<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Archiving and Interchange DTD v1.0 20120330//EN" "JATS-archivearticle1.dtd">
<article xmlns:xlink="http://www.w3.org/1999/xlink">
  <front>
    <journal-meta />
    <article-meta>
      <title-group>
        <article-title>The Justice Access Game: Crowd-Sourced Evaluation of Systems for Pro Se Litigants</article-title>
      </title-group>
      <contrib-group>
        <contrib contrib-type="author">
          <string-name>Karl Branting</string-name>
          <xref ref-type="aff" rid="aff0">0</xref>
        </contrib>
        <aff id="aff0">
          <label>0</label>
          <institution>The MITRE Corporation</institution>
          ,
          <addr-line>McLean, VA</addr-line>
          ,
          <country country="US">USA</country>
        </aff>
      </contrib-group>
      <abstract>
        <p>A variety of computational tools and systems have become available in recent years to assist pro se (self-represented) litigants in formulating claims and defenses in legal forums. However, there has been little exploration of techniques for evaluation of these systems to establish their relative efectiveness and identify specific strengths and weaknesses of each approach. This paper proposes the Justice Access Game, a crowd-sourced method for comparative evaluation of legal information systems. The Justice Access Game is a framework applicable to wide range of domains and systems.</p>
      </abstract>
      <kwd-group>
        <kwd>eol&gt;Human-computer interface</kwd>
        <kwd>evaluation</kwd>
        <kwd>crowd-sourcing</kwd>
        <kwd>pro se litigants</kwd>
      </kwd-group>
    </article-meta>
  </front>
  <body>
    <sec id="sec-1">
      <title>1. Introduction</title>
      <p>
        evaluation [6]. As a result, it is dificult to measure the
performance of alternative technical approaches in terms
Worldwide, a significant proportion of litigants appear either of accuracy, usability, efectiveness, or success rate.
pro se, i.e., without the assistance of an attorney, in courts Instead, relative performance is typical estimated
anecand other decision forums [
        <xref ref-type="bibr" rid="ref1">1</xref>
        ] [11]. Pro se litigants are dotally.
usually at a significant disadvantage in legal proceedings This paper describes the Justice Access Game (JAG), a
compared to parties represented by an attorney [
        <xref ref-type="bibr" rid="ref5">5</xref>
        ]. general, domain-independent approach intended to
ad
      </p>
      <p>
        Development of approaches to assisting pro se litigants dress the challenges of evaluating legal assistant systems.
is an area of active research and exploration. One ap- The next section describes the design of the JAG,
Secproach has been to replace conventional decision bodies, tion 3 sets forth the pragmatics of the game, the possible
which are often extremely challenging for non-experts to role of Large Language Models is discussed in Section 4,
understand and navigate, with alternative dispute resolu- and a summary and proposal for future work is set forth
tion forums, such as ODR platforms, which are designed in Section 5.
to be less disadvantageous to pro se litigants [
        <xref ref-type="bibr" rid="ref3">3</xref>
        ] [15]. In
the context of conventional courts, online information
and forms provided by courts and legal aid institutions 2. Game Structure
are the most widely used information source, but most
pro se litigants struggle to understand and efectively use The Justice Access Game is method for comparative
evalthese resources [
        <xref ref-type="bibr" rid="ref4">4</xref>
        ] [9]. uation of systems to assist self-represented individuals.
      </p>
      <p>
        Currently, the most common alternatives to simple Players compete to identify a legal claim, identify all
online form and instruction resources automated legal facts supporting the claim, and assert the claim via a
leassistant systems consist of some form of guided form- gal filing. Points are awarded for achieving these goals
iflling software, sometimes implemented as a chatbot and for incorporating all relevant but no irrelevant facts.
