "Fifth Committee Recommends $5.4 Billion Budget for 2016-2017 Biennium as It Concludes
Main Part of Seventieth Session" (23 December 2015) United Nations <www.un.org/press>.
295
Auckland UniversityLaw Review
More broadly, the technological frameworks on which algorithmic
law can exist are undergoing rapid change - or are yet to be conceived.
Previously-simple machines are now becoming intelligent and
interconnected.2 4
Where, formerly, algorithmic law would have needed to be
stored and promulgated from a central database, blockchain technology has
shown that algorithmic law could be implemented as a decentralised
system.25 Laws might not require human enforcement and, instead, be an
inherent restriction within technology itself.26 Where applications of
algorithmic law might seem far-fetched or unrealistic, with the rapid pace of
technological advancement it is not unfeasible to assume that such law will
become viable in the near future.
Scholarship on Algorithmic Law
The growing use and importance of prediction technology has prompted a
number of suggestions for how such technology might be used in law. Casey
and Niblett propose regulation by micro-directives: context-specific legal
commands that specify the exact behaviour required in a given situation. A
predictive algorithm defines precise rules for individuals that are optimised
to achieve a particular outcome. In this way, micro-directives are the hybrid
of a rule and a standard. Each micro-directive is prospective, yet microdirectives
as a whole will tend towards efficient outcomes and be "perfectly
calibrated" to achieve some balance of objectives.27 As well as predictive
technology, micro-directives are enabled by developments in communicative
technology that allow a micro-directive to be transmitted or queried in real
28
time.
Aaron Wright and Primavera De Filippi suggest a similar idea of
algorithmic governance: a "highly optimized" means of efficiently
regulating society. 29 Such law might be self-enforcing; and Wright and De
Filippi warn of the risk of "highly prescriptive and deterministic" law. 30
Ariel Porat and Lior Jacob Strahilevitz suggest that redictive
algorithms should be used to determine defaultrulesfor contracts. ' Default
rules which are personalised to an individual's characteristics or past
behaviour will better and more efficiently achieve outcomes than universal
default rules. Such personalised default rules are the best prediction of the
rules an individual would actually choose. Universal default contract rules
can act against minority interests - for example, a person in a wheelchair
who requires home delivery might be subject to a majority default position
24 Chui, Liffler and Roberts, above n 17.
25 Satoshi Nakamoto Bitcoin: A Peer-to-PeerElectronic Cash System (2008). See also "Ethereum
Project" Ethereum <www.ethereum.org>.
26 Aaron Wright and Primavera De Filippi "Decentralized Blockchain Technology and the Rise of
Lex Cryptographia"(12 March 2015) Social Science Research Network <www.ssrn.com> at 43.
27 Casey and Niblett, above n 2, at 1402.
28 At 1403-1404.
29 Wright and De Filippi, above n 26, at 41-44.
30 At 43 and 44.
31 Porat and Strahilevitz, above n 5.
296 Vol 23 (2017)
A NormativeEvaluationofAlgorithmic Law
that assumes pick-up from store.32
Porat and Strahilevitz suggest that default
rules can be determined by machine learning techniques. Through
collecting data on "guinea pigs" - and using this to categorise new
individuals - the default rules of similar persons can be applied.34
This is an
established artificial intelligence technique called instance-basedlearning.3 5
Omri Ben-Shahar and Ariel Porat argue for a personalisedstandard
of negligence - the standard of the "reasonable you". 36
Information about a
person is used to identify the actual risk produced by their actions
("personalized expected risk").37
A personal standard is then imposed, given
what might be broadly considered as aggravatingand mitigating factors
affecting risk. A young person could have a standard that permits driving at
a higher speed than an elderly person because the young person has a faster
reaction time.38
The actual risk created by each person at different speeds
might be the same. It is suggested that "Big Data" can aid in the formulation
of such personal standards and might use information relating to "other
characteristics" that are not unique to the standard at hand - such as
carelessness or a general disposition to risk.39
The above examples relate to different areas of law, and each will
have different critiques. However, at the same time, these ideas share a
common theme - the use of precise, algorithmic predictions to tailor law to
achieve better outcomes. These ideas will be broadly discussed as
algorithmiclaw.
Algorithmic law might be effected without explicit enactment, on
top of traditional law - explained by John 0 McGinnis and Steven
Wasick's concept of legal search.4 0
A system of micro-directives can
emerge through software that outperforms lawyers at providing accurate
legal advice. 4 1
This might not be law, but would be the best estimate of how
any discretion within the law will be applied. Through a Holmes' prediction
theory perspective, McGinnis and Wasick argue that such a legal "search
engine" effectively then becomes the law:4 2
The question of"what is the case law on the discovery rule in X situation"
will not be scattered among random case numbers or found in legal
encyclopedias, but will instead be entirely contained within the phrase ...
the search engine itself will effectively become the law.
32 At 1427.
33 At 1434-1435.
34 At 1450-1451.
35 Stuart J Russell and Peter Norvig Artificial Intelligence: A Modern Approach (3rd ed, Prentice
Hall, Upper Saddle River (NJ), 2010) at 737-739.
36 Omri Ben-Shahar and Ariel Porat "Personalizing Negligence Law" (2016) 91 NYU L Rev 627 at
631.
37 At 633.
38 At 630.
39 At 631.
40 John 0 McGinnis and Steven Wasick "Law's Algorithm" (2014) 66 Fla L Rev 991.
41 Casey and Niblett, above n 2, at 1422-1423.
42 McGinnis and Wasick "Law's Algorithm", above n 40, at 1023. See also 1023-1026.
297
Auckland UniversityLaw Review
Algorithmic law might also be established by private parties through the
"interplay of reasonableness, industry standards, and technology". 43
If legal
prediction software is used by private parties to guide decision making, it
might become seen as an industry standard.44 Failure to consult, or a
significant deviation from a recommended directive, might then be
45
considered per se unreasonable in assessing standards of care.
III DEFINING ALGORITHMIC LAW
Algorithmic law describes laws that have two aspects in common. First,
rules are highly tailoredto individuals and circumstances - a characteristic
facilitated by a predictive model. Secondly, the model for producing tailored
laws is dynamic and will adapt as outcomes are measured against some goal.
Both of these aspects are present in regular law, although to a lesser extent.
It will be shown that algorithmic law is different in degree rather than in
kind.
The Tailoring of Laws
The first aspect of algorithmic law is that laws are highly tailored to
individuals or circumstances according to a predictive model that seeks to
achieve some outcome. Laws might be prescribed to individuals within
discrete categories or might vary along a continuum. 4 6
Tailoring is not a novel characteristic of law. Laws are already
tailored. A law applying to a group, such as company directors, will not
affect someone who is not a part of that group. Current laws target people
who do a particular activity, or are of a certain age,48
or buy a certain
product.49
Standards of negligence can depend on the capabilities,so or the
professional skill" of the defendant.
The tailoring that defines algorithmic law is, therefore, a matter of
degree. Where traditional law is tailored to broad categories, algorithmic law
is finely calibrated and makes granular categorisations5 2
of people and
situations. Such granularity is enabled by - and can only realistically be
achieved with - the use of a predictive algorithm.
However, prediction alone cannot be a sufficient criterion to
distinguish law as algorithmic. Already, legislation is enacted after
considering predicted outcomes. In debating a Bill, Members of Parliament
43 Casey and Niblett, above n 2, at 1421.
44 At 1422.
45 At 1421-1422.
46 Cass R Sunstein "Deciding by Default" (2013) 162 U Pa L Rev I at 48.
47 See, for example, s 129A ofthe Sentencing Act 2002.
48 See, for example, the definition of "minor" in s 5 of the Sale and Supply ofAlcohol Act 2012.
49 See, for example, s 213 ofthe Customs and Excise Act 1996.
50 See generally Billy Higgs andSons Ltd v Baddeley [1950] NZLR 605 (CA).
51 See generally Bolitho v City and HackneyHealthAuthority [1998] AC 232 (HL); and New Zealand
GuardianTrust Co Ltdv Brooks [1995] 1 WLR 96 (PC).
52 Sunstein uses the termfine-grained. See Sunstein, above n 46, at 48.
298 Vol 23 (2017)
A NormativeEvaluationofAlgorithmic Law
will anticipate the effects of that Bill - a process that might include
economic analysis. Predictions might not be quantifiable - such as how a
certain principle of law could be affected - but a prediction is still made.
