Dispute Resolution Enhanced
How Arbitrators and Mediators Can Harness Generative AI
David L. Evans, Stacy Guillon, Ralph Losey, Valdemar Washington, and Laurel G. Yancey1
This article examines contemporary applications for generative artificial intelligence within the domain of arbitration and medi- ation and explores prospective applications, offering insights into potential future developments at the intersection of technology and dispute resolution.
Introduction2
The legal world has been buzzing with discussions about the potentials of artificial intelligence (AI), especially in its aid to advocates. However, a significant yet less discussed transforma- tion is unfolding within the oMces of arbitrators and mediators. This article, penned by the AI Working Group assembled by the American Arbitration Association® (AAA®), seeks to shine a light on this nuanced application of AI technology in the world of neutral decision-making and alternative dispute resolution (ADR). The authors are arbitrators serving on AAA's panels, and we have collectively put these tools to use in arbitrations and mediations, both as neutrals and as advocates.
While fears and speculations abound regarding AI's potential
to replace human judgment, our current perspective is not about
1 Together with AAA staff members, the authors, who are all AAA pan- elists, comprise AAA's AI Working Group. The content of this article was finalized in early November 2023.
2 In the spirit of our exploration, we used ChatGPT-4 to assist us in drafting this article's title, introduction, and conclusion, exemplifying the very blend of human expertise and machine capabilities we discuss in this article.
Dispute Resolution Journal® | May-June 2024, Vol. 78, No. 1, pp. 57–92.
Ⓢ 2024 American Arbitration Association®. All rights reserved.
ISSN 1074-8105 (print) |ISSN 25733-606X (digital).
replacement, but rather enhancement. For the arbitrator and mediator, generative AI stands as a promising tool to enhance eMciency, offer deeper insights, and provide a level of precision previously unattainable.
This article begins with an examination of contemporary applications for generative AI within the domain of arbitration and mediation, identifying specific purposes for which neutrals can use this technology to benefit the parties before them. We also compare various AI tools' effectiveness at these tasks. While OpenAI's ChatGPT may be the most widely recognized generative AI platform, there are others, including some that have been tailored for the legal industry. Then, we transition to prospective applications, offering insights into potential future developments in this intersection of technology and dispute resolution.
Current Uses
Large language models (LLMs) represent a remarkable advancement in the realm of generative AI. These models are trained on vast and diverse data sets sourced from a wide array of materials, including books, images, and web content. LLMs, such as the generative pre-trained transformers (GPT), serve as the foundation for platforms like ChatGPT and Anthropic's Claude. Generative AI's Natural Language Processing capacity means it can “understand,” interpret, and generate human language at an advanced level. Its level of “understanding” can be quite deep; it can identify patterns, relationships, and anomalies within document training sets. And, importantly, AI can perform those tasks at a herculean scale within a matter of minutes (or less).
1.Document Analysis and Comparison
Document Analysis Applications
Like most legal work, arbitrations and mediations can be document intensive, which creates enormous opportunities to use LLMs. Here, we offer a few ways in which arbitrators and
mediators can use AI to improve the quality and eMciency of
their work.3
Search and Index: At the most basic level—perhaps offering an ideal entry point for AI newbies—neutrals can use this feature to answer substantive questions about a large document set's con- tents. Because AI retrieves relevant content based on meaning, it is far more effective than a tool that simply finds exact word matches. For example, in preparing for a hearing, an arbitrator could upload the case documents and ask AI to comb them all for a particular fact. Some AI tools, as we discuss in more detail below, link directly to their sources within the document set, directing the neutral to the relevant page(s), so she can read the information firsthand. These tools will even compile the basis for their answers from various sources when the subject matter has been addressed in multiple exhibits and filings. AI's speed allows a neutral to use these tools in real time; for instance, by typing in a question raised about an expert's report during a hearing to navigate directly to the responsive section rather than flipping through pages or having to pause the hearing to ask for page numbers.
Compare Content: These tools can also analyze the sim- ilarities and differences between positions taken in various documents. An arbitrator could prompt the tool to identify evi- dentiary inconsistencies across multiple deposition transcripts, or a mediator could prepare for a session by asking the tool to identify the parties' overlapping interests.
Contract Review: Some AI platforms have been specially trained to analyze contracts. With these tools, neutrals can home in on contract provisions that govern the dispute resolution process, even if that text is scattered throughout the contract.
Summarize: AI tools can summarize documents with varying levels of detail. Neutrals can put this to use to create outlines that guide neutrals to key sections without requiring them to spend (and bill) time digging through extraneous text. For instance, a mediator could generate a detailed summary of a preliminary injunction hearing transcript to grasp eMciently the procedural
3 We are not addressing the separate issues of disclosure or indepen- dence associated with the use of AI by neutrals. We are identifying use cases.
history of the case. Or, where counsel have pointed to excerpts of a deposition, an arbitrator could generate a comprehensive sum- mary of the full deposition transcript to direct him to additional sections he should read before deciding a substantive motion.
Timeline Generation: During an arbitration hearing or a mediation, many neutrals like to refer to timelines. When coun- sel did not provide them, we used to have to manually identify, compile, and order the various dates. But, after reviewing the entire document set, AI can digest that information for us and generate a timeline of the case's events.
Platform Comparison: Document Summary
This section compares query results from the Generative AI models Casetext's CoCounsel,4 Anthropic's Claude Pro,5 and OpenAI's ChatGPT-4.6
First, we compare CoCounsel's and Claude's summaries of the proposed “Restricting the Emergence of Security Threats that Risk Information and Communications Technology Act” or “The RESTRICT Act.”7 Then, we compare CoCounsel's and ChatGPT-4's summaries of an Intellectual Property Cross-Li- cense Agreement.
CoCounsel vs. Claude: Summarizing
CoCounsel Summary—The RESTRICT Act
CoCounsel acts as an AI legal assistant and has been iden-
tified as a generative AI model currently used in arbitration.8
4 Casetext's CoCounsel 2.0 Beta, October 2023 version, https://casetext
.com/v2/cocounsel. Casetext provided AAA's AI Working Group, including the authors of this report, CoCounsel subscriptions in conjunction with our exploration of using AI in arbitrations and mediations.
5 Anthropic's Claude Pro, October 2023 version, https://claude.ai.
6 OpenAI's ChatGPT-4 Pro, October 2023 version, https://chat.openai
.com.
7 The RESTRICT Act, S.686, available at www.congress.gov/118/bills/ s686/BILLS-118s686is.pdf.
8 Andrea Seet et al., Arbitration Tech Toolbox: Looking Beyond the
Black Box of AI in Disputes Over AI's Use, Kluwer Arbitration Blog, May
From the CoCounsel Skills options, the Summarize Skill was chosen. As shown in the screenshot below, the skill generates three levels of detail: (1) Brief Summary (High level overview of document in 1-3 paragraphs), (2) Detailed Summary (Multi-page detailed summary that includes all important information from the documents), or (3) Comprehensive Summary (Page by page summary of the document, for comprehensive overview). The Comprehensive Summary was chosen.