[12], in computable logic [14], or combining these ele- Points are reduced based on errors and the duration or
ments with predictive models [17]. In the United States number of conversational turns required to achieve the
and Canada, many such systems have been developed goal. While individual participants compete for points,
through a collective process involving communities of alternative systems are compared via the average number
attorneys and other public-interest advocates [16] [7] of points that competitors are able to achieve using those
[
        <xref ref-type="bibr" rid="ref2">2</xref>
        ]. Unfortunately, informal development processes of systems.
this type have not shown themselves to be conducive to
rigorous system validation, vericfiation, maintenance, or 2.1. Procedure
In each round of a game, a player is provided a scenario
involving multiple issues and individuals, is assigned a
role and a goal, and is provided access to the system to
be evaluated. The system can be a logic-based or
guided
      </p>
      <p>After being provided with a factual scenario and
assigned a role and a goal, the player must accomplish the
following tasks using the system:
1. Determine what legal action would achieve the</p>
      <p>goal, if successful
2. Identify all facts necessary for each element of</p>
      <p>the legal action
3. Draft a claim or petition setting forth the claim</p>
      <p>and all elements needed to support the claim
At the conclusion of each round, the player is assigned a
score based on a weighted combination of the following
factors:
1. Is the legal action correct, i.e., if the claim
suc</p>
      <p>ceeded, would it achieve the player’s goal?
2. Were the elements required for the action
identi</p>
      <p>ifed?
3. Were all the facts relevant to those elements
iden</p>
      <p>tified?
4. Were there extraneous facts?
5. How many steps or how long did the process</p>
      <p>take?</p>
      <p>Systems are ranked based on the average number of
points that competitors are able to achieve using each
system.</p>
      <sec id="sec-1-1">
        <title>2.2. Example Game</title>
        <p>Consider an example game. Suppose that the factual
scenario were as follows:</p>
        <p>Under this set of facts, a player could be assigned the
role of Mary, John, or Bill. Each of these roles is associated
with diferent set of possible legal goals, as shown in
Table 1.</p>
        <sec id="sec-1-1-1">
          <title>Role</title>
          <p>Mary
John
Bill</p>
        </sec>
        <sec id="sec-1-1-2">
          <title>Goal</title>
          <p>Keep John out of their apartment
Keep John from being alone with Nell.</p>
          <p>Prevent eviction
Return to apartment.</p>
          <p>Spend time alone with Nell.</p>
          <p>Prevent eviction, if he is allowed
to return to apartment
Evict John and Nell.</p>
          <p>Keep John away from Bill</p>
          <p>For example, if a player were given a certain system,
assigned the role of Mary, and given Mary’s Goal 1
(keeping John out of the apartment), the player’s performance
using the system would be scored based on the the
following:
• Identifying that Goal 1 could be achieved by a</p>
          <p>protection order
• Listing requirements for a protection order:
– Domestic relation (cohabiting and having</p>
          <p>a child together)
– Actual or threatened harm (injury to Nell)
• Completion of all fields of a petition for a
protec</p>
          <p>tion order as Mary</p>
          <p>Facts
John and Mary started dating in September 2017, and in
January 2018 they moved into an apartment that they A fact pattern, together with a single role and a
sinrented from Bill. Baby Nell was born in 2019. Starting in gle goal constitutes a problem. A fact pattern together
2021 the radiator in the apartment began failing during with the sets of roles and goals arising from that fact
cold snaps. John and Mary complained about the cold, pattern, constitutes a problem set. The example above
and without telling Mary, John stopped paying rent in illustrates that a single problem set may include problems
February 2021 to protest the condition of the apartment. in multiple areas of law.</p>
          <p>When John and Mary received an eviction notice in
March, Mary became very angry at John. In the ensuing
argument, a window in the apartment was broken and 2.3. Scoring
Nell received a cut on the forehead, which Mary blamed
on John. When Bill came down to investigate the noise
and saw the broken window, John threatened to hurt
Bill unless he fixed the heat. Bill left the apartment
and called the police, but John left before the arrival
of the police, who listened to the arguments and left
without taking any action. Mary refuses to let John in
the apartment or see Nell. Bill insists that Mary leave
the apartment immediately. John refuses to pay the rent
until Bill fixes the radiator and insists that he has a right
to live in the apartment with Mary and to spend time
alone with his daughter.</p>
          <p>As described above, the scoring of each individual
problem is based on correctness, completeness, accuracy, and
duration. The Justice Access Game does not depend on
any particular procedure for performing this scoring. The
