IIA-A. AI THEORY: HOW AI UNDERSTANDS ITSELF – MACHINE LEARNING

When humans are in their preliminary stages of infancy and childhood, it may be observed that those infants and children learn constantly about their surroundings through physical and language-based investigation and about their own internal selves through introspection. This constant refrain – learning – seeps into any number of tasks that those children or infants learn to perform because the skills of understanding and comprehension that those children or infants are developing are applicable to as many tasks as exist in their comprehension. In other words, because the learned skills are applicable to as many tasks as the infant/child understands, that infant/child can learn the connections between their instructions and the task they are performing such to make the learning long-term.

To illustrate this, imagine an infant who has been crawling around the house for months whose parents will soon want to teach the infant how to take their first steps. An established method for teaching oneself how to walk has existed as long as humans have possessed knowledge of that method, but infants generally do not understand or comprehend human speech in a technical way and as such their parents cannot speak to them and tell them how to walk. Walking cannot effectively be taught to that infant until they begin to understand it through repetition in a safe environment, which their parents can provide. Once that toddler first toddles their way out of crawling, that toddler’s parents rejoice that their toddler has learned how to perform this key mobility task in a basic and repeatable way. If the training of that infant or child on how to walk successfully was comprehended, then not only did that child or infant learn how to successfully complete a task that it did not understand how to complete prior to that point, but they also, in an almost instinctual sense, learned how to learn.

In other words, applying the learning to the new task ipso facto instructed the infant/child on how such a task was applicable in other areas of learning. In an analogous way to the younger humans described above, LRATs (such as Lexis+ AI/Protege) are and have been learning tasks like legal research and drafting for the last few years. However, for legal researchers to effectively use LRATs and perform AIALR in an efficient manner, it would behove them to understand how AI learns how to learn through the process of machine learning, or ML.

ML is the process whereby algorithms teach Large Language Models, or LLMs, how to identify specific patterns in data and how to perform specific tasks to improve their performance through data-driven learning in either a supervised or unsupervised manner.Footnote ⁶ In other words, ML-influenced algorithms teach LLMs how to identify patterns in data, or how to learn how to learn, much like the infant/child above learning how to move their limbs to enable their own locomotion. This pattern identification in AI functions similarly to humanity’s pattern identification ability. In the context of legal research, ML teaches LRATs how to learn the relational information contained in a database/dataset or other informational resource, but within the context of AI, ML is a subset of AI focused on the use of various self-learning algorithms that derive knowledge from data to predict outcomes.Footnote ⁷ This distinction between AI terms as they are generally understood and AI terms in the context of legal research harks back to the common lack of a definition for AI discussed above.

It is not necessary for legal researchers to understand the minutiae of this distinction, but the meaning of it is that ML effectively allows LRATs to learn a task and not forget how to do it, like the toddler learning to walk analogy. In addition, just as ML can mean different things in different learning contexts as seen above, it is important for a researcher to distinguish between the import of an AI theory or AI-assisted tool on legal research and the import of those theories/tools on AI generally. As an example, although ChatGPT as an AI tool is often excellent at drafting human-sounding language and is an excellent example of a tool with a large import on AI generally, its usage in legal research may be less effective given the black box nature of AI epistemology, or its theory of knowledge/understanding. Because of the black box nature of the epistemology of current LRATs like ChatGPT and Lexis+ AI/Protege, legal researchers have no effective way, outside of the prompting methods discussed in Section IV-B below, to ensure the LRAT truly understands the prompts from the legal researcher such that it would not confabulate/hallucinate legal resources. As such, it is important for researchers to understand that not all AI tools generally or LRATs specifically are appropriate research tools in every context.

IIA-B. AI THEORY: HOW AI UNDERSTANDS ITSELF – DEEP LEARNING

If ML can be described as the process by which an LLM or LRAT ‘learns how to learn’, similar to the process of a toddler learning how to walk, then deep learning (DL, hereinafter) can be described as the process by which an LLM or LRAT develops informational intelligence, similar to the process by which a human brain learns and synthesises an individual’s knowledge of data into conceptual intelligence over a course of time.Footnote ⁸ In simpler terms, DL is analogous to an algorithmic theory of knowledge, or an algorithmic epistemology, for LLMs generally and LRATs specifically. To continue the earlier toddling analogy, a human individual that possesses the intelligence to know how to walk can then utilise that intelligence in different ways; for example, that individual can use their knowledge of how to walk to inform their process of learning how to run or even to sprint. In an analogous way, the process of deep learning in AI allows for an LLM or LRAT to be able to develop a sort of limited ‘intelligence’ that itself possesses the knowledge of how to perform distinct tasks, including summarising or comparing/contrasting.

These processes of running ML and DL algorithms on information allow LRATs (as currently designed) to function as AI-powered search tools, whereby human researchers may prompt the LRATs but where the LRATs do not themselves function as a research partner, but more as a sounding board; however, applied in the area of legal research, these types of ML and DL algorithms are not enough to provide the type of dynamic legal reasoning systems/legal epistemic models that would be necessary to allow LRATs to function as true research partners rather than sounding boards. Although this is beyond the scope of the current article, the author is in the process of drafting a further article on his ideas for such dynamic legal reasoning systems/legal epistemic models, focusing on AI as a recursive intelligence that (1) encodes data through processing statutes, case law, secondary sources, and other legal information resources; (2) recognises patterns through identifying trends in judicial reasoning; and (3) adapts contextually by adjusting its responses based on both legal context and user interaction.