CoCounsel produced a 16-page Comprehensive Summary report of The RESTRICT Act, a 55-page document. The Compre- hensive Summary listed pages 1 to 55 with each page providing defined key items in three to four bulleted paragraphs.
Key points excerpted from the summary:
The bill, S. 686, would authorize the Secretary of Commerce to review and prohibit certain transactions between the U.S. and foreign adversaries. The bill further clarifies the definition of “covered transaction” to include current, past, or potential future transactions.
25, 2023, arbitrationblog.kluwerarbitration.com/2023/05/25/arbitration- tech-toolbox-looking-beyond-the-black-box-of-ai-in-disputes-over-ais-use/.
The bill also defines “entity” broadly to include a variety of organizations, such as firms, governments, labor unions, and corporations. The bill defines “holding” as an equity interest, stock, security, share, partnership interest, interest in a lim- ited liability company, membership interest, or any participation, right, or other equivalent. The bill defines “information and communications technology products or services” as any hardware, software, or other product or service that enables information or data processing, storage, retrieval, or communication. The bill defines “transaction” broadly, encompassing any acquisition, importation, transfer, installation, dealing in, or use of any information and communi- cations technology (ICT) product or service. The document also outlines civil forfeiture procedures, which allow for the seizure and forfeiture of any prop- erty used to commit or facilitate a violation, as well as any proceeds from the violation. The document outlines criminal forfeiture procedures, which include the forfeiture of any property used to commit or facilitate a violation, as well as any proceeds from the violation.
Claude Summary—The RESTRICT Act
When prompted to summarize the Act and report results, Claude produced a one-and-one-half-page summary.
Key points excerpted from the summary:
The Restrict Act is a proposed bill that would authorize the Secretary of Commerce to review and prohibit certain transactions between persons in the United States and foreign adversaries that pose risks to US national security. Some key points:
It defines “covered transactions” broadly to include any transactions with foreign adversar- ies that could pose national security risks. This
includes China, Russia, Iran, North Korea, Cuba,
and Venezuela.
The Secretary of Commerce can prohibit or impose mitigation measures on covered transactions that pose undue risks. This includes risks of sabotage, effects on critical infrastructure, election interfer- ence, or other national security risks. The Secretary can also identify “covered hold- ings”—controlling holdings in ICT companies by foreign adversaries—that pose risks and refer them to the President for potential divestment or other mitigation measures. The bill establishes procedures for reviews, enforcement, and judicial oversight of the Sec- retary's prohibitions or mitigation measures. However, the scope of judicial review is limited. It aims to address national security risks from foreign ownership and control of ICT and critical infrastructure companies. The goal is to authorize stricter prohibitions and divestment requirements for covered transactions and holdings.
In summary, The Restrict Act would significantly expand the government's ability to block or mitigate foreign investment in US ICT companies that is deemed to pose national security risks. It reflects broader concerns about foreign influence in critical technology sectors.
Observations About CoCounsel’s and Claude’s Summary Functions
Both CoCounsel and Claude summarized the provisions of The RESTRICT Act, underscoring salient points applicable to pro- posed prohibitions on certain transactions. Each model reported results with eMciency. The speed of results from the Claude model was quicker than that of CoCounsel. The difference may be attributed to the selection of the CoCounsel Comprehensive Summary skill that reports an extensive page-by-page summary of the document for a comprehensive overview. Regardless of the speed of output, both generative AI models produced
results that would inform neutrals and enhance their specialized expertise, where stakeholders identify that “specialized/expert decision-making” is a top benefit of arbitration.9
The CoCounsel report outlined in more specific detail the terms, requirements, and sanctions of The RESTRICT Act as those sections relate to information and communications technol- ogy products and services transactions. The proposed sanctions are relevant to the eMcacy of arbitral processes. “Increasingly, US sanctions are giving rise to commercial disputes (many of which are subject to mandatory arbitration agreements) and even beginning to interfere with the smooth functioning of ongoing arbitration proceedings.”10
CoCounsel vs. ChatGPT-4: Summarizing and Analyzing
Here, we compare the abilities of CoCounsel and ChatGPT-4 when prompted to summarize and analyze a complex 15-page document, “Intellectual Property Cross-License Agreement,”11 dated June 30, 2023, which was selected at random for the test from the Edgar Archives.
The upload process for CoCounsel was simple and straight- forward, unlike the upload process for ChatGPT-4, which was difficult and complex. OpenAI is working on that, but as of October 2023, they still have a long way to go. Here is a screen- shot of the CoCounsel upload page. As you can see, they make it very easy.
9 Gary Benton et al., Cost Is the Top Tech Litigation Problem, Survey
Shows, Silicon Valley Arbitration and Mediation Center, 2017, https://svamc
.org/wp-content/uploads/SVAMC-2017-Survey-Report.pdf.
10 Anton A. Ware and Soo Mi Rhee, United States: Economic Sanctions And Arbitration: Are We Ready?, Mondaq, May 3, 2022, https://www.mondaq
.com/unitedstates/export-controls-trade-investment-sanctions/1189170/
economic-sanctions-and-arbitration-are-we-ready.
11 https://www.sec.gov/Archives/edgar/data/908255/000110465923 079180/tm2320751d1_ex10-4.htm.
We started by asking both CoCounsel and ChatGPT-4 to pre- pare a detailed summary of the License Agreement. As described above, the CoCounsel software offers a choice between a brief, detailed, or comprehensive summary of a document. ChatGPT-4 does not have such guidance, or limitation, depending how you look at it. You can ask anything about a document, once you finally get it uploaded.
Both AIs provided a good, detailed summary, but, in actuality, that was not too helpful for this document. We could skim the standard terms and conditions of the document manually in the same time and learn as much. It was mostly familiar boilerplate that we had seen many times before.
A shorter summary of a document would be more helpful for a busy attorney or arbitrator. So, next we asked CoCounsel for a Brief Summary. Here is the result:
This document is an Intellectual Property Cross-License Agreement between four companies. It defines key terms such as “Intellectual Property,” “Licensed IP,” “Licensee,” “SpinCo Field,” and “SpinCo Licensed IP.” The agreement outlines the license grants to SpinCo Licensee and Parent Licensee, as well as the rights and restrictions related to sublicensing, reservation of rights, and assignment of licensed IP. It also addresses intellectual property own- ership, the prohibition against challenging ownership,
and the prosecution, maintenance, and enforcement of intellectual property rights. The agreement covers confidentiality, disclaimers of warranties, limitations of liability, and indemnification. It also outlines the term and termination of the agreement, and contains miscel- laneous provisions such as the entire agreement clause, third-party beneficiaries, and notices.
We also prompted ChatGPT-4 for a brief summary. Here is the result:
The document is an “Intellectual Property Cross-License Agreement” dated June 30, 2023, involving multiple parties: BORGWARNER INC., DELPHI TECHNOLO- GIES IP LIMITED, PHINIA TECHNOLOGIES, INC., and BORGWARNER LUXEMBOURG OPERATIONS S.A.R.L.