most informative scoring would be individual ratings by
experts in the particular areas of law involved in the
problem set, e.g., legal-aid attorneys or law professors. A
more practical alternative would be a heuristic match to
model solutions, e.g., ROUGE scores [8] or similarity in
semantic-embedding space [10] between each solution
element and the corresponding element of the model
solution.</p>
        </sec>
      </sec>
    </sec>
    <sec id="sec-2">
      <title>3. Pragmatics</title>
      <p>3.1. Game Requirements and Preparation
Applying the Justice Access Game requires that the
systems to be compared share a common legal domain. For
example, a chatbot could be compared with a
guidedpathway system only if both were designed to address a
common legal problem.</p>
      <p>A more challenging pragmatic issue is that a model
solution is needed for each problem. Justice Access Game
problems are intended to be representative of the
challenges routinely faced by pro se litigators, so the solutions
should be straightforward for legal-aid attorneys or law
professors. However, preparing all the solutions for a
entire problem set could require an investment of many
hours.</p>
      <sec id="sec-2-1">
        <title>3.2. Game Participants</title>
        <p>Participation in the Justice Access Game is intended to
approximate the experience of actual pro se litigants,
who typically have little or no experience with judicial
proceedings. Appropriate participants could therefore
include anyone with limited legal experience, e.g.,
Mechanical Turk crowd-workers, participants in the NSF
Timesharing Experiment,1 or college students. By
contrast, law students would probably not be appropriate.
For the same reason, players should not be given multiple
problems involving the same area of law.</p>
      </sec>
      <sec id="sec-2-2">
        <title>3.3. Baselines</title>
        <p>The key objective of the Justice Access Game is to
empirically establish the efectiveness of alternative paradigms
for assisting pro se litigants. Accordingly, it is
appropriate to start with the most widely used baselines. The
ifrst would consists of court forms plus the Internet, the
resources traditionally available to pro se litigants in a
typical court. A second proposed baseline would consist
of the online form technologies currently promoted by
Legal Services Corporation, e.g., Court Forms Online2 or
Maryland Court Help.3</p>
      </sec>
    </sec>
    <sec id="sec-3">
      <title>4. Large Language Models</title>
      <p>The capabilities of Large Language Models (LLMs), such
as ChatGPT [13] have recently advanced dramatically.
When JAG problems are posed to these systems, they
are capable of providing advice that is mostly useful and
accurate, but occasionally disastrously incorrect. In an
1https://www.tessexperiments.org/
2https://courtformsonline.org/
3https://perma.cc/F654-KMNE
example session, when ChatGPT (3.5) was given the
example fact pattern from Subsection 2.2 above and was
asked what advice it would give if it were a legal aid
attorney, it provided the following response:</p>
      <p>Mary can take legal action to keep John
out of the apartment by obtaining a
protective order or a restraining order.</p>
      <p>A protective order is a court order that
prohibits an abuser from having
contact with the victim and may require the
abuser to stay away from the victim’s
home or workplace. In Virginia, a
victim of domestic violence, such as Mary,
can apply for a protective order from the
court.</p>
      <p>To obtain a protective order, Mary would
need to file a petition with the circuit
court in the county where she lives or
where the abuse occurred. She would
need to provide detailed information
about the abuse, including dates and
specific incidents, and explain why she needs
protection. She should also provide any
evidence she has, such as medical records,
photographs, or witness statements, to
support her allegations.</p>
      <p>If the court finds that Mary is in danger
of further abuse, it may grant a
protective order that prohibits John from having
any contact with her or coming within a
certain distance of her and her daughter.</p>
      <p>Violating a protective order is a criminal
ofense and can result in John being
arrested.</p>
      <p>Alternatively, Mary can also seek a
restraining order, which is a court order that
prohibits someone from coming near or
contacting the protected person. In
Virginia, a restraining order can be obtained
through the Juvenile and Domestic
Relations District Court.</p>
      <p>Mary should consider consulting with a
local attorney who can help her assess
her options and determine the best course
of action for keeping John out of the
apartment and protecting herself and her
daughter.</p>
      <p>When asked to draft a sample protection order, the result
was the following:</p>
      <p>As an AI language model, I’m not
authorized to provide a completed legal
document as it requires legal expertise and
access to specific information about the
case. However, I can provide an example
of the information that Mary may include
in a Virginia petition for a restraining
order based on the facts provided:
[Header information: Virginia Circuit
Court, [Name of County], Civil Case No.
