System 1 and System 2 on Legal Reasoning

February 15th, 2012

Dual-Process Cognition and Legal Reasoning by Anna Ronkainenapplies insights from Danny Kahneman’s Thinking Fast and Slow to the process of legal reasoning. It looks fascinating. Here is the abstract:

The dual-process framework is a set of theories on human cognition in which cognition is seen as consisting of (at least) two substantially different yet interdependent systems: the older, faster, partly unconscious and automatic System 1 and the newer, slower, fully conscious and considered System 2. When viewing legal reasoning through the dual-process model, we can easily see that System 1 is primarily responsible for deciding a case (or finding the best line of arguments in support of a party) with the help of aligning the particulars of the case with the preexisting framework of statute and case law, whereas System 2 is responsible for generating and evaluating arguments in support of the outcome determined by System 1, thereby opening up an individual’s reasoning process for external critique. System 2 may also override System 1 altogether, but this is only possible in easy cases. In part thanks to the dual-process framework we can take a scientific look into the often discussed but substantially neglected question of Right Answers in law through empirically testable hypotheses. This also has significant implications for artificial intelligence and law. By acknowledging the differences between the two, we can better use the most suitable computational models for each of them individually.

This article ties together so many of my interests (and I am thrilled to see someone else has connected them!)

From the article, on the people’s inability to think of legal reasoning in terms of System 1 and System 2:

Unfortunately, this aspect of legal thought has so far not been explained
satisfactorily by legal theory. In part, this failure is caused by a general
tendency to see System 2 as the only form of human reasoning and therefore
to only try to explain its functioning in System 2 terms. This however is
symptomatic of an even broader problem, namely a general refusal to see
legal reasoning as something performed by actual flesh-and-blood humans
rather than something preferably only taking place in the abstract realm of
ideas. Dworkin himself with his superhuman Judge Hercules as the paragon
of legal reasoning is not without blame in this respect, either.

When studying human reasoning within the legal
domain, we are free to consider the actual practice of legal reasoning, even
with all its flaws, as the gold standard. Legal theory by itself is not in a
position to show conclusively that some particular precedent was decided
incorrectly, as much as certain legal theorists might want it to be. Even in
the case of the most unquestionably dysfunctional before/after-lunch-type
variation, other actual cases must be invoked in order to argue that some
particular decision is incorrect.

On System 1:

The function of System 1 in the process of legal reasoning is straightforward:
it is responsible for seeing the facts of the case specifically as legal facts and
then translating those facts into a tentative decision regarding the outcome
of the case. This outcome can be, in the mind of a judge, the actual verdict,
or, in the mind of an advocate, the most favourable line of argument for
advancing the case of one’s party. Once the big picture is clear, the meaning
of the details become more apparent.
System 1 processing is massively parallel. As such, it is impossible to
fully describe System 1 processing in terms of individual rules, whether for
the purposes of identifying individual rules (on the System 1 side) used
in arriving at some particular decision or to make a ‘dump’ of the entire
rule base used by some individual reasoner. System 1 processing can be
understood in terms of heuristics, but it is important not to confuse the
two. Heuristics can be produced as explanations of System 1 processes in
System 2 terms as more or less simplified representations of the underlying
System 1 processes. On the other hand, heuristics can also be learned as
conscious rules of thumb, in which case they work as rules in System 2. If
a heuristic of this type is a complete description of the problem domain, it
may also remain in System 2 only. If it is a simplification of a more complex
problem, however, it will be assimilated to the previous System 1 model of
the domain.

Potter Stewart was big on System 1 (also ties into social costs):

Another major problem with System 1 is subjectivity. System 1 is
subjective in the everyday sense: it varies from person to person at least to
some extent. Depending on the amplitude of this variation it can be seen
either as inevitable noise or as a harmful interfering signal which should
be suppressed. In any case this kind of variation should be considered
dysfunctional, the only question is whether eliminating it is necessary or
indeed even possible in practice. Even though some particular form of
variation might fall into the noise category as far as the system as a whole
is concerned, at least for the parties involved it is always significant.
Another and perhaps a more important side of subjectivity is however
that the massive parallelization makes it difficult to say why one chooses
A rather than B. System 1 is intersubjectively opaque, in the worst case
making it impossible to justify one’s line of reasoning in any kind of generally
understandable terms, as aptly evidenced by Justice Stewart’s famous
dictum regarding obscenity: ‘I know it when I see it.’18 In order to deal
with these problems, we must turn to System 2.