Key highlights:
Definitions: The agreement meticulously defines key terms such as “AMliate,” “Confidential Information,” “Intellectual Property,” “Licensee,” and “Licensor,” among others. License Grants: Both Parent Licensors and SpinCo Licensors grant nonexclusive, perpetual, nontransfer- able, sublicensable, worldwide, royalty-free licenses to operate in specific fields. The licenses are also bound by certain limitations. Sublicensing: The agreement permits sublicensing under stringent conditions, including the protection of Confidential Information. Indemnification: Both Parent and SpinCo parties are obligated to indemnify each other against losses arising from breaches of the agreement. Miscellaneous Provisions: This section serves as a catch-all for various legal and procedural elements, including the entire agreement clause, third-party beneficiaries, and notices.
Interlocking with Separation Agreement: The document also refers to a separate “Separation Agree- ment,” and in the event of any conflict, the terms of this IP License Agreement will control.
The document is a comprehensive legal framework that outlines the scope, limitations, and responsibilities associated with intellectual property licensing between the involved parties. It also integrates terms from a separate “Separation Agreement,” making it a complex, interlocking legal document.
Would you like to delve into any other aspects?
The summary by ChatGPT-4 provided more information, but was still concise. We would give the edge this time to ChatGPT-4 over CoCounsel. Still, we do not think short summaries like this have much value for experienced practitioners, except perhaps for client consumption. Anyone with expertise in license agree- ments could dispense with this summary and jump right into the analysis—in the form of Q&A—about terms of the agreement of interest. Here, the AI summary and analysis abilities can be of real help.
To ask good questions, we needed to know more about the overall deal. We saw the license agreement was part of a larger corporate spin-off by BorgWarner.12 A quick Google search led to a business article with practical insights into the deal,13 which is the kind of information a lawyer and client would probably already have as a background to begin asking questions about
12 The license agreement was part of SEC Form 10-12B, Filing Date: May 18, 2023, found at http://pdf.secdatabase.com/536/0001628280-23-0188 28.pdf. This Intellectual Property Cross-License Agreement is Exhibit 10.4 to the 10-12B Registration disclosure, and is found at http://pdf.secdata base.com/536/0001628280-23-018828.pdf#%5B%7B%22num%22%3 A307%2C%22gen%22%3A0%7D%2C%7B%22name%22%3A%22XYZ%22
%7D%2C0%2C751%2C0%5D.
13 PHINIA Spin-Off Deep Dive (Stock Spin-Off Investing, June 30, 2023), https://stockspinoMnvesting.com/spin-offs/phinia-spin-off-deep-dive/.
a document. We wanted our Q&A to be as realistic as possible. The article on the Phinia deal stated:
BorgWarner (BWA) will spin-off [sic] PHINIA (PHIN), its Fuel Systems and Aftermarket segments, by the end of the day on July 3, 2023. To put it simply, the
rationale for the spin-off is to shed exposure to internal combustion engines and increase exposure to electric vehicles. Borgwarner is on track to meet or exceed its will spin-off goal of 25% of revenue from EV by 2025 . . . .
PHINIA end markets are perceived to be in secular decline.
With that background information, we could ask CoCounsel a useful test question, one that a practitioner might want to know in a transaction like this: “What information is specifically excluded from this assignment?” Here is the response.
Trademarks, power electronics, improvements
The document specifically excludes trademarks, service marks, trade dress, trade names, domain names, social media accounts or usernames, or other indicia of source or origin from the definition of “Intellectual Prop- erty” (pgs. 5,7), as well as power electronics applications for any electrified stationary applications or electrified mobility applications (pgs. 5,7). Additionally, improve- ments to the Licensed Parent IP or the Licensed SpinCo IP are not granted as part of this agreement (pg. 11), and any prior agreements or undertakings are superseded by this agreement (pg. 16).
We asked the same question of ChatGPT and received the following response.
The document contains several provisions that specify what is explicitly excluded or limited within the scope of the agreement. Here are some key points:
Third-Party Rights: The terms and conditions of the IP License Agreement, including the licenses granted,
are subject to any and all rights of and obligations owed to any third parties with respect to the Licensed IP under any contracts (Page 10-11).
No Other Rights or Obligations: Except as expressly set forth in the agreement, no right or license is granted, whether by implication, estoppel, or otherwise, under any intellectual property rights that any party or any of their aMliates may have (Page 11). Limitation of Liability: Neither party nor any of its aMliates will be liable for any consequential, inciden- tal, indirect, exemplary, special, punitive, or enhanced damages, or for any loss of actual or anticipated profits (Page 13-14). Provision of Information: The agreement specifies that no licensor or its aMliates are required to provide any written, electronic, computerized, digital, or other tangible or intangible media that has already been provided to, or is in the possession of, the licensee or its aMliates (Page 10-11). Assignment of Licensed IP: Any assignment of Licensed IP will include a transfer of any applicable license and similar contractual rights or permissions granted to the applicable licensee in this IP License Agreement (Page 10). Relationship of the Parties: The agreement states that nothing contained creates any agency, partnership, joint venture, or other forms of joint enterprise, employment, or fiduciary relationship between the parties (Page 17-18).
These exclusions and limitations are crucial for under- standing the scope and boundaries of the agreement, as they delineate the responsibilities and liabilities of each party involved. Would you like to delve into any other aspects?
Again, although both responses were good, and both pro- vided references to the agreement page numbers, we thought ChatGPT-4's response was more useful.
Finally, we asked what we thought was a key question for this deal, “Are hybrid vehicle parts covered by this agreement?”
CoCounsel's response:
No.
Hybrid vehicle parts are likely not covered by the
agreement, as the agreement specifically excludes power electronics applications for any electrified mobility appli- cations, which would include hybrid vehicles (pg. 7).
ChatGPT's response to the same question:
The document explicitly states that power electronics applications for any electrified stationary applications or electrified mobility applications, including any hybrid vehicles, are excluded from the SpinCo Field (Page 7).
In essence, hybrid vehicle parts are not covered by this Intellectual Property Cross-License Agreement. This exclusion is significant as it delineates the scope of the agreement and clarifies that the licensing does not extend to hybrid vehicle technologies.
Would you like to explore further?
Again, we preferred ChatGPT-4's answer, as it was more complete. Moreover, it did not qualify its answer to this straight- forward question with the phrase “likely not covered.”
Overall, both CoCounsel and ChatGPT-4 perform admirably well in summarizing documents and providing analysis.
A Note About Privacy
As using the document summary and analysis function of any of these tools involves uploading information, we want to address concerns about privacy. All three systems purport to protect the confidentiality of all documents submitted for analysis.
We are aware of many misconceptions that the paid version of ChatGPT offers no privacy protections, and we suspect those stem in part from conflating the free version of ChatGPT-3.5 and ChatGPT-4. Even in ChatGPT-4, however, the data control settings are off by default and must be turned on manually. Once
enabled, the privacy settings purport to provide for the auto-de- struction of all information after 30 days and allow OpenAI to use the information only “when needed to monitor for abuse.”14 When history is turned off, as it should be for any legal work, you will see this notice in the upper left corner of your browser screen.