The tailored aspect of algorithmic law then is no different in kind
from any regular law. Rather, algorithmic law is different in degree. There
becomes a point where law is so fine-grained that it takes on a new character
and ought to be considered a distinct category oflaw.
To illustrate this, consider a basic algorithm that prescribes a speed
limit given the following variables:
(1) time of day (in hours only);
(2) weather (good, neutral, poor);
(3) road type (motorway, rural, semi-rural, residential, urban);
and
(4) driver history (good, neutral, poor).
There will be 1,080 tailored situations with rules prescribed - far
more than any lawmaker would traditionally attempt to define.53
In form this
is no different than having five tailored speed limits according to road type.
However, the sheer level of tailoring that can be reached by an algorithm
means the law takes on a new character.
Dynamism: Adapting the Predictive Model
The second defining aspect of algorithmic law is that it is dynamic. A
prediction algorithm will gradually change and respond to new information
to refine prediction accuracy. This aspect is implicit in the nature of a
prediction algorithm and machine learning itself.5 4
The dynamism of algorithmic law allows laws to adapt to external
changes. For example, a speed limit might increase to meet a long-term
improvement in car safety or reduce to respond to a sudden fluctuation in
risk:"
Algorithm-driven laws will automatically and rapidly adapt to the
circumstances, optimizing according to the objective of the law. But
changes to the law result in winners and losers. Frequent changes to the
law may impose additional risks on individuals and may affect the
willingness of individuals to invest in projects that may be subject to legal
uncertainty. A smart machine will, however, be able to take into account
any effects on the values of reliance investments to find a global
optimum, rather than merely a local optimum.
Like tailoring, the dynamism of algorithmic law reflects the law-making that
already occurs, but to a greater extent. Law change takes place by lawmakers
predicting that the change will bring about a better outcome. In this regard,
law is already outcome-focused. Algorithmic law might merely shift the
mechanism for change from a political, discretionary assessment of outcome,
53 Calculated from the possible combinations of the given variables: 24 x 3 x 5 x 3 = 1,080.
54 See generally Russell and Norvig, above n 35, at chs 18-21.
55 Casey and Niblett, above n 2, at 1437-1438 (footnotes omitted).
299
Auckland UniversityLaw Review
to a pre-defined and precise scientific assessment. Political assessment of
outcome might often be reached through scientific analysis anyway - and
algorithmic law might automate choices that are already guided by economic
analysis.
It is possible to avoid dynamism by using a pre-defined prediction
model that is unaffected by new information. However, this accepts some
level of prediction inaccuracy. Such a prediction model would produce more
certain law in the short term, but would also require periodic repeal and
replacement to avoid gross inaccuracy.
Overcoming Practicalities
A natural criticism of prediction-based law is to point to some practical
problem in the accuracy of the prediction model, for example, faulty data or
a failure to consider information.56
For the purposes of presenting a
conceptual normative evaluation, I will overlook these concerns and assume
the algorithm is highly accurate. I also avoid criticisms that the wrong, or
overly narrow, goal would be prescribed to a prediction algorithm.5 7
This is
best left as the domain of the lawmaker. Finally, I avoid the idea that
individuals might game the system and change their behaviour to trick the
prediction model. I assume that implicit in accurate predictions is the need to
account for this possibility and create a model that cannot be tricked by
minor changes in behaviour. Of course, incentivising an individual to
change their behaviour to be more desirable - for example, by taking fewer
risks - is the entire purpose of such algorithmic law.
IV THE ARGUMENT FOR ALGORITHMIC LAW
The Death of Rules and Standards
The promise of algorithmic law is that by tailoring law to specific groups or
situations - and updating tailored law with new information over time
the
law can better achieve some functional goal. This goal might be
anything. Maximising wealth is a natural goal for algorithmic law, but so too
might be other objectives decided by the lawmaker, such as reducing traffic
56 See Anna Chalton "Rape Myths and Invisible Crime: The Use of Actuarial Tools to Predict Sexual
Recidivism" (2015) 2 PILJNZ 19 at 45-48.
57 See O'Neil, above n 4, at 50-59. O'Neil describes the problem of narrow goals by using the
example of the college ranking system developed by US News & World Report in the United
States. O'Neil argues the rankings use proxies to assess college quality that cannot fully capture
educational experience. Rankings might be useful at identifying a top-tier college from a mid-tier
college, but cannot helpfully distinguish between two mid-tier colleges. If these rankings guide
students' college selections, a self-fulfilling prophecy is created in that top students seek to attend a
top-ranked college, which further increases that college's ranking. In consequence, colleges will
focus their efforts on the particular criteria assessed, even if those criteria do not capture the full
picture of college quality.
58 Casey and Niblett, above n 2, at 1423.
Vol 23 (2017)300
A NormativeEvaluationofAlgorithmicLaw
fatalities. How this functionalism is achieved is best understood from a law
and economics analysis of rules and standards.
Algorithmic law is a "hybrid" of rules and standards, meaning that
functional goals can be better achieved while minimising other costs. 59
The
"canonical" 60 law and economics approach to rules and standards is given by
Louis Kaplow.6
1 Rules are precise and ex ante in nature, while standards
leave open ambiguity that is clarified retrospectively after an individual acts
in an unprecedented manner.62
An example of a rule is the "bright-line test"
that deems profit from disposing land within two years as taxable income.63
This rule overrides the general standard that capital gain from a property is
only taxable if the property was acquired for the purpose or intention of
disposal. 4 The distinction between rule and standard is one of degree.65
A
law prohibiting "vulgar behaviour" might be considered to be a rule or a
standard depending on whether the set of acts that are "vulgar" are
predominantly defined ex ante or ex post.66 Rules are typically expensive to
create because all possible acts and areas of application need to be
considered. However, because rules are certain, their enforcement is
inexpensive.67 Standards are cheap to create and can be applied flexibly to
adhere to a goal, but because of this will be expensive to enforce. Individuals
subjected to a standard will also incur costs in learning - and seeking
advice on - the scope of a standard.
Whether a rule or a standard is desirable for lawmakers will depend
on the particular error, decision and uncertainty costs. 6 8
Lawmakers
effectively "trade-off between certainty and calibration". 69 A standard will
be most calibrated to achieve a functional goal, but will lack certainty and
incur ongoing enforcement costs. This is why standards are typically
preferred only when behaviour is diverse and infrequent.70
Rules are certain,
but their lack of calibration means they will be over-inclusive and apply
even when the underlying legal justification does not - for example a traffic
light will continue to order cars to stop even when there is no other traffic on
the road.
Algorithmic law changes this dichotomy in a way that Casey and
Niblett describe as the death of rules and standards: 7 2
59 Casey and Niblett, above n 2, at 1410.
60 McGinnis and Wasick, above n 40, at 1029.
61 Louis Kaplow "Rules Versus Standards: An Economic Analysis" (1992) 42 Duke U 557 as cited
in McGinnis and Wasick, above n 40, at 1029.
62 Casey and Niblett, above n 2, at 1407.
63 Income Tax Act 2007, s CB 6A.
64 Section CB 6A(6). See also s CB 6.
65 Kaplow, above n 61, at 600-601.
66 At 600-601.
67 At 563.
68 Casey and Niblett, above n 2, at 1407-1408.
69 At 1402.
70 At 1408.
71 See Kimberly Kessler Ferzan "Prevention, Wrongdoing, and the Harm Principle's Breaking Point"
(2013) 10 Ohio St J Crim L 685 at 688.
72 Casey and Niblett, above n 2, at 1410.
301
Auckland University Law Review
... the result will be a new hybrid form of law that is both rule and
standard. The lawmaker can set a broad objective, which might look like a
standard. But the predictive technology will take the standard and
engineer a vast catalog of context-specific rules for every scenario.
To illustrate that this hybrid can be preferable to a rule or standard, Casey
and Niblett imagine an algorithm that predicts whether a patient requires
surgery. Two scenarios are given, with algorithms of different accuracies:73
Under scenario I [a bad prediction algorithm], the technology should
have no effect on your decision as a regulator to implement a rule or a
standard. You should implement a standard and determine liability on a
case-by-case basis, learning more about doctors' behavior over time.
Under scenario 2 [an accurate prediction algorithm], however, the optimal
form of the law will be different. The machine's predictions provide the
exact content of the law. The machine provides microdirectives for each
and every scenario. The over- and underinclusivity associated with simple
rules have disappeared. ... The justification for relying on ex post
adjudication of standards - reducing the error costs of rules - is gone.
