Figure 1: Image generated by Dall-E 3
This article was prepared by the ASTC AI Task Force
This is the second part of the two-part series.
Uses and Limitations of AI Chatbots in Trial Consulting
In Part I of this article we addressed some fundamentals concerning AI and AI chatbots. Now, we turn to the use of AI chatbots in trial consulting.
If you ask AI chatbots about how they might be used in trial consulting (and we did!), the scope of applications is anything that trial consultants currently do, with some exceptions made for in-court consulting such as in-court jury selection.
However, it is necessary to keep in mind the knowledge base upon which the LLMs were trained. As we discussed in Part I, transparency concerning the training of the major LLMs used today is poor and it seems unlikely that they have been trained on a wide universe of scholarly and academic research–partly because huge amounts of such research have never been digitized, partly because much that is digitized is behind paywalls. Some AI chatbots are unable to reliably process PDFs, charts/tables, images/photographs, and other nontext/narrative content that is publicly available, leading to potentially incomplete, misleading, or inaccurate results. The value and comprehensiveness of the content available from generally oriented, publicly trained programs like ChatGPT, Gemini, and Copilot is therefore unclear. In the current competitive environment, companies are constantly tweaking and updating their programs so the models are changing rapidly and becoming increasingly sophisticated. Nevertheless, the old principle of garbage in, garbage out applies–which is not to say that these tools produce garbage, but that if you want to make the best use of these tools, their greatest power will likely come from applying them to your own curated sets of texts/information that reflect your specific needs and interests.
To do this, you need to have a subscription (e.g., ChatGPT Plus) or access to an enterprise version (e.g., ChatGPT Enterprise or Microsoft’s Azure AI) that allows you to upload your own data into the chatbot’s engine, giving it specialized training on the universe of texts that you care about. Doing this entails security and confidentiality risks, however, so don’t create a custom chatbot casually. You need to understand exactly how any texts/information that you “add” to the chatbot will be used by its company (e.g., model training).
Because AI chatbots have been programmed to provide responses within specific guidelines or “guardrails,” as we mentioned in Part I, they may refuse to answer certain queries.
As we discussed in Part I of this article when considering AI “hallucinations,” currently, there are no strict truth or accuracy guardrails in LLM programming (and that may never be possible). In fact, recent research has demonstrated that current claims of hallucination-free legal AI legal research tools have proven overoptimistic, with 17% to 33% incidences of hallucinations in the tools tested.[1] Therefore, you must always independently verify AI chatbot results–which may limit the usefulness of an AI chatbot in situations where time is a constraint.
Finally, although AI chatbots are trained to understand natural language, learning how to “talk” with a chatbot to obtain the results you want takes time and there is skill involved in crafting queries that get quality answers. This takes us to the topic of “prompt engineering.”
Like many of the topics considered here, prompt engineering could be an article in itself. In basic terms, prompts are the core tool users have to guide the activity and output produced by the AI chatbot. It provides the task instructions, task goals, and the information domain to be considered, along with the desired output/product (e.g., summary, analysis, stories, memos, or list of voir dire questions). Creating good prompts enables AI Chatbots to provide the most useful results. As such, the following are some useful guidelines.
Role adoption. Ask your AI chatbot to adopt a specific role or orientation (e.g., defense or plaintiff’s lawyer or trial consultant) when addressing its task. This will help your AI chatbot provide more refined/relevant answers.
Be clear. Keep prompts clear and focused. General instructions or ambiguous language tend to produce less useful answers and often require more refined follow-up prompts to get the information you seek.
Provide context. To ensure the answers you receive are most useful, provide information in the prompt (or attachments of relevant documents) that gives the AI chatbot the appropriate context for your query.
Keep tasks simple. Break down complex tasks into a series of steps or subtasks. This will help AI chatbots focus on each step needed to provide quality results.
Provide examples. Providing examples of what you are interested in will lead to more useful results. If you want the AI chatbot to incorporate rhetorical questions in an argument, provide an example in the instructions to illustrate how this could be done.
Specify the output desired. Direct the AI chatbot to produce the results in the desired format. If you want charts, a bullet point outline, a generated image, a word-count limited narrative, a summary report with designated subparts, or a spreadsheet containing the desired information, explicitly state what form your product should take.