[insert case number]]
PETITION FOR A PROTECTIVE ORDER
My name is [Mary], and I reside at
[address]. The person I am seeking
protection from is [John], and his last known
address is [address]. John and I were in
a relationship from September 2017 to
March 2021. We have a daughter together,
named [Nell], who was born in 2019. In
February 2021, John stopped paying rent
on our apartment to protest the condition
of the apartment. Despite my requests to
pay the rent, John refused to do so,
resulting in us receiving an eviction notice in
March. On [date], John and I had an
argument about the eviction notice, during
which a window in the apartment was
broken, and Nell received a cut on her
forehead. I believe that John broke the
window and caused the injury to our
daughter. When our landlord, Bill, came down
to investigate the noise and saw the
broken window, John threatened to hurt Bill
unless he fixed the heat. Bill left the
apartment and called the police, but John left
before their arrival. I am in immediate and
present danger of family abuse by John,
and I am seeking a protective order to
ensure my safety and the safety of my
daughter. I request that the court grant a
protective order that prohibits John from having
any contact with me or my daughter and
from coming within a certain distance of
us. [Signature of petitioner, Mary]
Again, this is just an example of the type
of information that Mary may include in
a petition for a restraining order in
Virginia. It is important to note that legal
documents such as this require a high
degree of legal expertise and accuracy, and
it is recommended that Mary consult with
a local attorney for assistance in drafting
and filing the petition.</p>
      <p>It seems almost certain that we stand on the threshold
of an era in which pro se litigants will increasingly turn to
LLMs for legal advice, notwithstanding any disclaimers
such as those appearing above, against providing legal
advice. The Justice Access Game may provide one
mechanism for assessing the quality and usefulness of this
advice to pro se litigants.</p>
    </sec>
    <sec id="sec-4">
      <title>5. Summary</title>
      <p>This paper has described a new crowd-sourced method
for comparative evaluation of legal information systems,
the Justice Access Game. The Justice Access Game
provides a framework applicable to wide range of domains
and systems and can provide a realistic assessment of the
capabilities of systems to assist pro se systems in a
realistic setting. This paper has briefly set forth the details
of the Justice Access Game and provided two example
games. It is hoped that this work will be of value in
understanding the capabilities of emerging technologies,
including not just extensively engineered solutions such
as the JusticeBot [17] but also rapidly changing ad hoc
capabilities typified by ChatGPT [13].</p>
    </sec>
    <sec id="sec-5">
      <title>Acknowledgments</title>
      <p>I gratefully acknowledge the assistance of Treyvon
Jordan and Stephen Steward in developing Justice Access
Game scenarios. The MITRE Corporation is a
not-forprofit company, chartered in the public interest. This
document is approved for Public Release; Distribution
Unlimited. Case Number 23-1584. ©2023 The MITRE
Corporation. All rights reserved.</p>
      <p>A. Justice Access Game Example 2
Psychology, Public Policy, and Law 26(2), 198–212
(2020)
[6] Lauritsen, M., Steenhuis, Q.: Substantive legal soft- Facts
ware quality: A gathering storm? In: Proceedings Tracy let an ex-boyfriend, Mark, borrow her car after
of the Seventeenth International Conference on Ar- many assurances that it would be a quick trip “down the
tificial Intelligence and Law. pp. 52–62. ICAIL ’19, block”. An hour later, Tracy received a call from Mark
ACM, New York, NY, USA (2019) that he was pulled over by the police under a suspended
[7] LawHelp Interactive (2023), https:// driver’s license. The police advised Mark that he was
lawhelpinteractive.org/, last visited 15 April traveling through a known drug corridor as the reason