Further, we have an added benefit of eliminating uncertainty for the
doctors. If they follow the directive of the machine, they know they will
not be held liable.
The desirability of a system of micro-directives does not depend on a perfect
prediction algorithm.74
Rather, there reaches what Casey and Niblett
describe as:"
... a point where the technology is good enough that the costs of using a
microdirective are sufficiently low so that there is no longer any need to
use traditional rules or standards
Highly calibrated law - akin to that of a standard - then becomes possible
in areas where it would have previously been inefficient.
Calibrated Law: Eliminating Risk Subsidisation
To illustrate the benefits of calibration, consider how calibrated law can
make individuals liable for their own risk. Porat and Strahilevitz discuss this
in the context of personalised default rules for contracts.7 6
Some laws that
apply generally cause "cross subsidies" between groups in society. 7 One
group will face higher costs, or greater restrictions on liberty, because of the
behaviour of another group. Speed limits are an example of this: it might be
73 At 1414-1415.
74 At 1415.
75 At 1415 (emphasis in original).
76 Porat and Strahilevitz, above n 5, at 1453-1454.
77 At 1453-1454.
302 Vol 23 (2017)
A Normative EvaluationofAlgorithmicLaw
safe for some people to drive at a high speed, but the general prescribed limit
must be lower than this to account for drivers who are unskilled or easily
distracted.
Algorithmic law reduces this cross subsidisation. Consider the
Consumer Guarantees Act 1993, which mandates a guarantee for the quality
of products bought in trade. 8 Businesses pass the costs of this to consumers
uniformly, but, naturally, some consumers will be more reckless in how they
treat their product and will have a higher probability of claiming their
guarantee. This means that other, more cautious, consumers effectively
subsidise the behaviour of the group of risk-taking consumers prone to using
their guarantee. In contrast, algorithmic law can be calibrated so that costs
are borne more proportionally. There might be a less stringent guarantee for
risky consumers - meaning that cautious consumers can benefit from the
same guarantee at a reduced cost.7 9
A more complex example can be found in New Zealand's 90 day
trial law for new employees.so Employers, facing the risk that a potential
employee is not as they appear in an interview, can contract with the
employee to undergo a 90 day trial. During a trial, the employee can be
dismissed at any time and cannot bring legal proceedings in respect of that
dismissal.81 This trial period has become common practice, with 66 per cent
of hiring employers employing one or more of their new staff on a trial.82
The trial benefits "people at the margins of the labour market" who appear to
be risky employees (but would remain employed at the end of a trial) as well
as the employers who would like to give such people a chance.83
However,
the trial also creates a general risk for non-risky new employees, who, during
the trial, might be dismissed with reduced rights and in a way that would
otherwise be illegal - for example, they could be immediately dismissed
following an economic downturn.
Algorithmic law can limit this general risk to only the people "at the
margins" who the law purports to help. It might dictate that a trial period
only be available for the set of employees who, according to a predictive
algorithm, appear risky (for example, those who are youthful or have a
transient employment history). These employees can then retain the benefit
of the trial, while regular employees, who would be employed regardless of
the trial, are not subjected to the risk of sudden dismissal. This is law that
could otherwise be implemented as a standard but would likely be
prohibitively uncertain.
The following table sets this idea out in more detail:
78 Consumer Guarantees Act 1993, ss 6-7.
79 In this example, all consumers would then pay less for a product, but risk-averse consumers would
benefit by retaining the same guarantee. It is feasible that a company could implement its own
personalisation by tailoring price to the risk for each guarantee.
80 Employment Relations Act 2000, s 67A.
81 Employment Relations Act 2000, s 67A(2)(c).
82 Ministry ofBusiness, Innovation & Employment NationalSurvey ofEmployers 2014/15: Summary
Findings(April 2016) at 4.1.1. The number increased from 63 per cent in 2013/2014 and 59 per
cent in 2012/2013. It was not recorded in 2015/2016.
83 See comments made by Hon Kate Wilkinson MP. (9 December 2008) 651 NZPD 318 and 319.
303
Auckland UniversityLaw Review
Employed with
Employed 90 day trial law? Employed with an
Employee without 90 day (with additional algorithmic law
trial law? risk of dismissal trial period?
within 90 days)
Appears risky, is a No Yes (then Yes (then
bad employee dismissed within dismissed within
the trial period). the trial period).
Appears risky, is a No Yes (additional Yes (additional
good employee risk of dismissal risk of dismissal
within 90 days). within 90 days).
Appears non-risky, Yes (then Yes (then Yes (then
but is a bad dismissed under dismissed within dismissed under
employee normal laws). the trial period). normal laws).
Appears non-risky, Yes Yes (additional Yes; no additional
and is a good risk of dismissal risk of other
employee within 90 days). dismissal.
Calibrated Law: Personalisation
Much of the appeal of algorithmic law relates to its potential for
personalisation. While it is possible for algorithmic law to be unrelated to
personal attributes (for example, a speed limit that varies with weather), it is
through personalising law that the greatest reduction in risk subsidisation is
likely to occur.
The efficiency of personalised standards for negligence is examined
by Ben-Shahar and Porat.84
They analyse two types of personalisation
skill-based
and risk-based - and, broadly, conclude that both types of
personalisation are efficient and optimise levels of care.
Skill-based personalisation reflects the different costs that
individuals incur to reach different standards of care. Ben-Shahar and Porat
identify that "[m]ore skilled injurers can achieve the same reduction in risk
as unskilled injurers by spending less on care."8 5
The efficiency of this is
"wholly intuitive":86
It pays to impose higher burdens on the more competent actors to take
advantage of their greater productivity. Thus, the driver who is more
competent in operating sophisticated technical equipment should probably
use it, while the less competent driver perhaps should not. ...
Personalized standards, although imposing more differentiated levels of
care, impose less differentiated costs of care on the various types of
injurers.
84
85
86
Ben-Shahar and Porat, above n 36.
At 647.
At 649 (emphasis in original).
304 Vol 23 (2017)
A NormativeEvaluationofAlgorithmicLaw
Risk-based personalisation reflects the variations in risk of harm that are
created by people operating at the same standard of care. Two people
exercising the same level of care will not necessarily produce the same
outcome due to differing knowledge and experience.87
Risk-based
personalisation is also efficient:88
It pays to impose higher burdens on the more risky actors since any
additional burden would produce more risk reduction for the high-risk
actor than for the low-risk actor. Thus, the high-risk driver with the poor
instincts should take more care than the driver with the sharper instincts.
Similarly, the high-risk doctor with less experience and knowledge should
take more care than the more experienced and knowledgeable doctor.
It appears to contradict that it is efficient for both unskilled individuals to
take less care and riskier individuals to take more care; however this is
reconciled in two ways. 89
Care and precaution can be viewed as having
"multiple dimensions":90
A doctor should take a low level of the type of precautions that she is
unskillful in deploying. That, in turn, makes her relatively riskier and
justifies imposing upon her a higher level of care with respect to other
precautions.
The alternative reconciliation is to view skill and risk as "pulling in a
different direction".91 The net combination of the forces then dominates: 9 2
If, for example, only one type of precaution is available, a driver who is
both riskier and low skill may in the end be required to take either a
higher or lower level of care, depending on which effect dominates. It is
therefore possible that despite the low skill in applying this singledimensional
precaution, the high-risk driver may be required after all to
take a higher level of care.
Prediction of Choice
Algorithmic law can be used to substitute the choices of individuals. This is
suggested by Porat and Strahilevitz for default contract rules.93
Consider
New Zealand laws requiring informed consent for medical procedure. In
situations where informed consent is not possible - for example, the patient
is unconscious and there are no persons with the power to consent for the
patient - a doctor may decide to carry out the procedure. This is permitted
if the doctor, having ascertained the views of the patient, believes "on
reasonable grounds, that the provision of the services is consistent with the
87 At 650.
88 At 651-652.
89 At 652.
90 At 652.
91 At 652.
92 At 652.
93 See Porat and Strahilevitz, above n 5, at 1442-1444.
305
Auckland UniversityLaw Review
informed choice the [patient] would make if [they] were competent".9
Algorithmic law could provide an alternative to this approach and issue a
directive to the doctor, predicting consent from known information (such as
prior consent given by the patient). This might be a more accurate prediction
mechanism than the doctor's own views and may align better with the
desired role and liabilities of doctors.