Build on past prompts. The prompts within a given “conversation” tend to be cumulative in nature. You can refine prompts to get better results either by asking for the task to be repeated (when the answer was less than satisfactory) or by systematically modifying the prompts used to obtain more useful results. In addition, you can structure your series of prompts to take advantage of previous prompts within the same conversation. For example, you can ask for a summary of certain files/documents/information, including strengths and weaknesses of the parties’ cases and, if the summary meets your needs, you can then prompt it to provide an opening statement based on this analysis. Such an approach will provide a better quality response and greater transparency as to the basis of the later response.
Be willing to start a new conversation. Should you start to encounter unreliable information (e.g., hallucinations), it is helpful to stop the current conversation and start a new conversation.
Validate or verify the results. Given the propensity of AI chatbots to occasionally provide inaccurate information or hallucinate, it is necessary to be vigilant in verifying the results produced and not simply accept the results as accurate.
Even with the current caveats and constraints we have discussed, AI chatbots are evolving so rapidly that we believe with tweaks and changes, they will soon be indispensable aids in many core consulting tasks, and so we turn to discussing how they might be used in (a) jury selection; (b) communication and persuasion; (c) witness preparation/testimony; and (d) jury research.
Jury Selection
There are four areas where we see AI chatbots being potentially useful in jury selection:
- internet research on potential jurors
- developing juror profiles
- developing voir dire questions
- developing supplemental juror questionnaires
Because of guardrails that have been programmed in, AI Chatbots often exclude themselves from providing in-court assistance in jury selection. There are still ways to use them to aid in background research, but the results currently provided by AI chatbots tend to be more general and lack the more specific detail needed in these areas.
To make an AI chatbot maximally useful, it needs to “understand” the information most relevant to your case–it needs, in other words, to “train” on your case. This means you need to provide it with some case information–either actual case materials; disguised information that still conveys significant specifics; or hypothetical or summary information that identifies important details. This raises obvious ethical, security and confidentiality issues that we discuss later in this article.
Internet research on potential jurors. One obvious application for AI chatbots is their use in conducting internet research on potential jurors. Their ability to search multiple social media platforms quickly makes it seem like a no-brainer. True, they can conduct searches in rapid order. However, as one of the authors has found, some AI chatbots have refused to do such searches or discontinued searches and only resumed juror searches with some creative cajoling. In addition, numerous instances of hallucinations were found in such searches and their performance against trial consultants’ efforts was less than stellar. It therefore appears that at the moment, AI chatbots have limited utility for this task.
Juror profile development. A second area of interest is juror profile development. AI chatbots will provide profiles for favorable and unfavorable jurors based on the information provided to them. As noted above, these profiles tend to be general and may not be based on research and theory relevant to a given case. As a result, it is important to ask for the rationales for these profiles and evaluate them against your research and knowledge in the field and the instant jurisdiction. However, they can be helpful in the same way that looking at recipes for a dish that you already know how to cook can be, helping bring to mind topics or questions that you had not included but could be helpful.
The long-term potential here is in using AI technology to create a custom program trained on proprietary data; we understand that there are jury consulting companies already doing this.
Supplemental juror questionnaires. AI chatbots can produce versions of supplemental juror questionnaires (SJQs) and in a format that can be adapted to fit the trial/jurisdiction. These SJQs tend to be basic and general and in a simple format and need significant adjustments to meet case-specific demands and the content/format constraints of the court/trial jurisdiction, but they can be helpful starting places or points of comparison. In addition, there is the potential for AI with handwriting-to-text conversion capabilities to assist in processing completed hard-copy (and online) SJQs.
Voir dire development. AI chatbots can produce voir dire questions. These questions tend to be general and global in nature. These questions should not be taken at face value. Professional intervention is needed to tie questions to the case more closely, ensure that critical opinions, values, experiences, and backgrounds are addressed, and capitalize on effective question phrasing.
Communication and Persuasion
Key to success in jury trials is the ability to communicate clearly and persuasively with jurors. AI chatbots can process information provided to them and produce content and recommendations addressing themes and case theories, opening statements, and closing arguments.
Themes and case theories. AI chatbots can provide themes and case theories based on the information they receive. These themes can be useful, even if general, such as “defendant putting profits over safety” or “taking responsibility for one’s actions,” provided the essence of the case can be appropriately captured by such general statements. Case theories can also be generated, again, noting the potential for more generalized recommendations.
Opening statements. AI chatbots can also generate opening statements, again of a more general nature. The degree of detail you provide to the AI chatbots about your case will influence the subsequent detail and usefulness of any openings provided. However, such general structures for opening may not take full advantage of rhetorical tools, such as rhetorical questions or inoculation, that professionals can apply to the same set of facts.