CoCounsel has privacy automatically engaged. According to Casetext, data input into CoCounsel is “never used to train the AI model. The model is accessed through dedicated, secure servers. Data is encrypted in transit and at rest, and never stored by our AI partner. Clients retain all rights to their data. Your data is only used by Casetext to serve the product to your users.”15
Anthropic reports that Claude will only “use your prompts and conversations” if you give explicit permission, if they are “flagged for trust and safety review,” or “to improve [Claude's] ability to detect and enforce Acceptable Use Policy violations, including to train Trust and Safety classifiers in order to make our services safer.”16 Anthropic also “retains your personal data
14 Joshua J., Data Controls FAQ, OpenAI, https://help.openai.com/en/ articles/7730893-data-controls-faq.
15 CoCounsel, the First Legal Assistant: Security Overview, provided
directly to the authors on May 9, 2023.
16 I would like to input sensitive data into Claude Pro. Who can view my conversations?, Anthropic, https://support.anthropic.com/en/articles/ 8325621-i-would-like-to-input-sensitive-data-into-claude-pro-who-can- view-my-conversations.
for as long as reasonably necessary for the purposes and criteria outlined in our Privacy Policy.”17
Before uploading confidential data, we urge users to look into the privacy protections of the platform. We would not use any of these programs to summarize and ask questions about classified documents. But, for more standard confidential case documents, there are generative AI tools available that can satisfy many arbitrators' and mediators' privacy needs, especially when appropriate disclosures are made in advance.
2.Research
Research Applications
LLMs excel in combing and analyzing large quantities of data quickly—which, in the internet age, is the very essence of legal and factual research. Often, independent research is not a sig- nificant part of a neutral's work, including because it is often not economical. When used prudently, AI helps solve that problem. Although computerized legal research tools have assisted with legal research for years, generative AI is indeed a game changer. For example, AI's concept-based search produces more thorough and often more targeted results than platforms that do only a search for specified keywords. The user can ask an initial question and follow-up questions, allowing her to engage in a conversation with the bot in the same way she would a law firm
associate, honing the query as the thinking progresses.
For example, neutrals can use AI's research function to explore the law on a particular topic in preparation for a medi- ation, as the parties' mediation memoranda often do not brief all legal issues in detail. Where appropriate,18 arbitrators can review the legal contentions made in arbitration filings. And, though likely precluded in arbitrations,19 mediators can also use
17 How long do you store personal data?, Anthropic, https://support
.anthropic.com/en/articles/7996866-how-long-do-you-store-personal-data.
18 For purposes of this paper, the authors take no position on whether an arbitrator should or may conduct independent legal research.
19 See, e.g., American Arbitration Association Commercial Rule 36.
generative AI to verify the accuracy of claims or facts presented when publicly available information can be used to do so.
Platform Comparison: Legal Research
LLMs started off with something of a bad name in the context of legal research. Not only has there been prolific news coverage of the Roberto Mata v. Avianca Inc. lawyer's reliance on fake cases,20 judges and others have expressed skepticism about the careless use of AI in litigation, while in general still encouraging its use.21 However, different AI platforms offer different guardrails relevant to legal research, including to reduce hallucinations.22 This section compares ChatGPT-4's, Claude's, and CoCounsel's legal research functions.
ChatGPT-4's Legal Research
As the Avianca filing exemplifies, ChatGPT-3's answers to legal research questions, in these authors' experience, often contained citations to and quotes from cases that the software had fabricated (hallucinated).
ChatGPT-4 has been updated to disclose its own limitations. For example, in response to the prompt “Which First Circuit cases address circulation of religious images in public schools?,” ChatGPT-4 responded:
As of my last training data in January 2022, I cannot pro- vide an exhaustive or up-to-date list of all First Circuit cases specifically addressing the circulation of religious images in public schools. The First Circuit, like other federal appellate
20 Brandon Kimura, ChatGPT Ethics Case Summary, Blawg 401, August 11, 2023, https://blawg401.com/chatgpt-ethics-case-summary/ #:~:text=One%20of%20the%20first%20cases,The%20plaintiff%2C%20Mr.
21 See, e.g., retired Judge Ralph Artigliere, Ethical Guideposts for Law- yers Using Generative AI, EDRM, Oct. 31, 2023, https://edrm.net/2023/10/ ethical-ai-guideposts-for-lawyers-using-generative-ai/ (describing the Florida Bar's proposal to regulate attorneys' use of generative AI in a case).
22 Cade Metz, Chatbots May “Hallucinate” More Often Than Many
Realize, New York Times, November 6, 2023.
courts, has certainly addressed First Amendment issues in the public school context.
If you are specifically interested in cases from the First Circuit on this topic, it would be best to consult a legal database like Westlaw or LexisNexis. These databases provide up-to-date and specific case law on such nuanced topics. Alternatively, reaching out to a legal professional or scholar familiar with First Amendment issues in the First Circuit might provide more detailed insights.
Remember, case law evolves, and new decisions can impact interpretations and precedents. Always ensure that the information you're using is current and relevant to your specific context.
When we framed the question differently,23 however, ChatGPT-4 did perform legal research. In response to the prompt “Can New Hampshire public schools circulate religious imag- ery?,” GPT-4 offered a substantive response:
Public schools in the United States, including those in New Hampshire, are subject to the Establishment Clause of the First Amendment, which has been interpreted to
23 If AI tools are adopted, training on how to properly construct prompts/ queries and how to evaluate any results is necessary. Akin to Boolean searches that require some knowledge of how to construct a “good” search, AI tools require “good” prompts. One advantage of generative AI prompts and responses is that the tool has “thread” conversations. A person can ask clari- fying questions. Users can ask the AI tool to clarify previous responses or ask the AI tool to customize the tone or persona of the response. Training on com- pliance with confidentiality concerns, as well as considerations involving bias is recommended. See Hon. Xavier Rodriquez, Artificial Intelligence (AI) and the Practice of Law, 24 Sedona Conf. 783 at 817-18 (Sept. 2023), https://the sedonaconference.org/sites/default/files/announcements/Artificial-Intelli gence-and-the-Practice-of-Law-Xavier-Rodriguez_1.pdf. Jus Mundi offers training to arbitrators to educate themselves on how the tool works and its fea- tures, including Jus-AI. Part of the training includes the limitations of the tool and its benefits. See AI in Arbitration: Insights from Jus-AI's Creators, Daily Jus, July 20, 2023, daily.jusconnect.com/in-conversation-with/2023/07/ ai-in-arbitration-insights-from-jus-ais-creators.
mean that schools may not promote or endorse any par- ticular religion. The Supreme Court has ruled that public schools cannot sponsor religious activities or symbolism because they are government institutions.