V CRITIQUES OF ALGORITHMIC LAW
This part offers two critiques of algorithmic law. The first critique is that
algorithmic law undermines aspects of the rule of law, specifically: equality
before the law; and certainty of law. The second critique is that algorithmic
law can constrain individual rights to freedom.
Rule of Law
1 EqualityBefore the Law
Equality has long been a fundamental value of the rule of law. AV Dicey's
famous conception of the rule of law prescribed the "equal subjection of all
classes" to the law of the land.9 5
Lord Bingham has described equality before
the law as a "cornerstone" of society,9 6
while John Locke wrote against law
varied "in particular cases" and promoted "one rule for rich and poor, for the
favourite at Court, and the country man at plough". Rule of law was a
value called upon by Lord Scarman in 1983 to reject the argument that
habeas corpus protection did not apply to non-British nationals:9 8
Every person within the jurisdiction enjoys the equal protection of our
laws ... He who is subject to English law is entitled to its protection. This
principle has been in the law at least since Lord Mansfield freed "the
black" in Sommersett's Case (1772) 20 St.Tr. 1.
Equality in law reflects the idea that equal treatment is just; a universal value
described by Justice Scalia as a "motivating force of the human spirit": 99
Parents know that children will accept quite readily all sorts of arbitrary
substantive dispositions ... But try to let one brother or sister watch
television when the others do not, and you will feel the fury of the
fundamental sense ofjustice unleashed.
94 Code ofHealth and Disability Services Consumers' Rights, Right 7(4)(c)(i).
95 AV Dicey Introduction to the Study of the Law of the Constitution (10th ed, Macmillan, London,
1959) at 202.
96 Tom Bingham The Rule ofLaw (Penguin Books, London, 2011) at 55.
97 Peter Laslett (ed) John Locke: Two Treatisesof Government (2nd ed, Cambridge University Press,
London, 1967) at [142].
98 R v SecretaryofStatefor theHome Department, exparteKhaivaja [1984] 1 AC 74 (HL) at 111.
99 Antonin Scalia "The Rule of Law as a Law of Rules" (1989) 56 U Chi L Rev 1175 at 1178.
306 Vol 23 (2017)
A Normative EvaluationofAlgorithmic Law
Equality also serves a constitutional purpose: manifesting in concepts of
legislative generality and bills of attainder which recognise that unequal and
specific law can undermine the judiciary.' 0
As Justice Jackson of the
Supreme Court ofthe United States described in 1949:'01
... equality is not merely abstract justice ... there is no more effective
practical guaranty against arbitrary and unreasonable government than to
require that the principles of law which officials would impose upon a
minority must be imposed generally.
However, the nature of law makes equality before the law difficult to define.
Laws apply conditionally: all members of society are subject to a law, yet
only some might trigger the conditions that impose substantive outcomes.
Consider New Zealand's Health and Safety Act, which imposes
health and safety duties on workers, officers and "person[s] conducting a
business or undertaking".' 02
In form this law applies to all people, but in
substanceonly a person meeting the definition of a relevant category could
ever be burdened with its effect. This is the same for other laws, including
those commonly held up as an aberration to equality before the law: laws
that apply only to citizens of a particular country,' 03
laws excluding women
from the military draft,'" laws that are applied randomly at the toss of a
coin.s0 5
There must, therefore, be some substantive measure for determining
acceptable and unacceptable inequality before the law. This is not
straightforward, as different views exist on the limitations of legal equality.
As Bingham observes, the principle of equality before the law has, in liberal
countries, not deterred the enactment of anti-terrorism laws that target
groups by nationality. 0 6
A starting point is the scholarship of Bruno Leoni, who in Freedom
andthe Law argued that any legal categorisation undermines equality of law.
Drawing on Dicey's denouncement of separate administrative tribunals, he
wrote:
07
Within each category people will all be "equal" before the particular law
that applies to them, regardless of the fact that other people, grouped in
other categories, will be treated quite differently by other laws. ... Thus,
by a slight change in the meaning of the principle of "equality," we can
pretend to have preserved it. Instead of "equality before the law," all that
we shall have will then be equality before each ofthe two systems of law
100 Duane L Ostler "Legislative Judging: Bills Of Attainder in New Zealand, Australia, Canada and
the United States" (2014) 22 Wai L Rev 78 at 78.
101 RailwayExpress Agency v PeopleofState ofNev York 336 US 106 (1949) at 466-467.
102 Health and Safety at Work Act 2015, ss 17-19.
103 See generally commentary on Japanese internment camps in World War Two, for example, Diane
P Wood "The Rule of Law in Times of Stress" (2003) 70 U Chi L Rev 455 at 462.
104 Rostker, DirectorofSelective Service v Goldberg453 US 57 (198 1).
105 See Ronald Dworkin Law's Empire (Belknap Press, Cambridge (Massachusetts), 1986) at 178-
184.
106 Bingham, above n 96, at 58.
107 Bruno Leoni Freedom and the Law (3rd ed, D van Nostrand, Princeton (NJ), 1961) at 68-69
(emphasis in original).
307
Auckland UniversityLaw Review
enacted in the same country, or, if we want to use the language of the
Dicean formula, we shall have two laws of the land instead of one.
Of course, we can, in the same way, have three or four or thousands of
laws of the land-one for landlords, one for tenants, one for employers,
one for employees, etc. This is exactly what is happening today in many
Western countries where lip service is still paid to the principle of "the
rule of law" and hence of "equality before the law."
Such a strict view of equality cannot be said to prevail in modern theory. For
Leoni, progressive taxation was incompatible with the ideals of equality in
law, yet such taxation continues to be a core feature of a liberal state over
fifty years later.108
That specific rules continue to exist for employers and
landlords shows that there must be some modern compromise on the ideal of
equality before the law. For most scholars, this compromise depends on the
extent to which inequality in law is justified; a view described by Richard
Posner as the law having a "rational structure".1 09
There must be a "convincing distinction" between two categories in
law for there to be acceptable inequality.' 10
Bingham argues this distinction
is met if a categorisation is "unobjectionable" and "relevant" to the
distinction made between groups.'1
Therefore, non-nationals may be subject
to deportation because citizenship is a defining feature of their group, but
they cannot be excluded from the protection of habeas corpus because that
would be arbitrary.
Yet any inequality prescribed by an algorithm will be relevant or
justified along statistically convincing lines. This is the argument for
economic justice presented by Posner:'
Economic theory is a system of deductive logic: when correctly applied, it
yields results that are consistent with one another. Insofar as the law has
an implicit economic structure, it must be rational; it must treat like cases
alike.
That there is some mathematical explanation for differential treatment does
not mean such treatment is unobjectionable. If non-nationals have a higher
risk of terrorism than nationals, must this be accepted as a statistical feature
of their group that justifies alternative law? It is Bingham's second idea of
"objectionable" differentiation that is relevant for algorithmic law.
(a) Objectionable Categorisations
Laws prohibiting discrimination have long provided for certain rights of
equal treatment to mitigate wrongful biases in society. In New Zealand,
108 At 70. The work was originally published in 1961.
109 Richard A Posner The Economics of Justice (Harvard University Press, Cambridge
(Massachusetts), 1981) at 75.
110 George P Fletcher "Equality and the Rule ofLaw" (1990) 10 Tel Aviv U Stud L 71 at 80-81.
Ill Bingham, above n 96, at 57.
112 Posner, above n 109, at 75.
308 Vol 23 (2017)
A NormativeEvaluation ofAlgorithmic Law
prohibited grounds of discrimination are listed in s 21 of the Human Rights
Act 1993 and include sex, race, ethical belief, age, political orientation and
sexual orientation.
As stated in Brown v Board of Education of Topeka, even
differential treatment of a group in name only can be detrimental:"
Segregation ... has a detrimental effect upon the colored children. The
impact is greater when it has the sanction of the law; for the policy of
separating the races is usually interpreted as denoting the inferiority of the
negro group.
Discrimination is a complex area of law and ethics, and differential
treatment will not always be discriminatory:' 14
On the one hand we think disabled people should be treated the same as
other people. They should be able to compete fairly on merit for prized
positions in society, in employment and in public life. On the other hand
we think some accommodation should be made for their disabilities.
This article does not intend to delve into the limitations of discrimination
law. It is sufficient to observe that there are categories of established
discrimination where differential treatment is presumptively undesirable.