Closing arguments. Closing arguments offer a unique opportunity in that the facts, arguments, and jury instructions are known and can be summarized along with trial transcripts for use by AI chatbots. It may be possible to provide AI chatbots with real-time transcripts from critical parts of the trial, (e.g., opening statements, certain documentation or descriptions of documents, testimony transcripts from critical witnesses, and jury instructions) for analysis and eventual closing arguments guidance. As with opening statements, AI chatbots generally lack a sophisticated understanding of elements of persuasion and may provide more general guidance as compared to legal professionals.
Witness Preparation/Testimony
Witness preparation proves difficult for AI chatbots. While AI chatbots may tout the use of facial recognition software (which they do not use), the use of FER (Facial Emotion Recognition) software presents a mixed picture, with at least one study showing that machine learning facial recognition software did not fare well with recognizing emotions (66% correct classification between FER and human raters), while another study showing a relatively higher level of accuracy (76.6% correct classification between FER and human raters) in a clinical setting using video from therapy sessions. However, AI chatbots can examine transcripts or witness statements and propose questions to be asked of witnesses in preparation for deposition or during depositions and for use at trial during direct and cross-examination. The likelihood of these questions surpassing professional capabilities remains to be seen but the results could serve as food for thought in the preparation process.
Jury Research
There are a number of other uses of AI chatbots in the field of trial consulting, but the final area of significance that we will cover in this article is jury research, in particular, small group research (e.g., focus groups and mock trials) and opinion survey research. Generative AI presents a transformative opportunity in this realm, with the potential to support consultants at various points in the jury research process, from recruiting research participants and the preparation of study materials to data analysis and reporting. This could increase efficiency and enable consultants to focus more on the unautomatable aspects of consulting, such as building client rapport, managing expectations, and providing recommendations based on the research findings, among other functions.
Focus groups and mock trial research. AI chatbots may assist with the development of materials, such as questionnaires and recruit specifications, in preparation for focus groups and mock trials. In addition, moment-to-moment analysis of jurors’ persuasion ratings can be helpful in identifying the least and most persuasive arguments in real time.
Furthermore, AI chatbots can generate summaries of participants’ comments and hours of group deliberations or discussions which can be helpful in condensing and analyzing vast amounts of qualitative data. There are even AI features embedded in video conferencing platforms, such as Zoom, or software vendors providing video-to-text transcriptions which can provide transcripts and summaries of group discussions based on online discussions (or summaries based on verbatim transcripts to avoid quality control issues); though they may not be 100% accurate, they provide a good starting point to digest the large amount of content generated from group discussions. These summaries could provide an overall picture of case themes as well as the types of jurors most persuaded by different arguments. This can be further developed by AI to support the creation of a juror profile (keeping in mind sample size issues) which would provide a foundational level of understanding of the jurors to potentially strike in jury selection.
AI chatbots can also be used for damages simulations using individual damages’ awards and simulating a group verdict, which can be repeated numerous times to achieve hundreds of group verdict simulations. This might inform trial strategy and even settlement discussions.
It is important to note that while AI chatbots may be useful in qualitative content analysis of jurors’ responses, it is not yet reliably used in quantitative data analysis; though, we do expect that to change as technologies continue to advance. As with its application in other aspects of trial consulting, generative AI should be used with caution until clearer regulations are established with respect to privacy and confidentiality of information.
Opinion surveys. Relatedly, AI chatbots can assist with the creation of surveys and the evaluation of survey results. Provided with a survey topic, target audience, and research goals, current chatbots are capable of generating questions suitable for survey use. Additional details regarding the aims, goals, or constraints of the survey can further refine the questions provided by the chatbot. AI can also provide suggestions on the types of responses that should be offered for a specific question, such as whether the question should be open-ended, closed-ended, or Likert-type. As with any other AI-generated content, the survey questions should be evaluated for clarity, relevance, and bias prior to finalizing the survey. The use of generative AI can potentially significantly reduce the time spent creating survey content for new topic areas.
AI chatbots can also be beneficial in analyzing data collected via surveys. As noted in the sections above, AI chatbots can summarize vast amounts of data, including survey responses. Responses to open-ended questions and other qualitative survey responses can quickly be summarized and broken down by theme to allow for a quick overview of survey respondents’ thoughts and beliefs. Companies are already building generative AI models to analyze qualitative survey data further. These models will compile, summarize, and provide feedback that can be then used to inform witness preparation strategies, develop focus group and mock trial materials, and assist with theme development and case evaluation.