2023 for the stop. Mark was not aware that his license was
sus[8] Lin, C.Y.: ROUGE: A package for automatic eval- pended prior to borrowing the car. Mark called the DMV
uation of summaries. In: Text Summarization and was informed that his license was suspended due
Branches Out. pp. 74–81. Association for Compu- to the lack of insurance coverage on his vehicle. Mark
tational Linguistics, Barcelona, Spain (Jul 2004), had switched insurance providers recently, and believed
https://aclanthology.org/W04-1013 this is the reason for the error. Mark told Tracy that the
[9] Macfarlane, J.: The national self-represented liti- vehicle was impounded by the police department and
gants project: Identifying and meeting the needs of was told that it was a 30-day impoundment and would
self-represented litigants: Final report. CanLIIDocs cost $300 for release. Tracy was stunned because she
p. 493 (2013) only had one vehicle and not having one will
substan[10] Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efi- tially afect her ability to work as an Uber driver. Tracy
cient estimation of word representations in vector had recently moved into her current apartment and after
space. CoRR abs/1301.3781 (2013), http://arxiv.org/ incurring expenses for new furniture, a security deposit,
abs/1301.3781 and other costs associated with moving, a $300 loss would
[11] Moore, S., Nwebury, A.: Legal aid in crisis: Assess- have a significant impact on her family’s finances. James,
ing the impact of reform. Bristol University Press, Tracy’s husband (who thinks they overpaid on moving
ex1st edn. (2017), https://doi.org/10.2307/j.ctt1t8988q penses and furniture) became angry after hearing about
[12] Morgan, J., Paiement, A., Seisenberger, M., the additional $300 expense for the vehicle’s return.
AfWilliams, J., Wyner, A.: A chatbot framework for ter the argument, James asked for a divorce, took their
the children’s legal centre. In: Palmirani, M. (ed.) three-year-old son, and left the apartment.
Legal Knowledge and Information Systems -
JURIX 2018. Frontiers in Artificial Intelligence and</p>
      <p>Applications, vol. 313, pp. 205–209. IOS Press (2018)
[13] chatGPT, https://platform.openai.com/
[14] Pyle, J.: Docassemble (2019), https://docassemble.</p>
      <p>org/ last accessed on APril 19, 2023
[15] Schmitz, A., Zeleznikow, J.: Intelligent legal tech
to empower self-represented litigants. Columbia
Science and Technology Law Review pp. 142–190
(March 2 2022)
[16] Steenhuis, Q., Colarusso, D.: Digital curb cuts:
Towards an inclusive open forms ecosystem. Akron</p>
      <p>Law Review (August 25, 2021)
[17] Westermann, H., Walker, V., Ashley, K., Benyekhlef,</p>
      <p>K.: Using factors to predict and analyze
landlordtenant decisions to increase access to justice.</p>
      <p>In: Proceedings of the Seventeenth International
Conference on Artificial Intelligence and Law. p.
133–142. ICAIL ’19, Association for Computing
Machinery, New York, NY, USA (2019)
Tracy
James
James</p>
      <p>Goal
Retrieve the vehicle
before the 30-day deadline
Avoid paying for the
vehicle’s return or be
compensated for doing so
Divorce Tracy
Obtain Joint-Legal
Custody of Son
If DC-499 is too lengthy a
process, pay for the
vehicles return, and sue Mark
for the losses. This can be
done in small claims court
by filing a DC-402.</p>
      <p>If no fault, submit a VS-4
Report of Divorce or
Annulment
Either the required forms
in J&amp;DR Court or file
complaint from scratch to be
included with the divorce
in Circuit Court
Provide the DMV with a
Certification of Insurance
Pay the DMC
reinstatement fees
Submit a Petition for
Proceeding in Civil Case
Without Payment of Fees
or Costs to the General
District Court</p>
      <sec id="sec-5-1">
        <title>Elements</title>
        <p>Prove you did not know
the ofender’s license was
suspended.</p>
        <p>Prove that the immediate
family only has one
vehicle and will sufer
hardship if not returned
Demonstrate that Mark is
the cause of Tracy’s injury
(the costs incurred by the
vehicle’s impoundment).