Algorithmic law might be presented as a solution to differential
treatment - a fair and mathematical test that eliminates human discretion
and unconscious bias. " Yet, as observed by Cathy O'Neil, bias might
merely be "camouflaged" by technology." 6
Embedded within algorithmic
models are "a host of assumptions, some of them prejudicial"." 7
Predictive
algorithms are constructed, "not just from data but from the choices we make
about which data to pay attention to", and involve assumptions that are
"fundamentally moral".'"s
The idea that algorithmic law cannot be discriminatory if it is blind
to prohibited grounds of discrimination lacks truth for two reasons. First, as
a practical feature of prediction algorithms, discrimination can occur by
inference of a prohibited ground, even if this is not an explicit variable. Race
can be predicted from its correlation with other information, such as
postcode or online behaviour." 9
Secondly, as critics of neutrality have argued, even neutral law can
produce coercive or discriminatory results because of underlying societal
113 Brown v BoardofEducation of Topeka 347 US 483 (1954) at 494.
114 Nicholas Mark Smith "Basic Equality and its Applications" (PhD Thesis (Law), University of
Auckland, 2006) at 73.
115 Casey and Niblett, above n 2, at 1428.
116 O'Neil, above n 4, at 25.
117 At 25.
118 At 218.
119 Alistair Croll "Big data is our generation's civil rights issue, and we don't know it" (2 August
2012) O'Reilly Radar <http://radar.oreilly.com> as cited in Porat and Strahilevitz, above n 5, at
1435.
309
Auckland UniversityLaw Review
inequalities. Mary Joe Frug gives an example of the inequalities caused by
neutral prostitution law because "most sex workers are women":1 2 0
... even though anti-prostitution rules could, in theory, generate parallel
meanings for male and female bodies, in practice they just don't. ... By
characterizing certain sexual practices as illegal, these rules sexualize the
female body. They invite a sexual interrogation of every female body: Is
it for or against prostitution?
The above two points are reflected in criticisms of the Northpointe crime
prediction software used in Florida to assess criminal risk.12 1
The software
was found to falsely identify black people as future offenders almost twice
as frequently as it falsely identified white people. While the software does
not have any data on race, it takes information on family members who have
been imprisoned, acquaintances who have been arrested or have convictions
for illegal drugs, and a person's socioeconomic status. 122
In this way the
software, and algorithmic law more generally, might codify police and
judicial bias, and reinforce pre-existing social inequalities. These concerns
were expressed in 2014 by then Attorney General of the United States, Eric
Holder: 23
Here in Pennsylvania and elsewhere, legislators have introduced the
concept of "risk assessments" that seek to assign a probability to an
individual's likelihood of committing future crimes and, based on those
risk assessments, make sentencing determinations. Although these
measures were crafted with the best of intentions, I am concerned that
they may inadvertently undermine our efforts to ensure individualized and
equal justice. By basing sentencing decisions on static factors and
immutable characteristics - like the defendant's education level,
socioeconomic background, or neighborhood - they may exacerbate
unwarranted and unjust disparities that are already far too common in our
criminal justice system and in our society.
(b) Seemingly Random Categorisations
Aside from discriminatory categories, it is possible that other, seemingly
random, categorisations are objectionable. A law might be different for
attractive and unattractive people. Cricket players, if seen as prone to some
crime, might be prescribed harsher sentences than rugby players. Ronald
Dworkin describes laws enforced along arbitrary grounds as "checkerboard"
laws.124
120 Mary Joe Frug PostmodernLegal Feminism (Routledge, New York, 1992) at 131-132.
121 Julia Angwin and others "Machine Bias" ProPublica(online ed, New York, 23 May 2016). See
also their analysis: Julia Angwin and others "How We Analyzed the COMPAS
Recidivism Algorithm" ProPublica(online ed, New York, 23 May 2016).
122 "Sample-COMPAS-Risk-Assessment" ProPublica <www.propublica.org>.
123 Eric Holder, Attorney-General of the United States (speech at the National Association ofCriminal
Defense Lawyers 57th Annual Meeting and 13th State Criminal Justice Network Conference,
Philadelphia, 1 August 2014).
124 Dworkin, above n 105, at 179 (footnotes omitted).
310 Vol 23 (2017)
A Normative EvaluationofAlgorithmicLaw
Most of us, I think, would be dismayed by "checkerboard" laws that treat
similar accidents or occasions of racial discrimination or abortion
differently on arbitrary grounds. Of course we do accept arbitrary
distinctions about some matters: zoning, for example. ... But we reject a
division between parties ofopinion when matters of principle are at stake.
For Dworkin, checkerboard laws are repugnant because they offend the
political virtue of integrity, a value that stands alongside fairness and
justice.1 2 5
Integrity commands consistency in the principles underlying the
law. Arbitrary law lacks integrity "because it must endorse principles to
justify part of what it has done that it must reject to justify the rest".' 26
Through integrity, equality before the law is substantively protected by
commanding consistency in values:1 2 7
We insist on integrity because we believe that internal compromises
would deny what is often called "equality before the law" and sometimes
"formal equality." ... The equal protection cases show how important
formal equality becomes when it is understood to require integrity as well
as bare logical consistency, when it demands fidelity not just to rules but
to the theories of fairness and justice that these rules presuppose by way
ofjustification.
Algorithmic law might lack integrity because of inconsistency in principle.
Tailoring laws allows for values to be qualified and applied differently to
different individuals. The laws provided for some are also laws rejected for
others.
However, such inconsistency might also be justified as having
integrity. Dworkin describes checkerboard laws as "internal compromises"
ofprinciple that exist arbitrarily or to resolve political division.1 28
It might be
reasoned for algorithmic law that seemingly checkerboard laws exist
because of integrity in principle. Consider a law that issues the death penalty
only to educated and wealthy persons. This might be seen as inconsistent in
the sense that the death penalty is both endorsed as a criminal sanction and
rejected as immoral. But if the motivating principle is deterrence of crime at
all costs, a death penalty might be consistent! On this view, if the death
penalty was restricted to cases where deterrence was expected to be
successful (perhaps educated, wealthy persons), this would be consistent
with legal integrity.
In this way, the legitimacy of unequal categorisation returns to the
territory of justification. Checkerboard laws are a useful analogy of
undesirable law. However, the definition of such law hinges on arbitrariness,
and so whether algorithmic law can be said to be checkerboard (and offend
legal integrity) will depend on the force of justification for such a
categorisation in law.
125 At 183-184.
126 At 184.
127 At 185.
128 At 178-179.
311
Auckland UniversityLaw Review
(c) Equality Conclusion: Trade-off with Accuracy
Algorithmic law must balance the accuracy of its predictions with equality
before the law. Porat and Strahilevitz argue that, in the context of default
rules, accuracy is more important:1 29
... because gender and race can be reliable predictors of current
preferences and future behavior, entirely excluding these variables from
an algorithm leaves a great deal of predictive power on the table. Most
people would probably prefer an algorithm that knows their race and
gender and, as a result, more accurately predicts their preferences...
This argument is likely to be unpersuasive beyond opt-out default rules. For
law with more significant consequences, some accuracy will need to be
sacrificed to prevent categorisations that are discriminatory or
objectionable. 130
It is difficult to find a clear explanation of permissible inequality in
law that holds when applied to algorithmic law. Concepts such as
justification and objectionable quickly fall away or become circular when
applied beyond the situations we instinctively feel are acceptable or
repugnant.
However, implicit in these criteria seems to be the broader idea that
people should not be unfairly judged on something they cannot change. This
is the product of two variables: the degree of impact a law has on a person;
and how easy it is for that person to move between the groups the law
distinguishes. As technology allows tailored groups to be more precisely
defined, there is an increasing need for the law to catch up and develop
clearer principles of what exactly is required by equality before the law.
2 CertaintyofLaw
Legal certainty is the second aspect of the rule of law that could be
undermined by algorithmic law.
(a) Rationale for Certainty
The rationale for certainty in private law is best described by Friedrich
Hayek's argument for rules ofjust conduct. According to Hayek, rules ought
to be independent of ends but provide certainty of means - protecting the
"recognizable private domains" of individuals and enabling spontaneous
order.' 3
Certain law facilitates transactions by providing information for
decision-making and clarifying "the particular things [individuals] can count
129 Porat and Strabilevitz, above n 5, at 1467 (footnotes omitted).
130 O'Neil, above n 4, at 210.
131 FA Hayek "The Principles ofa Liberal Social Order" (1966) 31 IL Politico 601 at 603.
312 Vol 23 (2017)
A NormativeEvaluationofAlgorithmic Law
on being able to use" in the market.1 3 2
For individuals to be free and rational
actors, they need to be able to make long-term decisions - a prerequisite of
which is certain law.