Ethical and Security Issues
The ethical and security issues involved in using AI differ depending on whether we are talking about programs like Copilot and Gemini that are connected to the Internet and are theoretically accessible by everyone; programs like Claude, that are publicly available but are not connected to the Internet and are searching for responses only within their own dataset, perhaps supplemented by data that you have provided; or privately developed programs that may be built on top of engines like ChatGPT but exist behind firewalls and are accessible only to select users. This article focuses on publicly available chatbots in free and subscription versions because that is what is available right now to all of our members. However, we note that because of the security and confidentiality issues their use raises, we believe privately created, custom AI chatbots and AI programs are likely the future of AI programs in trial consulting.
Ethical and security concerns of using AI chatbots can be divided between three different categories:
- Obligations regarding use and disclosure of use
- Monitoring of results
- Ensuring confidentiality of your own and client confidential data
Use and Disclosure Obligations
At one time, conducting a mock trial was an unusual and exotic undertaking. Today, mock trials are routine and in many instances, failing to conduct a mock trial could be considered malpractice. We are at the same point with AI tools right now that lawyers were with mock trials back in the 1970s and 1980s. It is not currently routine nor obligatory to use them. However, that could quickly change as tools develop, so it’s important to keep abreast of developments. At this point, where we are all running to catch up to the technology and are experimenting to understand how to use it ethically and productively, it’s most likely clients do not assume you are using AI technology—unless you tell them. Given that AI chatbots are still an experimental technology and present security risks, we think it is advisable to discuss any potential uses of AI technology with your client, in particular, if the use of AI will entail sharing any non-public information about the case with a chatbot.
Monitoring Results
By their nature and design, it is known that AI chatbots both lie and plagiarize. It is therefore incumbent on every consultant to check factual information provided by a chatbot to ensure it is accurate, and to evaluate images and texts produced in response to creative prompts to determine if they might qualify as copying rather than creation.
Ensuring Confidentiality
Internet AI chatbots like Copilot and ChatGPT may incorporate your queries and any information you input into the chat into their LLMs as part of their training to improve their responses. This means that any data you input is not confidential–it is now part of a publicly available universe of data. All of these companies say in their fine print that they save your conversations – the information you type into a chat with a chatbot – for a period of months, and may incorporate some of that data into their models. In addition to potentially releasing private information publicly, your queries could be used to train the chatbot, making it give better answers to your opponents if they choose to use the same tool after you.
Given that documents filed with the court such as complaints and motions for summary judgment are typically public documents, you should be safe using information from these to form queries, but we believe the ethically responsible course of action is to inform your client of your intention and to work in conjunction with your client to create an agreed-upon universe of information about the case that might be used for public searches and queries.
Some chatbots provide extra levels of data security and prohibit their use for training purposes for an extra fee or as part of a premium service.
You therefore must carefully read the fine print when you sign up to use a service to understand how data you input might be used or protected from use.
As a general practice, we recommend not entering any attorney-client privileged or court-restricted information into a publicly maintained chatbot unless you know it has been configured for data security by your organization.
Conclusions
It’s been little more than a year since generative AI chatbots stormed onto the scene. These tools present opportunities but also have limits, and raise important ethical issues. Currently, we see AI chatbots as a complementary tool, not a replacement for trial consultants. Trial consultants need to be transparent in their use of AI. We need to accurately characterize the work we do, which includes not making false claims about the use of AI, being honest about what AI can and cannot do, and indicating what steps, if any, we are taking to ensure the confidentiality of client-sensitive information.
The long-term opportunities–and the long-term dangers–lie in developing customized, proprietary tools that are trained on the data that trial consultants and trial lawyers care about, offering more specialized guidance and data analysis. It will take time and money to develop such tools, and it’s likely to be difficult for smaller operators, as many ASTC members are, to develop these tools independently. Such tools offer the potential to improve the quality of insights and guidance trial consultants can provide—and the potential to deepen social inequities and reinforce stereotypes and biases that exist in our justice system, as well as to create inequities within the trial consulting profession itself. We believe that the ASTC has an important role to play in ensuring that AI tools for trial consulting are created in accordance with our professional and ethical standards and allow for healthy competition in the profession.
This article was prepared by (in alphabetical order) Erica Baer, Jeffrey Frederick, Anupama Gulati, Kristi Harrington, and Sarah Murray.
[1] See https://dho.stanford.edu/wp-content/uploads/Legal_RAG_Hallucinations.pdf.