And on System 2:

If the outcome of a case is decided by this opaque System 1, does it mean
that judicial decisions are arbitrary and individual judges may decide cases
however they please? No. In this regard, there are two important safe

guards. First of all, we do not let just any System 1 make the decisions, but
rather almost always there are requirements as to the professional qualifications
of judges to ensure at least some degree of consistency. By itself this
is however not sufficient, as the well-known (but far from well-understood)
phenomenon of judicial temperament illustrates. And this is where System
2 takes over.
When System 1 has come up with a tentative outcome for the case at
hand and at the same time has aligned it with the legal system as a whole,
it becomes the task of System 2 to explain the details of that alignment.
That is, the role of System 2 is to generate legal arguments in familiar terms
of statute and precedent after the fact in support for a conclusion that
has already been reached. This process of justification has two important

You see, System 2 comes after System 1. A judge decides the outcome, and then backfills the reasoning. Most judges are totally oblivious this is even happening. This feeds strongly into my writings on social.

Firstly, this makes it possible to ensure that one’s decision is correct.
After all, there might be some statutory provision one’s System 1 has overlooked,
or a precedent in favour of a different outcome that is closer to
the present case than the one suggesting the tentative outcome one had in
mind. In these situations, the tentative outcome must be modified accordingly.
After eliminating the first tentative outcome, it is possible that only
one other alternative outcome remains, or the correct revised outcome has
otherwise become apparent in the justification process, or possibly System
1 must be consulted for a second (third…) time.
Secondly, justification opens up one’s line of thought for external critique.
Explaining one’s line of reasoning to others forces one to consider
aspects which might otherwise remain unconscious. There is always at least
some degree of blindness towards one’s own cognitive biases, so adding more
people to the equation helps, especially if there is the kind of diversity that
results in substantially differences in individual biases. In law, however,
this by itself is not enough. Legal decisions must have a basis in law, that
is, decisions must be made based on some specifically acknowledged legal
criteria whereas some other potentially influential criteria must explicitly
be ignored. In this respect, we can view the task of System 2 in the light

of the theory of legal argumentation as proposed by Alexy and Aarnio.19
The arguments used to justify a particular outcome in a case must live up
to the standards of rationality acknowledged by the narrow legal community
of expert professionals. As to specific points of law, these standards
must include some theory of sources of law, but that alone is not sufficient,
a substantial element of common-sense knowledge is always present
as well. System 1 may also respond predictably to many unorthodox forms
of argumentation, such as PROOF BY LOUDNESS20, which in some circumstances
may by themselves be sufficiently persuasive, but which do not
live up to agreed standards of rationality in argumentation and as such
cannot be used as justification.

So how do Systems 1 and 2 work together.

As described above, Systems 1 and 2 work together in the legal reasoning
process: System 1 produces the decision and System 2 produces the
arguments in support of the decision while also verifying its correctness.
This implies a particular temporal order in which System 1 has to act before
System 2. This claim should however be taken with two important

Firstly, even though the decision is what we are most interested in in law,
System 1 only has to make a decision, or perhaps more properly a decision
candidate, for System 2 to start doing its part, and multiple decision candidates
can and should be considered in parallel before committing oneself to
a particular one as the decision. In a typical case, from the perspective of a
judge, these decision candidates are provided (and thus a considerable part
of the decision-making process has already been dealt with) by the parties
of the case, and they, together with any other plausible decision candidates
should be given equal consideration from the beginning. At some point,
which can, depending on the case, happen very early or quite late in the
process, one does however have to pick one of them.

Secondly, some of the characteristics of System 1 processing make it
extremely difficult to pinpoint the moment at which this happens. This is
because System 1 processing is mostly unconscious and massively parallel.
That is, one may consider different alternatives equally at the same time
and not be aware of it. This is of course not a problem. What can be
a problem is making up one’s mind far too early in the process and then
interpreting all the remaining incoming information with a confirmation
bias, that is, in a way that is most favourable for the decision one has
already made.

She also discusses this in terms of System 1 as being emotional and System 2 as being rational based on brain imagery:

There is already also some support for the dual-process model in legal
reasoning from neuroimaging studies.25 There is always an element of
‘emotional’ (System 1) processing, but with lawyers, ‘rational’ (System 2)
processing takes over faster and to a greater extent than with non-lawyers.