More broadly, certainty of law also protects against arbitrary
discretion and abuse of public power. In criminal law, this is reflected in the
expectation that offences are prospective and citizens are provided with "fair
warning".133
Non-publication or unavailability of a statutory instrument may
be a defence to a crime that is committed unknowingly.' 3 4
It was the importance of legal certainty that led Jeremy Bentham to
call for widespread codification of law.' 35
The grand utility of the law is certainty: unwritten law does not-it
cannot-possess this quality; the citizen can find no part of it, cannot take
it for his guide; be is reduced to consultations-he assembles the
lawyers-he collects as many opinions as his fortune will permit; and all
this ruinous procedure often serves only to create new doubts
However, despite algorithmic law being heavily codified, uncertainty might
arise from the plurality of tailored laws, changes in law over time and issues
relating to the accessibility of law.
(b) Uncertainty Due to the Plurality ofTailored Laws
Porat and Strahilevitz discuss uncertainty in the context of personalised
default contract rules, where they focus on the uncertainty resulting from the
numberof tailored default rules.
A system of tailored laws, calibrated "too finely along a continuous
range", might increase uncertainty.136
An individual will need to invest effort
to learn the particular laws that apply to them. While a system of uniformly
applied law also requires an individual to learn the relevant rules, these
learning costs are generally reduced if the rules also apply to everyone
else.
"
However, tailoring might make it easier for individuals to anticipate
the law and increasecertainty. In the context of default contract rules, Porat
and Strahilevitz observe that an individual "alreadyknows a great deal about
[their] preferences and characteristics, which are the factors driving the
choice among multiple personalized default rules".' 3 8
The same point is
made by Ben-Shahar and Porat in regard to personalised negligence law:'39
132 FA Hayek Law, Legislation and Liberty: The Mirage of Social Justice (University of Chicago
Press, Chicago, 1976) vol 2 at 123.
133 AP Simester and WJBrookbanks PrinciplesofCriminalLaw (4th ed, Brookers, Wellington, 2012)
at 28-29.
134 At 28-29.
135 Jeremy Bentham The Works ofJeremy Bentham (William Tait, Edinburgh, 1843) vol 3 at 206.
136 Ben-Shahar and Porat, above n 36, at 675.
137 Porat and Strahilevitz, above n 5, at 1458.
138 At 1458 (emphasis in original).
139 Ben-Shahar and Porat, above n 36, at 675.
313
Auckland UniversityLaw Review
... it often would be easier for injurers to anticipate personalized standards
than uniform ones, because they know more about their own
characteristics than about the general distribution of characteristics in
society.
Someone who knows they are prone to risk-taking will also know that they
are likely subject to more restrictive laws and can assume they need to take
greater precautions to stay within the law. Technology further enables the
certainty of tailored law. For instance, micro-directives that communicate
legal commands to an individual in real time leave no doubt about the
required course of action for complying with the law.
Finally, the success of algorithmic law requires certain, accessible
law. If it is too expensive for an individual to determine the law, they will
instead accept liability equal to the average harm. Kaplow makes this
argument in regard to standards that are too complex. If it is too expensive
for an individual to determine their liability, they will instead pay the
average of the harm caused rather than try to follow the standard - and the
law will not achieve any change in behaviour:1 4 0
In this instance, their behavior will be the same under both formulations
of the law. At the enforcement stage, applying the complex standard will
be more costly. But this will be a waste, because behavior will not be
improved by avoiding over and underinclusiveness. As a result, the
simple rule would be superior. Achieving a better fit between the law and
behavior is accomplished only if individuals are induced to conform their
behaviour to the legal norm.
For these reasons it is likely that the uncertainty caused by the plurality of
tailored laws is minimal, and would be offset by the promulgation potential
for algorithmic law.
(c) Uncertainty Due to Dynamism
The dynamic aspect of algorithmic law also has the potential to increase
uncertainty in law. Inherent in the prediction models that prescribe law is a
need for ongoing refinement. This means that law might vary from one day
to another and gradually update to better align with the desired goal.
Bruno Leoni describes how variations in law over time produce
uncertainty. Legislation produces "short-run" certainty of the law, but in the
long-term the ability to repeal legislation causes some base level of
uncertainty in law:14
1
The certainty of the law, in the sense of a written formula, refers to a state
of affairs inevitably conditioned by the possibility that the present law
may be replaced at any moment by a subsequent law. The more intense
and accelerated is the process of law-making, the more uncertain will it be
that present legislation will last for any length of time.
140 Kaplow, above n 61, at 592.
141 Leoni, above n 107, at 80.
314 Vol 23 (2017)
A NormativeEvaluationofAlgorithmic Law
This uncertainty also exists with algorithmic law, which, like normal law,
can be repealed. However, algorithmic law has additional uncertainty in that
its application might vary from day to day. While a predictive algorithm
must be codified, the future data that it will use to refine its predictions
cannot be known in advance. Therefore, an individual cannot know for
certain which tailored category they will fall into and what the applicable
law will be. Individuals must predict the results of the legal algorithm by
guessing future data.
The uncertainty of algorithmic law is less pronounced than it first
seems. An established prediction model is unlikely to be heavily influenced
by new data, given the amount of previous data that has already been
incorporated into the model. Change is likely to be gradual and infrequent,
responding to changing behaviours over time. However, for long-term
planning, even gradual change is significant. In addition, there is always a
risk that some external, behaviour-changing event will sharply redefine the
law.
Uncertainty is most pronounced when dynamism is combined with
tailored categories along lines of past choices. An individual, making a
choice, cannot know for certain the category in which their choice will place
them in the future, particularly when the weighting given to past choices is
indeterminate.
(d) Accessibility of Algorithmic Law
Even if algorithmic law is precisely codified and does not change over time,
there is still a concern that such law might lack cogency. Bingham discusses
certainty of law within the larger idea of accessibility of law.142 Legal
certainty fits within a broader need for the governed to understandthe law in
order for it to have effect. This idea is reflected in comments made by Lord
Mansfield in 1761:143
The daily [negotiations] and property of merchants ought not to depend
upon subtleties and niceties; but upon rules, easily learned and easily
retained, because they are the dictates of common sense, drawn from the
truth of the case.
Algorithmic law might both enhance and suppress different aspects of law's
accessibility. On one hand, predictive algorithms are complicated and their
workings are beyond the understanding of most people (including lawyers!).
It might be difficult to discern the principles of the law or anticipate how a
predictive algorithm will operate in the future.
However, the ability of algorithmic law to be queried and respond
with a simple and clear directive holds merit. Laws are already complex and
beyond many people's understanding. Bingham describes a 2008 United
142 Bingham, above n 96, at ch 3.
143 Hamilton v Mendes (1761) 2 Burr 1198 at 1214, 97 ER 787 at 795 as cited in Bingham, above n
96, at 38.
315
Auckland UniversityLaw Review
Kingdom Court of Appeal case in which, on the eve ofjudgment, the parties
and judges discovered subsequently-enacted regulations which rendered the
debated regulations no longer relevant.'" While noting the lack of a statute
search engine, Lord Justice Toulson made the Kafkaesque observation that
"the courts are in many cases unable to discover what the law is". 145
Although better legal databases and software can exist
independently of algorithmic law, the necessity of these for functioning
algorithmic law aids legal accessibility. That an individual can query facts
and be issued with precise commands would make law significantly more
accessible for those currently unable to navigate the complex network of
statutes and case law. Whether this produces the same effect as "rules, easily
learned and easily retained" 46 is unclear. Lord Mansfield might be dismayed
at the deterioration of "common sense" principles, but perhaps also pleased
that common sense no longer must be a necessary heuristic for legality,
given the query-able ability of algorithmic law.
Freedom and Rights
The second critique of algorithmic law offered relates to individual freedom
and rights. The purpose of algorithmic law is to guide individuals (through
precisely calibrated rules and sanctions) to make choices that align with an
overarching societal goal. In this way, algorithmic law can constrain an
individual's freedom: by tying outcomes to the behavioural predictions of
other similar individuals; and by the past choices made by the individual
themselves.' 4 7
Wright and De Filippi describe such a system of law as
limiting the "realm of choices" available to an individual and presenting the
"illusion of free will".14 8
1 SubjugationofFreedom to the MajorityInterest
The algorithmic narrowing of law subjugates individual freedom to the
majority interest - a point best explained through an analysis of rights.