Circulating religious imagery, if it is part of a curriculum that is educational and not devotional in nature, such as in a world religions or art history class, might be permissible as long as it is presented in an objective, non-proselytizing manner. However, using religious imagery in a way that could be seen as pro- moting or endorsing a particular religion would likely be unconstitutional.
The precise legal boundaries can sometimes be complex and subject to interpretation, and local school policies may also provide additional guidance on these matters. It's often best for school administrators to con- sult with legal counsel to ensure that they comply with the relevant constitutional and statutory requirements.
When we asked for the cases that informed ChatGPT-4's answer, it identified and very cursorily described six of the United States Supreme Court's Establishment Clause cases24 but then wrote, “If you need details or have a specific situation in mind, consulting legal experts or reviewing recent court decisions would be valuable.”
Claude's Legal Research
User Prompt: “Can New Hampshire public schools cir- culate religious imagery?”
24 Engel v. Vitale, 370 U.S. 421 (1962); Abington School District v. Schempp, 374 U.S. 203 (1963); Lemon v. Kurtzman, 403 U.S. 602 (1971); Stone v. Graham, 449 U.S. 39 (1980); Board of Education of Kiryas Joel Village School District v. Grumet, 512 U.S. 687 (1994), Santa Fe Independent School District v. Doe, 530 U.S. 290 (2000).
Response:
I do not have enough context to definitively answer whether public schools in New Hampshire can circulate religious imagery. However, I can provide some general information about the legality of religious imagery in U.S. public schools:
The First Amendment of the U.S. Constitution prohibits the government from establishing an oMcial religion. This is generally interpreted to mean that public schools cannot promote or endorse any particular religion. At the same time, the First Amendment protects private religious expression, including by students in public schools. So students are typically allowed to express their religious beliefs through things like clothing, jewelry, etc. Whether or not a public school can display religious imagery like crosses or nativity scenes is a complex issue that depends on the specific context and how the display is handled. Courts have ruled both for and against certain types of religious displays. In general, religious displays are more likely to be upheld if they have a clear secular purpose (like displaying art or celebrating a holiday), are initiated by students rather than school oMcials, and do not dominate the school setting. Displays that seem to endorse a particular religion are more vulnerable to legal challenge.
So in summary, while some religious expression is allowed in public schools, there are legal limits. Whether any particular religious image or display would be permis- sible in New Hampshire public schools would depend on the specific details of the situation. The school would need to be careful not to appear to be endorsing or promoting religion in a way that violates the First Amendment.
Definitive judgments require looking at the totality of the
circumstances around the display.
When asked to cite its sources, Claude, like GPT-4, identified and briefly summarized United States Supreme Court cases.25
CoCounsel's Legal Research
CoCounsel's built-in database of case opinions includes all 50 states, the United States Supreme Court, Federal Circuit Courts, District Courts, Bankruptcy Courts, the Tax Court, administrative materials, and legal articles published by JD Supra.26
When you click on the Legal Research Memo button, CoCoun- sel will ask you to “Explain your legal research request-using normal sentences like you would when talking to a colleague. Include as much relevant information for the search such as jurisdiction, background facts, and anything else that you feel is relevant to the query.” To ensure that it understands your query, CoCounsel will “interpret” your question and will in a few sec- onds give you its understanding of your legal research question. If CoCounsel has correctly understood your query, clicking the “Submit” button will cause CoCounsel to start processing your request. If there is an error or miscommunication between you and CoCounsel, you may return to your original question to add or subtract to it for clarity. CoCounsel will then re-interpret your request. There are filters by jurisdiction, date, and unpublished cases to narrow or expand the focus of your search.
Once your question has been submitted to CoCounsel, a prog- ress window and completion appear. This function allows you to run another legal research request while the first one is processing or tend to other work, and when the first legal research memo is
25 Lemon v. Kurtzman, 403 U.S. 602 (1971); Stone v. Graham, 449 U.S. 39 (1980); Lynch v. Donnelly, 465 U.S. 668 (1984); Allegheny County v. ACLU; 492 U.S. 573 (1989); Santa Fe Independent School District v. Doe, 530 U.S. 290 (2000).
26 Valerie McConnell, Which sources does CoCounsel search and review when generating a legal research memo?, Casetext, help.casetext.com/en/ articles/7207127-which-sources-does-cocounsel-search-and-review-when- generating-a-legal-research-memo.
finished (usually) in two to three minutes or less, it will send you an email telling you that your legal research memo is complete. The output from CoCounsel will be a Microsoft Word document with hyperlinks to the actual case, statute, regulation, or court rule that is relevant to your query. Those hyperlinks allow the neutral to conduct all necessary diligence using primary authority rather than AI-generated text.
User Prompt: “Can New Hampshire public schools cir- culate religious imagery?”
Response:
Answer
There is no clear answer to whether public schools in New Hampshire can distribute materials containing religious imagery. However, several authorities suggest that such distribution may be permissible if it does not advance or inhibit religion, does not foster excessive entanglement between the school and religion, and is not calculated to favor any particular religious sect.
Analysis
Several cases and statutes provide guidance on the question of whether public schools in New Hampshire can distribute materials containing religious imagery. The most relevant authorities are the New Hampshire Revised Statutes sections 189:26 and 194-B:7, and the cases Opinion of the Justices, 307 A.2d 558 (N.H. 1973) and State v. Mack, 173 N.H. 793 (N.H. 2020).
Section 189:26 of the New Hampshire Revised Stat- utes prohibits public schools from introducing any book that favors a particular religious sect. This statute sug- gests that the distribution of religious imagery would be permissible as long as it does not favor one particular sect over another. Section 194-B:7 sets out the criteria for determining whether a chartered public school is secular or nonsecular, and prohibits schools from advancing or prohibiting religion, or fostering excessive entanglement
between the school and religion. This statute also sug- gests that the distribution of religious imagery would be permissible as long as it does not violate these criteria.
The case Opinion of the Justices, 307 A.2d 558 (N.H. 1973) discusses the constitutionality of religious activities in public schools in New Hampshire. The court held that a law does not offend the First Amendment's prohibition on enactments respecting an establishment of religion if it has a secular legislative purpose, does not principally advance or inhibit religion, and does not foster an exces- sive governmental entanglement with religion. This case suggests that the distribution of religious imagery in pub- lic schools would be permissible if it meets these criteria. Finally, the case State v. Mack, 173 N.H. 793 (N.H.
2020) discusses the free exercise of religion under the New Hampshire Constitution. The court held that when religious practices violate a generally applicable law, the state constitution requires a balancing of competing interests. This case suggests that if the distribution of religious imagery in public schools is challenged, the court will need to balance the interests of the school against the interests of those challenging the distribution.