Having been separated
for at least a year
Plead complaint in
alignment with the
“information considered in
custody proceedings.”
Provide the DMC with a
Certification of Insurance
Pay the DMC
reinstatement fees
Provide net income,
assets, and expenses to the
court.</p>
      </sec>
    </sec>
  </body>
  <back>
    <ref-list>
      <ref id="ref1">
        <mixed-citation>
          [1]
          <string-name>
            <surname>Cerniglia</surname>
            ,
            <given-names>C.E.</given-names>
          </string-name>
          :
          <article-title>The civil self-representation crisis: The need for more data and less complacency</article-title>
          .
          <source>Georgetown Journal on Poverty Law and Policy</source>
          <volume>27</volume>
          (
          <issue>3</issue>
          ),
          <fpage>355</fpage>
          -
          <lpage>388</lpage>
          (
          <year>2020</year>
          )
        </mixed-citation>
      </ref>
      <ref id="ref2">
        <mixed-citation>
          [2]
          <string-name>
            <given-names>CLEO</given-names>
            <surname>Connect</surname>
          </string-name>
          (
          <year>2023</year>
          ), https://cleoconnect.ca/,
          <source>last visited 15 April 2023</source>
        </mixed-citation>
      </ref>
      <ref id="ref3">
        <mixed-citation>
          [3]
          <string-name>
            <surname>Himonas</surname>
            ,
            <given-names>D.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Hubbard</surname>
            ,
            <given-names>T.</given-names>
          </string-name>
          :
          <article-title>Democratizing the rule of law</article-title>
          .
          <source>Stanford Journal of Civil Rights &amp; Civil Liberties</source>
          <volume>16</volume>
          (
          <issue>2</issue>
          ),
          <fpage>261</fpage>
          -
          <lpage>282</lpage>
          (
          <year>2020</year>
          )
        </mixed-citation>
      </ref>
      <ref id="ref4">
        <mixed-citation>
          [4]
          <string-name>
            <surname>Knowlton</surname>
            ,
            <given-names>N.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Cornett</surname>
            ,
            <given-names>L.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Gerety</surname>
            ,
            <given-names>C.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Drobinske</surname>
          </string-name>
          , J.:
          <article-title>Cases without counsel: Research on experiences of self-representation in u.s. family court</article-title>
          .
          <source>Tech. rep.</source>
          ,
          <article-title>Institute for the Advancement of the American Legal System (June 8,</article-title>
          <year>2016</year>
          ), https://iaals.du.edu/ sites/default/files/documents/publications/cases_ without_counsel_research_report.pdf
        </mixed-citation>
      </ref>
      <ref id="ref5">
        <mixed-citation>
          [5]
          <string-name>
            <surname>Kroeper</surname>
            ,
            <given-names>K.M.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Quintanilla</surname>
            ,
            <given-names>V.D.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Frisby</surname>
            ,
            <given-names>M.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Applegate</surname>
            ,
            <given-names>N.Y.A.G.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Sherman</surname>
            ,
            <given-names>S.J.</given-names>
          </string-name>
          ,
          <string-name>
            <surname>Murphy</surname>
            ,
            <given-names>M.C.</given-names>
          </string-name>
          :
          <article-title>Underestimating the unrepresented: Cognitive biases disadvantage pro se litigants in family law cases</article-title>
          .
        </mixed-citation>
      </ref>
    </ref-list>
  </back>
</article>