Rights are the mechanism that entitles individuals to freedom, whether this
be the privilege to act in a certain way, the power to exercise authority or a
passive entitlement to be free from interference by others.1 4 9
Law, then, is
the means of allocating rights across society. 150
Two competing ideas guide this allocation of rights: instrumental
theories and status theories. Instrumental theories view law as "an instrument
of social policy", where rights are tools for optimising society-wide
144 R v Chambers[2008] EWCA Crim 2467.
145 At 28 as cited in Bingham, above n 96, at 42.
146 Hamiltonv Mendes, above n 143, at 1214.
147 See Sunstein, above n 46, at 52.
148 Wright and De Filippi, above n 26, at 43-44.
149 Leif Wenar "Rights" (9 September 2015) Stanford Encyclopedia of Philosophy
<https://plato.stanford.edu> at [2.1] and [4].
150 At [4].
316 Vol 23 (2017)
A Normative EvaluationofAlgorithmicLaw
objectives."' Rights are justified because, in aggregate, they achieve good
consequences. By contrast, status theories take a deontological perspective
of rights. Law expresses rights that do not depend on instrumental value.'5 2
Rights reflect some human characteristic - they exist because "people
actually have them".15
An instrumental theory is presented in the work of John Stuart Mill,
who viewed rights as tools that could be used to bring the greatest utility to
society. For Mill, restrictions on freedom are justified according to the harm
principle. 154
The sole justification for interference with freedom is "selfprotection":
decisions that impose harm on others may be restricted by
law.'55
Thus, for Mill, restricting freedom under the harm principle is
consistent with freedom as a right as this maximises and upholds the
aggregate freedom of individuals in society.
As algorithmic law is better calibrated than traditional law to the
actual consequences of an act, it can better uphold the harm principle and
optimise freedom in society. Those predicted to harmlessly perform a
(previously) illegal act would be allowed to do so, while those who would
legally impose harm on others would have to bear the precise costs of that
harm.
However, algorithmic law can undermine the freedom of particular
individuals in the way that it is tied to predictedharms. Even highly accurate
predictions are generalisations that will not hold true for every individual.1 5 6
A prediction that is 99 per cent accurate will statistically uphold the harm
principle - minimising aggregate harm - and be efficient for lawmakers to
act on, but it will also undermine the freedom of the one per cent whose
decisions are wrongly predicted. People in that one per cent are then
subjugated to the average of the group of people who have similar attributes.
It has been argued this subjugation already exists, to a greater extent,
under normal rules that are uniformly applied. With these rules, a broader
group is subject to a less accurate average:
... any default rule, impersonal or personalized, is statistical in nature
because it assigns rights and duties to individuals according to the
averaged preferences of an entire population or a subset of people.
Personalized default rules are just a better proxy-based on more accurate
statistics-for the preferences ofthe specific party.
While this is true, the full political and practical picture of algorithmic law
must be considered.
151 Richard A Posner "Instrumental and Noninstrumental Theories of Tort Law" (2013) 88 Ind LJ 469
at 469.
152 At 469.
153 Warren Quinn MoralityandAction (Cambridge University Press, Cambridge, 1993) at 173 as cited
in Wenar, above n 149, at [6].
154 John Stuart Mill On Liberty (Yale University Press, New Haven, 2003) at 80.
155 At 80.
156 Porat and Strahilevitz, above n 5, at 1461.
157 At 1461-1462.
317
Auckland UniversityLawReview
Politically, algorithmic law increases the potential for law to be
more restrictive of individual freedom. First, this is because tailoring ensures
that oppressive restrictions only apply to few individuals and will not affect
many voters. But, secondly, the idea that an algorithm determines a labelled
group to be risky - for example, suspected terrorists - adds legitimacy to
oppressive restrictions. By tailoring laws to a denounced group, restrictions
on freedom might become more politically palatable. For the majority of
people determined to be in a group unaffected by restrictions on liberty, the
law may appear to be working correctly.
Practically, algorithmic law can allow for restrictions on freedom
that would otherwise be unable to exist. Consider an algorithmic law that
heavily restricts property rights for those with a high propensity to commit
theft. These rules would be undesirable to apply to the entire population
the
effect on commerce would be severe. It is only through tailoring that
such carve outs of property rights can be made, targeting thieves (and
subjugating the rights of some wrongly identified people) while preserving
the benefit of the right for the rest of society.
For an instrumental theorist, this subjugation is fine. If law tailored
to predictions on aggregate reduces harm, it is of no concern that the
freedom of some is subjugated. Under this conception, rights exist to the
extent that they benefit the collective.
Yet the idea that algorithmic law allows seemingly fundamental
rights to be so readily diced and repackaged across individuals is
confronting. Instrumentalists might have a defence to this concern by
adapting how the outcome is measured - for instance, by seeking other
goals, like egalitarianism. However, the criticism is not defeated. It is only
from a status theory perspective that a true defence to the optimisation of
algorithmic law is raised.
Consider the contemporary status theory of Robert Nozick, viewing
rights as grounded in individual liberty and the Lockean "state of nature".' 58
For Nozick, rights comprise a "moral space around an individual", a
boundary that is prohibited to be crossed.1 59 The core assumption is that any
"boundary crossing" infringement on rights can only be permitted if it is
duly compensated.1 6 0
This, it is observed, is particularly relevant for law
made on the basis of risk. Nozick gives the example of a law prohibiting
epileptics from driving to reduce the risk to other road users.' 6
1Forbidding
any particular person from driving "may not actually lessen the harm to
others" - it is an assumption of risk made by a prediction of how a person
might behave. 6 2
Therefore, a person who faces restrictions imposed for
reasons of risk ought to be compensated by those who benefit from that
reduction of risk. For algorithmic law, the core assumption is the opposite.
While it might be possible for a system of compensation, the assumed
158 See Robert Nozick Anarchy, State, and Utopia(Basic Books, New York, 1974) at 10-12.
159 At 57.
160 At 57-84.
161 At 78 and 79.
162 At 78.
318 Vol 23 (2017)
A NormativeEvaluationofAlgorithmic Law
position is that interference with rights is an entitlement of the lawmaker
a
means to some optimal end.
Casey and Niblett view the potential for algorithmic law to
"overreach" and restrict freedom as presenting "real concerns for individual
autonomy". 163 Traffic lights could "decide who goes first based on
productivity", and law might "dictate what a citizen is allowed to eat for
breakfast".16 Laws might be founded on predictions made on the basis of
genetic information. 165
However, Casey and Niblett argue this is an objection
to "reckless lawmaking" rather than any feature of algorithmic law.166 The
reality is that such "reckless" law-making occurs only from within the view
of the instrumentalist, optimised freedom that they are advocating. The
functional optimisation of law leads down a dangerous path, warned as being
potentially "totalitarian" by Wright and De Filippi.1 67
As algorithmic law allows rights to be reshaped in new ways, there
will be a greater role for status theorist protections of freedom. Something
feels instinctively wrong with the state pre-emptively controlling behaviour
(and making rights contingent) on the basis of a prediction. Such an
approach undermines the very autonomy that rights are intended to preserve.
As expressed by Quinn:1 6 8
We think there is something morally amiss when people are forced to be farmers or
flute players just because the balance of social needs tips in that direction. Barring
great emergencies, we think people's lives must be theirs to lead. Not because that
makes things go best in some independent sense, but because the alternative seems
to obliterate them as individuals.
2 FreedomandJustice:the Tyranny ofPastDecisions
In addition to limiting the freedom of those affected by incorrect predictions,
algorithmic law can limit freedom by ascribing high signalling value to an
individual's past decisions. Past choices take on a signalling value for
algorithms that determine future law, and the costs of harmful decisions
become disproportionate to the actualharm caused.
This is another form of subjugation to the majority.1 6 9
A prediction
made on the basis of an individual's past choices requires assumptions about
this behaviour to be made from the data of the past choices of others.
However, where past choices determine future applicable law for an
individual, this is also a justice concern.
The signalling value of an individual's past actions acts as a sanction
(or an incentive) for those actions. The dynamism of algorithmic law further
means that the exact signalling value of any act will be indeterminate - an
uncertainty which itself operates as a sanction. A prudent individual,
163 Casey and Niblett, above n 2, at 1443.
164 At 1443.
165 Ben-Shahar and Porat, above n 36, at 683-684.
166 At 52.
167 Wright and De Filippi, above n 26, at 43.
168 Quinn, above n 153, at 171.
169 1 discussed this above in terms of rights.