CoCounsel then summarized and linked directly to seven New Hampshire cases, two New Hampshire statues, and an article,27 and CoCounsel described how each of these sources are relevant to the research inquiry. In this instance, CoCounsel did not ask whether we wanted to include federal as well as state law in our research. When we added “under federal law,” to the beginning
27 In the order they appeared on CoCounsel: Opinion of the Justices, 108 N.H. 97 (N.H. 1967); Opinion of the Justices, 307 A.2d 558 (N.H. 1973); Opinion of the Justices, 233 A.2d 832 (N.H. 1967); State v. Mack, 173 N.H. 793 (N.H. 2020); Opinion of the Justices, 109 N.H. 578 (N.H. 1969); Lincoln
v. Page, 109 N.H. 30 (N.H. 1968); City of Concord v. New Testament Baptist Church, 118 N.H. 56 (N.H. 1978); N.H. Rev. Stat. §§ 189:26, 194-B:7; Fox Rothschild LLP, Clear as Mud: Navigating In-School Employee Expression in the Wake of Kennedy v. Bremerton School District, July 15, 2022, www
.foxrothschild.com/publications/clear-as-mud-navigating-in-school-em
ployee-expression-in-the-wake-of-kennedy-v-bremerton-school-district.
of the prompt, CoCounsel provided a different answer, citing 11 First Circuit and 3 District of New Hampshire cases.28 It also identified with a red flag graphic that one of the cited cases had been reversed.
iv. Analysis of the Platform's Legal Research Skill
Like ChatGPT and Claude themselves, we recognize the short- comings of conducting legal research using a general-purpose AI tool that does not include current source material in its training. Not only did those platforms deliver much more generalized answers to legal research questions, but they also acknowledged that their information is out of date. CoCounsel, which specializes in legal research specifically, surpasses the other platforms in terms of the breadth and currentness of its legal databases, as well as its ability to parse legal questions more like a lawyer would, acknowledging, for example, the differences in jurisdictions and the hierarchy of courts. CoCounsel also Shepardizes its citations and links directly to the source material, thereby guarding against inaccuracies and hallucinations. We caution readers, however, that, when we specified a jurisdiction in the prompt rather than using the filters, CoCounsel myopically focused on opinions from that jurisdiction's courts without acknowledging the relevance, for instance, of United States Supreme Court opinions.
28 In the order they appeared on CoCounsel: Freedom v. Hanover Sch. Dist., 626 F.3d 1 (1st Cir. 2010); Freedom from Religion Found. v. Hanover School, 665 F. Supp. 2d 58 (D.N.H. 2009); Weisman v. Lee, 908 F.2d 1090 (1st Cir. 1990); Local 8027 v. Frank Edelblut, Comm'r, 21-cv-1077-PB (D.N.H. Jan. 12, 2023); Parker v. Hurley, 514 F.3d 87 (1st Cir. 2008); Carson v. Makin, 979 F.3d 21 (1st Cir. 2020); Asociación v. García, 490 F.3d 1 (1st Cir. 2007); Knights of Columbus, # 94 v. Town of Lexington, 272 F.3d 25 (1st Cir. 2001); Rojas v. Fitch, 127 F.3d 184 (1st Cir. 1997); Governor Wentworth Reg. School v. Hendrickson, 421 F. Supp. 2d 410 (D.N.H. 2006); Gary S. v. Manchester Sch. Dist., 374 F.3d 15 (1st Cir. 2004); Shurtleff v. City of Bos- ton, 928 F.3d 166 (1st Cir. 2019); Norris ex rel. A.M. v. Cape Elizabeth Sch. Dist., 969 F.3d 12 (1st Cir. 2020); Members of Jamestown Sch. Comm. v. Schmidt, 699 F.2d 1 (1st Cir. 1983).
3.Document Generation
Document Generation Applications
In addition to analyzing complex texts for the purpose of summarization, analysis, or research, today's generative AI plat- forms can assist professionals, including arbitrators and medi- ators, in creating and editing documents. ChatGPT and Claude are two notable examples of LLMs designed to aid in drafting legal content.29 They can generate human-like text in response to user queries, sometimes incorporating information provided by the user. Moreover, these tools can refine their outputs based on user feedback.
ChatGPT and Claude can generate text in response to a prompt or a question by a user, sometimes based on documents, images, or other information provided by the user. ChatGPT and Claude can also iterate on the content they have generated based on feedback from the user.
Using LLMs such as ChatGPT and Claude for drafting legal
documents presents several advantages, including:
Time and Resource EMciency: LLMs can swiftly gen- erate text that resembles human-written content. This can substantially reduce the time spent on research, writing, editing, and proofreading of legal documents. Enhanced Creativity and Innovation: LLMs can gener- ate diverse, relevant, and novel text, which can foster fresh ideas, perspectives, and arguments that might otherwise remain unexplored. Improved Quality and Consistency: LLMs produce coherent, fluent, and grammatically correct text, mitigating errors, inconsistencies, and ambiguities in legal documents.
29 As of the date of this article, CoCounsel did not offer a drafting
function.
However, employing LLMs for legal drafting also comes with
specific challenges, which include:
Ensuring Accuracy and Reliability: LLMs, while power- ful, are not infallible. They can generate inaccuracies, outdated information, invent facts or cases, or misin- terpret user input. Therefore, users must cross-verify content generated by LLMs with credible sources and seek expert input. Maintaining Ethical Standards and Professional Responsibility: LLMs are not bound by ethical codes or professional regulations, and they lack moral or legal obligations. Users must exercise their judgment and discretion when using LLMs, ensuring compliance with data protection, confidentiality, intellectual property rights, plagiarism, and disclosure laws and regulations. Managing Expectations and Limitations: LLMs are tools that augment human capabilities, not substitutes for human lawyers. Users should not solely rely on LLMs for drafting legal documents and should employ their own knowledge, skills, experience, and intuition. Users must also be aware of LLM limitations, including data sources, training methods, time constraints, and domain specificity.
As applied to arbitration and mediation, the potential use cases for generative AI as a drafting tool are practically unlim- ited. They include:
Drafting arbitration agreements and mediation clauses in contracts. AI tools can help generate customized lan- guage based on industry standards and best practices. Preparing mediation statements and arbitration plead- ings. AI tools can analyze case facts and legal issues, and produce draft documents outlining each party's positions and arguments serving as a starting point for lawyers to refine. Drafting procedural orders and case management plans. LLMs can generate standard formats for
procedural orders based on prompts about the number of parties, issues in dispute, deadlines, etc.
Drafting settlement agreements and arbitration awards. LLMs can facilitate the creation of initial drafts of straightforward settlements or uncontested arbitration awards based on terms provided. Automating parts of mediation briefs and letters. LLMs can generate standard sections in mediation briefs, such as recitals and reservations of rights. Producing first drafts of form documents. When stan- dard contracts, agreements, or correspondence are needed, LLMs can produce a draft using templates and key case details.
Platform Comparison: Document Generation
Here are examples of drafting exercises we explored using ChatGPT-4 and Claude:
User Prompt: You are an attorney asked to draft an arbitration clause to be included in a contract between a Massachusetts software company and a New York investment bank. The parties wish the clause to invoke the Commercial Rules of the American Arbitration Association, and to cover the issues of the venue for the arbitration, a pre-arbitration mediation requirement, a reasonable time limit for completion of the arbitration, and reasonable discovery limitations. Please draft the arbitration clause.