At this point it is however far too early to make any substantial claims
based on imaging studies, in part because the studies so far have been designed
in terms of the ‘moral brain’ hypothesis rather than the dual-process
hypothesis, in part because of the general limitations of using imaging studies
to study cognitive processes combined with a problematic tendency to
make bold and unwarranted claims despite these limitations, often quite

Plus she ties this into AI:

As an AI & law researcher, my interest in legal theory arises from needs
which are very different from those of most other scholars of legal theory.
Considering the possibility of automated or computer-supported legal reasoning,
it is hardly realistic to expect that hard cases are the best place
to start. Instead, easy cases should, as the name suggests, be easier to
decide, and in them, computer-supported decision-making should prove to
be most beneficial both in terms of achievable economic efficiency and the
elimination of unwanted individual variation. In order to do this, a theory
of legal reasoning in easy cases is required. Unfortunately, since deciding
easy cases is indeed easy, at least for people, even some of the latest work
taking otherwise a reality-based approach38 fails to deliver in this respect.

Drawing from e-discovery practices, she proposes a starting point for legal decision-making:

I suggest that in light of the dual-process model of legal reasoning,
similar solutions can also be transferred from information retrieval to the
domain of legal decision-making. In this way it is possible to combine the
two competing dominant approaches in artificial intelligence: statistical
methods and rule-based methods. On the one hand, the idea of representing
law without rules just does not compute, at least as long as also people
are involved and have to trust the system. On the other hand, rule-based
approaches alone do not scale for the level of complexity encountered in reallife
problems, but rather the result becomes an unstable and unmanageable
pile of rules and exceptions on top of each other which may lead to serious
problems with novel cases not covered by the rule base

And she breaks down how a decision would be made:

A tentative outcome for a case could for example be derived from a
self-organizing map created from pleadings in previous cases within the
domain. This by itself is however not sufficient, as it only represents the
System 1 part of the process, and after that the System 2 part has to
take over. Once the most difficult aspects of the case have been settled,
aligning the facts of the case with some framework of arguments becomes
much easier. The System 2 part should of course also here check that
the outcome produced by the System 1 part is correct, and if it is not,
suggest that it is dealing with a case from outside its area of expertise
(possibly a hard case) and give up. It is most realistic to expect that
computerized decision-support systems would be most advantageous and
best implementable within domains with large numbers of cases which are
not being appealed, that is, presumably easy cases. Hard cases can wait,
and indeed they should wait until we understand the details of the legal
decision-making process much better than we do now. On the other hand,
it is neither necessary nor practical to require absolute perfection in terms
of performance before such a system can be deployed, at least in a court
of first instance, and for higher level courts the advantages of automated
decision support are much lower, anyway. After all, the whole court system
is built on the assumption that the first instance is not perfect, for why else
would we need courts of appeal at all?

And how to teach a computer to think like a lawyer?

From an ontogenetic perspective, on the other hand, the importance of
the professional legal education must be emphasized. Law school teaches
one to think like a lawyer. The development reasoning abilities especially
from the System 2 perspective has already been dealt with extensively in
the literature. Regarding System 1 processes specifically, it is clear that
the extensive use of cases plays a significant role. Cases, whether completely
made-up ones just for educational purposes, or actual precedents in
a simplified form relevant for the rule being taught, provide a temporary
structure for organizing the different facts around which a more elaborate

version can be built. The same phenomenon at much earlier stage has been
an object of study in the psychology of learning, where the currently preferred
technical term for such temporary structures is scaffolding.46 Scaffolding
is constructed in cooperation by the student and the teacher and it
provides a temporary frame of reference for understanding the legal issues
involved within some particular domain. Once the scaffolding has fulfilled
its purpose in supporting the learning process, it is discarded, with the
exception of actual precedents used in this manner, which are retained for
argumentative purposes.
The result of this process in Ross’s terms is a judicial ideology, that
is, ‘the normative ideology which animates the judge’47. Only viewing
it as a residual source of law used to decide hard cases gives however a
quite misleading picture of its function. In particular in hard cases, the
individual differences become apparent and even salient. This obscures the
fact that, by any account, more than 99% of the judicial ideology will be
identical across all individuals with a professional education within some
particular legal system, as otherwise any legal system would hardly be able
to function. What Ross really was trying to point out was that the judicial
ideology is the object of study for all scientific jurisprudence.48

This is a paper I would love to write.