319
Auckland UniversityLaw Review
realising the potential for a decision to restrict future options in unforeseen
ways, will choose the safer option to reduce the risk of negative signalling.
From an instrumentalist viewpoint of rights this is mostly fine. The
ongoing freedom of an individual ought to be limited by the signalling value
of the individual's previous decisions, because, on aggregate, this achieves
good outcomes. Risky people should face more legal restrictions. There may
be some concern that the indeterminacy of algorithmic law will make
individuals overcompensate and be too cautious; however, this can be offset
in the calculation of the signalling value itself. There will also be some need
to encourage risky behaviour where there is a corresponding utility - for
example, by limiting the extent that a bankruptcy can be used for signalling,
to not deter individuals from starting a business.
However, a status perspective of rights would see law made on the
basis of signalling as being an uncompensated interference with rights - a
punishment (or reward) for past decisions that is disproportionate to such
decisions. This is best explained with reference to corrective justice.
Corrective justice is a theory of justice in private law which
maintains that a wrong creates a transactional inequality in rights and
establishes a relationship between the wrongdoer and the wronged. 70 Such a
view of justice then mandates restoring the original equality between parties
by correcting the imbalance in rights - restoring the wronged party to their
rightful position.17
Corrective justice subscribes to a status theory of rights.
Rights are deontological. They are a Kantian unification of free individuals
in society - the "sum of the conditions under which the choice of one can
be united with the choice of another in accordance with a universal law of
freedom".1 72
The pure, personal rectification of a transactional inequality
under corrective justice does not take place when a wrong is also used to
signal more restrictive laws into the future. For example, the sanction for a
breach of tort becomes the cost of putting the victim back in the same
position and the cost of future, more restrictive laws.
When future sanctions are prescribed from past decisions, "feedback
loops" that further narrow individual freedom can occur. Feedback loops can
be explained by reference to social media "echo chambers", in which a
person is narrowed into only seeing similar views to their own.1 7 3
An
algorithm takes a person's past information - the content they interact with,
their friend group, the news sources they read - and predicts their interests.
It then presents back an "informational universe that is entirely selfselected",
further narrowing content choices into the future.' 74
O'Neil gives
an example of a feedback loop that is created when employers use credit
score algorithms to evaluate potential employees.'17
Individuals with low
170 Ernest J Weinrib The Idea ofPrivateLaw (Oxford University Press, Oxford, 2012) at 62-63.
171 Ernest J Weinrib "Corrective Justice in a Nutshell" (2002) 52 UTLJ 349 at 349.
172 Immanuel Kant Die Metaphysik der Sitten (1797) (translated ed: Mary Gregor (translator) The
Metaphysics ofMorals (Cambridge University Press, Cambridge, 1996)) at [6:230].
173 Cass R Sunstein Republic.com 2.0 (Princeton University Press, New Jersey, 2007) at 43-44.
174 Sunstein, above n 46, at 50.
175 O'Neil, above n 4, at 7.
320 Vol 23 (2017)
A Normative EvaluationofAlgorithmic Law
credit scores are less likely to find work, making it even harder for them to
improve their credit score.176
It is not difficult to see how algorithmic law can then also have this
effect. Consider Uber, a personal driving service where passengers rate their
driver. A driver falling below a particular average will have their account
deactivated.177
Similar scoring models could be widely implemented in law.
For example, imagine a rating system for surgeons (combining objective
surgery results with patient feedback), where a surgeon with a low rating is
restricted to performing low-risk surgeries. This system would be beneficial
for patients, who would, on aggregate, face less risk. However, a surgeon
with a low rating would have a reduced opportunity to gain experience
actively
preventing them from increasing their rating in the future.
Under such a system, the costs of making the wrong choice can
become exponentially high and disproportionate to the harm caused. These
concerns are reflected in criticisms of risk-based sentencing - that certain
groups in the community can be marginalised.' 7
8A person who makes a bad
decision will not only bear the retributive consequences for that decision, but
suffer consequences from the signalling value about the kind ofperson they
are, sent to algorithms that affect their life. We might accept this for
insurance,1 79
but to what extent is this acceptable in law? The law might
move from a general position of proportional consequences to a position of
harsh limitations on freedom for some, because of how it has been signalled
they will act.
To some degree, a right to be free of the past is inherent in freedom.
Individuals need to face consequences for their decisions, but in a
proportional manner that allows them the opportunity to move on and make
different decisions. It was this idea that saw the enactment of the Criminal
Records (Clean Slate) Act 2004 to recognise that, after seven law-abiding
years, some past convictions ought to be hidden from the record. As stated
by Hon Phil Goff in the parliamentary debates, these people "have met the
penalty that was imposed on them" and "should be given a fair go to put that
offending behind them".' 80
Ideas like corrective justice have a role to play in
the proliferation of algorithmic law - ensuring that consequences for
actions remain proportionate and are not entirely functional to advance some
wider, unrelated goal.
176 At7.
177 "Driver Deactivation Policy - Australia & New Zealand Only" (28 June 2016) Uber
<www.uber.com>. Uber only look to the most recent 500 trips to give the driver "the chance to
improve [their] rating over time".
178 Discussed at the beginning of Part V.
179 See Human Rights Act 1993, s 48. We might equally not accept this for insurance!
180 (4 May 2004) 617 NZPD 12566.
321
Auckland UniversityLaw Review
VI CONCLUSION
Algorithmic law might be viewed as law and economics taken to the furthest
extent. It is the optimisation of function, enabled by technology. Such
optimisation inevitably calls for an appraisal of the legal values it might
subvert. To what exact societal cost should rights be protected for any one
individual? What is the extent of permissible inequality before the law?
Such questions go to the very heart of legal philosophy and cannot
be easily answered. Rather, this article has attempted to define algorithmic
law as a discrete category of law - and to reconcile the different scholarship
advocating for this law. It has then, against the background of arguments in
favour of algorithmic law, presented preliminary critiques of algorithmic law
with reference to normative values of rule of law and freedom.
Robin Paul Malloy describes the law and economics theory of
wealth maximisation as "analogous to the Frankenstein Monster".181 It can
"lead even the most noble and good spirited people to conclusions that
affront our basic social values" - an observation accepted by Posner, who
observed that moral concerns may override wealth maximisation (but that
moral philosophy was a "weak field" unable to obtain consensus).1 82
For
Malloy, the values of freedom, liberty and human dignity are important, and
"something more than the cold hard variables of economic calculus":18 3
While economics, capitalism, and democracy are all important factors in
helping us understand and promote human dignity, we cannot escape our
own personal moral obligation to make each of these factors subservient
to the higher values they can promote ... freedom, liberty, and human
dignity.
With algorithmic law, the usual assumptions that restrain the wealthmaximisation
of law and economics do not apply. It becomes possible for
efficient law to be made by personalising rules to each individual. Laws can
be far more intrusive on human autonomy and channel a person down a path
of restricted choices. Laws can undermine aspects of law that might have
previously been unconsciously assumed as inherent.
Because of this, it will be important to sometimes place "higher
values" above "cold hard" outcome-producing variables. This is a view well
expressed by the following quote from Milton Friedman:1 84
A free society, I believe, is a more productive society than any other; it
releases the energies of people, enables resources to be used more
effectively, and enables people to have a better life. But that is not why I
am in favor of a free society. I believe and hope that I would favor a free
181 Robin Paul Malloy and Jerry Evensky Adam Smith and the Philosophy of Law andEconomics
(Kluwer Academic Publishers, Dordrecht, 1995) at 160.
182 At 160 and 170.
183 At 160.
184 Milton Friedman "Free Markets and Free Speech" (1987) 10 Hary JL & Pub Poly 1 at 7 as cited in
Malloy and Evensky, above n 181, at 160.
322 Vol 23 (2017)
A NormativeEvaluation ofAlgorithmicLaw 323
society even if it were less productive than some alternative-say, a slave
society. ... I favor a free society because my basic value is freedom itself.
There are areas of law where algorithmic law might readily undermine core
values and be undesirable. Criminal law is one example. In other areas, such
as regulatory law, algorithmic law might be more appropriate. However, as
this article has argued, algorithmic law should be approached with caution.
That there is significant potential for society to benefit from efficient,
predictive law should not mean that the master of statistics and economics
can reduce to a mere number the higher values that law can promote.