ChatGPT 4 Response:
Here is a possible arbitration clause that meets the requirements of the parties:
Arbitration Clause
Any dispute, controversy or claim arising out of or relating to this contract, or the breach, termination or invalidity thereof, shall be settled by arbitration in accordance with the Commercial Arbitration Rules of
the American Arbitration Association (the “AAA Rules”) in effect on the date of this contract, except as modified herein.
The place of arbitration shall be Boston, Massachu- setts, unless the parties agree otherwise in writing.
Before initiating arbitration, the parties shall attempt to resolve their dispute amicably through mediation administered by the American Arbitration Association under its Commercial Mediation Procedures. Either party may commence mediation by providing to the other party a written request for mediation, setting forth the subject of the dispute and the relief requested. The parties will cooperate with the American Arbitration Association and with one another in selecting a mediator from the Amer- ican Arbitration Association's panel of neutrals and in scheduling the mediation proceedings. The parties agree that they will participate in the mediation in good faith and that they will share equally in its costs.
If the dispute is not resolved within 60 days after the request for mediation, either party may initiate arbitra- tion by filing a notice of arbitration with the American Arbitration Association and serving a copy on the other party. The notice of arbitration shall include a statement of the claim, the relief sought, and the names and contact information of the parties.
The arbitration shall be conducted by one arbitrator, unless the amount in dispute exceeds $1,000,000, in which case three arbitrators shall be appointed. If one arbitrator is to be appointed, the parties shall endeavor to agree on the arbitrator within 15 days after receipt of the notice of arbitration. If three arbitrators are to be appointed, each party shall appoint one arbitrator within 15 days after receipt of the notice of arbitration, and the two party-appointed arbitrators shall appoint the third arbitrator within 15 days thereafter. If a party fails to appoint an arbitrator within the specified time limit, or if the two party-appointed arbitrators fail to appoint the third arbitrator within the specified time limit, the Amer- ican Arbitration Association shall appoint the arbitrator.
The arbitrator(s) shall decide the dispute in accor- dance with the substantive law of New York, without regard to its conflict of laws rules.
The arbitration shall be completed within six months after the appointment of the arbitrator(s), unless other- wise agreed by the parties or extended by the arbitrator(s) for good cause.
The discovery in the arbitration shall be limited to document production and depositions of no more than three witnesses per party, unless otherwise agreed by the parties or ordered by the arbitrator(s) for good cause.
The award of the arbitrator(s) shall be final and bind- ing on the parties, and may be entered and enforced in any court having jurisdiction. The award shall state the reasons for the decision and apportion the costs of arbi- tration between the parties. The prevailing party in any arbitration or judicial proceeding arising out of or relating to this contract shall be entitled to recover its reasonable attorneys' fees and expenses from the other party.
Claude Response:
Here is a draft arbitration clause for the Massachu- setts software company and New York investment bank: Arbitration. Any dispute arising out of or relating to this Agreement shall be submitted to binding arbitration in accordance with the Commercial Arbitration Rules of the American Arbitration Association. The parties agree to first attempt mediation prior to initiating arbitration. The mediation shall take place in Boston, Massachusetts and shall conclude within 30 days of initiation unless extended by mutual written agreement. If mediation does not resolve the dispute, the arbitration shall take place in New York, New York. Discovery shall be limited to the request and production of documents directly relevant to the dispute, along with no more than two depositions per party, each lasting no more than four hours. The arbitrator shall issue a final award within 120 days of the initiation of arbitration unless this time is extended by mutual written agreement. The arbitrator shall have
authority to award compensatory damages only and shall not award punitive or exemplary damages. The arbitrator shall have no authority to award provisional remedies of any nature including, but not limited to, injunctive relief. Any award rendered shall be final and binding upon the parties. Judgment upon the award may be entered in any court having jurisdiction.
A second example:
User Prompt: You are a mediator who has been asked to provide the parties with an opinion on the merits of a dispute between a copywriter and her employer over ownership of an article written by the copywriter. The article was written at 10:00 p.m., after work hours, on her work computer. The article has some relevance to the business of the employer, but would never be published by the employer. Please advise the parties of your view on ownership of the article.
ChatGPT-4 Response:
Thank you for choosing me as your mediator. I have reviewed the facts of your dispute and searched the web for relevant information on the topic of work made for hire. Based on my research, I have formed an opinion on the merits of your case, which I will share with you below.
The main issue in your dispute is whether the article written by the copywriter is a work made for hire or not. If it is, then the employer owns the copyright in the article. If it is not, then the copywriter owns it.
According to the definition of work made for hire in the U.S. Copyright Act, a work is a work made for hire in two situations:
When it is created by an employee as part of the
employee's regular duties
When a certain type of work is created as a result of an express written agreement between the creator and a party specially ordering or commissioning it
In your case, the second situation does not apply, as there is no evidence of any written agreement between the copywriter and the employer regarding the article. Therefore, the only question is whether the article was created by the employee as part of her regular duties.
This question is not easy to answer, as it depends on various factors, such as the nature of the employment relationship, the scope of the employee's duties, the level of control and direction exercised by the employer over the employee's work, and the intent of the parties. [Links omitted]
Based on the information you provided, I think that there are arguments for both sides. On one hand, the employer could argue that the article was created by the employee within the scope of her employment, because:
The article was written on her work computer, which implies that she used the employer's resources and equipment to create it The article has some relevance to the business of the employer, which implies that she used her skills and knowledge acquired from her employ- ment to create it The article was written during a time period that could be considered part of her regular work hours, depending on her schedule and expectations
On the other hand, the copywriter could argue that the article was not created by her within the scope of her employment, because:
The article was written after her normal work hours, which implies that she did not create it as part of her assigned tasks or projects The article would never be published by the employer, which implies that she did not create it for the benefit or use of the employer The article was written for her own personal interest and expression, which implies that she did not create it under the control or direction of the employer
In my opinion, based on these arguments and facts, I think that the copywriter has a stronger case than the employer. I think that the article is not a work made for hire, because it does not meet the criteria of being created by an employee as part of her regular duties. Therefore, I think that the copywriter owns the copyright in the article.
However, this is only my opinion and not a legal advice. You should consult with a qualified attorney if you want to pursue this matter further. I hope that this opinion helps you resolve your dispute amicably and fairly. Thank you for your cooperation and trust.
Claude Response:
Dear Parties,
As the mediator in this dispute over ownership of the article written by the copywriter, here is my opinion on the merits based on the facts provided:
While the article was written after work hours, it was done so on the employer's computer with resources provided by the employer. However, the article has little relevance to the employer's business and would likely never be published by the employer.
The general rule is that works created by an employee within the scope of employment belong to the employer. But here, since the article was written after hours on a matter unrelated to the employer's business, there is a stronger argument that the copywriter herself owns the copyright to the article, not the employer.
My recommendation would be to consider a compro- mise—the copywriter maintains copyright ownership of the article, but grants the employer a limited license to use portions of the article for internal purposes only. This allows both parties to benefit from the work product to some degree.
I hope this assessment of the merits provides a rea- sonable middle ground to resolve this dispute. Please let
me know if you have any other questions as we continue the mediation process.
In both examples, ChatGPT-4 excels in providing nuanced and in-depth responses, offering users a comprehensive under- standing of the legal issues with links to sources. Claude, on the other hand, provides more concise responses, focusing on practical solutions. The choice between the two depends on the user's specific needs and preferences.
Future Uses
As new generative AI tools are developed and the existing ones learn and grow, we foresee additional uses in arbitration and mediation. While possibilities abound, we identify a few here.
1.Judicial and Arbitral Decision Making
The hot topic in the arbitration world is whether genera- tive AI will replace arbitrators as decision-makers. For years, companies like eBay have used learning computer systems to resolve disputes online.30 But, with the proliferation of LLMs, computers can digest large quantities of information rapidly and either suggest or predict the outcome of a dispute. Although we are not aware of a court or ADR institution that designates AI as the final decision maker,31 the technology is evolving rapidly.
30 Colin Rule, Resolving Disputes in the World's Largest Marketplace, ACResolution (Fall 2008), https://colinrule.com/writing/acr2008.pdf.
31 Automated Justice: Use of AI in the Resolution of Disputes, Tilleke & Gibbons, July 11, 2023, https://www.tilleke.com/insights/automated-jus tice-use-of-ai-in-the-resolution-of-disputes/ (“For now, there are few, if any, courts that use automated decision making. However, the use of AI to assist judges is not uncommon. In the US, many courts use the COMPAS system to help determine criminal sentences, a practice that has been upheld by the Wisconsin Supreme Court. Malaysian courts have experimented with sim- ilar AI systems. More recently, judges in Colombia and Pakistan have used ChatGPT to assist with preparing judgments. By far the most advanced case
For example, certain AI tools predict what a particular judge would decide or propose a decision for the arbitrator. Pre/Dicta, which launched in 2022, is a predictive litigation analytics tool “that looks for patterns in a judge's past decisions and potential areas of bias, using around 120 data points.”32 According to Pre/ Dicta's CEO, the tool can predict how a judge will rule with 86% accuracy.33 Bridget McCormick, president and CEO of the Amer- ican Arbitration Association, has commented that information tools like Pre/Dicta's are most useful to parties “trying to decide whether to litigate at all, as opposed to arbitrate or mediate.”34
Other AI tools are tailored for arbitration but are currently targeted at advocates rather than neutrals. In June 2023, Jus Mundi launched a beta version of “Jus-AI,” its AI-powered arbi- tration intelligence tool. The tool “is aimed at providing concise summaries from arbitral awards and court judgments,”35 and “interpret[s] specific provisions of arbitration rules by verifying the legal reasoning of thousands of arbitrators, counsel, and experts.”36 “The International Centre for Dispute Resolution, the international division of the American Arbitration Association, and Jus Mundi united to make non-confidential ICDR interna- tional arbitration awards freely available to the global legal and business community,”37 which could invite their use as training materials for Jus-AI and other arbitration-focused AI tools.
study is China, where AI is used in a number of ‘smart courts' to automate transcription, analyze evidence, recommend decisions, and monitor the consistency of judgments with past case law.”).
32 Ryan Heath, AI Tells Lawyers How Judges Are Likely to Rule, Axios,
Sept. 12, 2023.
33 Id.
34 Id.
35 Lucia Bizikova et al., IA Meets AI—Rise of the Machines, Daily Jus, Oct. 2, 2023, https://dailyjus.com/legal-tech/2023/10/ia-meets-ai-rise-of- the-machines.
36 Legal Intelligence Powered by AI Technology, Jus Mundi, jusmundi
.com/en.
37 A Key Partnership for the Publication of AAA-ICDR Arbitral Awards, Jus Mundi, https://jusmundi.com/en/partnership/icdr?utm_source=Website &utm_medium=why%20partner%20with%20us%20page&utm_campaign=
The available data is more limited for commercial arbitra- tions, which are generally confidential. Arbitration practitioners, nevertheless, still expect the use of AI tools to rise in the coming years, depending on more “initiatives to publish a greater number of commercial arbitration awards and related materials (with or without redaction/anonymization).”38
While predictive analysis tools are generally directed at advo- cates, they may “enable arbitrators to predict the possible factual outcomes of their awards, based on historical cases and juris- prudential data. Simply put, ChatGPT can predict the chances of an award being annulled, denied enforcement or voluntarily complied with, which, in turn can allow arbitrators to draft their awards in such a way, so as to ensure their longevity.”39
The same tools that currently review large quantities of past judicial and arbitral decisions to evaluate or predict decisions could eventually become the deciders. While that reality may seem far off, we anticipate that automated decision makers will be available shortly to provide first-cut, appealable resolutions in disputes with lower amounts in controversy.
2.Augmenting Information Provided to the Neutral
ChatGPT-4 suggests that, as generative AI technology evolves, it can assist arbitrators and mediators by providing the following types of information that might not be otherwise available:
Emotion Analysis: By analyzing voice tones, word choices, or facial expressions during hearings or ses- sions, AI could provide insights into parties' emotional
aaa_icdr_why_partner_with_us&utm_id=AAA-ICDR%20Landing%20page
%20&utm_content=aaa_icdr_why_partner_with_us.
38 Bizikova, supra note 35.
39 Leonardo F. Souza-McMurtrie, Arbitration Tech Toolbox: Will ChatGPT Change International Arbitration as We Know It?, Kluwer Arbitration Blog, February 26, 2023, arbitrationblog.kluwerarbitration
.com/2023/02/26/arbitration-tech-toolbox-will-chatgpt-change-interna tional-arbitration-as-we-know-it/.
states, aiding neutrals in navigating discussions more
effectively.
Automated Fact-Checking: In real time, AI could verify claims or facts presented during proceedings against a vast database of information, ensuring the accuracy of presented data. Predictive Behavior Analysis: Based on historical data and behavioral patterns, AI could predict the likely responses of parties to certain resolutions or decisions, allowing arbitrators and mediators to strategize their approaches.40
Conclusion
Machines are not ready to assume sole responsibility for legal analysis, decision-making, or dispute resolution functions. How- ever, they can adroitly support arbitrators and mediators who are well informed about the technology's strengths and potential pit- falls. As we look to the future, we see the potential for even more transformative and disruptive applications. These advancements, while promising, should be approached with judicious balance, ensuring that the essence of human judgment, empathy, and ethics remain paramount in arbitration and mediation. As we stride into this AI-augmented future, the arbitrator or mediator is not diminished but rather enhanced, equipped with a richer toolkit to better serve the evolving needs of parties in dispute.
40 ChatGPT-4's excerpted response to the October 30, 2023, prompt: “As generative AI evolves, what are some of the future, cutting-edge ways arbitrators and mediators could use it?”