As described in the introduction to this article, DL is a subset of ML, which is itself a subset of AI more broadly. Although AI as a field was limited in its usefulness as applied to the real world during its infancy due to its reliance on deterministic algorithms that were only applicable in certain supercomputer or defence contexts,Footnote ⁹ the advancements of ML in the 1990s and then DL in the 2010s helped to make the theoretical substructure of AI as it is currently understood operable for human researchers. AI could and did exist without DL at a basic level, like the tax software using AI from the 1990s which used ML algorithms.Footnote ¹⁰ However, for DL to be effective in developing informational intelligence, a certain organisational information paradigm had to be developed through which those algorithms could be run, and a technological innovation had to occur through which the resulting information, post running of those algorithms, could be made intelligible to human users – the paradigm being neural networks and the innovation being transformers.

IIA-C. AI THEORY: HOW AI UNDERSTANDS ITSELF – NEURAL NETWORKS AND LINEAR REGRESSION VIA GRADIENT DESCENT

Before the introduction of DL algorithms in the field of artificial intelligence, AI was a different beast entirely due to its reliance on ML alone because DL algorithms had not yet been developed. At the genesis of AI during the Dartmouth Summer Research Project in 1956 and because of the above fact, the apparent applications of AI science at the time to more generally useful real world use cases were more modest than the applications seen today. This is largely due to AI’s early focus on attacking practical problems in perception rather than on developing a general intelligence.Footnote ¹¹ However, some attendees of the Dartmouth Project signalled their openness to including the study of human brains and how they learned and developed intelligence as part of the field of AI to eventually approximate “computer brains,”Footnote ¹² similar to what DL eventually became. However, as the study of the brain proper was a relatively new field with a dearth of evidence to back it up at the time, the minds behind the Dartmouth Project concluded that the Project itself would not include brain studies as part of the field of AI. This limitation of AI’s applicability prevented the development of technology that would be able to learn and analyse concepts to such a degree that it could be useful in a variety of different use cases.Footnote ¹³

Given the technological limitations of using only ML algorithms and because the scope of AI-enabled tools was limited to solving problems and did not allow for the ability of those tools to ‘learn’ or ‘reason’ on a more advanced scale, any such tool using only ML algorithms could not be scaled up to allow for such learning and reasoning so as to give that tool the ability to answer questions that were new to it and not programmed into it or learned by it using ML algorithms. During the 1990s, the AI-enabled tools of the type described above were introduced to the public as expert tax systems; these expert systems could only help to organise the responses put into them by human users into categories in a deterministic sense,Footnote ¹⁴ like how a human thinker can organise responses to third party questions by using a deterministic decision tree. However, because these ML-powered tools lacked the reasoning capabilities that would come to be endowed later by the application of DL algorithms, they were not usable for the purposes of research. This limited usefulness led to a massive drop in R&D funding for AI technologies, and this helped to precipitate the AI Winter of the 1980s,Footnote ¹⁵ wherein AI received much less funding than it received in the prior three decades of AI R&D. This AI Winter helped to cause the public perception of AI in the 1980s, led by the perception of the characters of The Terminator, Skynet, etc., as incapable of processing thoughtful analysis and as inherently anti-human, which are biases that have staying power.

However, with the advent of DL algorithms in the mid-to-late-1980s, the ability of AI technology to mirror real life – due to that technology’s ability to apply its learning in more than domain-specific uses to answer more general questions – began to seem achievable and it became necessary to determine whether computers could approximate a human brain through technological adaptations so as to be able to develop narrow AI that would be useful in answering questions.Footnote ¹⁶ Although work on such a technology was called for in the AI community by a foresighted thinker among that community’s founders,Footnote ¹⁷ the existence of work to develop such technology took decades to work its way into the public consciousness given the earlier AI Winter and public bias against non-ethical AI, and this work culminated in data scientists developing narrow, extractive AI through the technological innovation of neural networks.

Neural networks are, in essence, computer ‘brains’ that mirror the real-world neural networks that are composed of billions and billions of neurons composing and surrounding the human brain and that allow the human nervous system to execute tasks such as physical movement, speech, or emotional recognition.Footnote ¹⁸ These nervous system neural networks possessed by human beings allow for advanced pattern recognition and the execution of complex thoughts and tasks. Neural networks as applied to LRATs perform a similar function, allowing for LRATs to perform much more complex pattern recognition and problem-solving tasks – tasks that would be difficult for a human to perform at a consistent level. These neural network ‘brains’ are complex, but researchers need to keep four main concepts in mind to understand how they function; namely (1) that neural networks are composed of node layers, (2) that they use linear regression models, (3) that data is passed between the nodes in the network in a feed-forward manner, and (4) that neural networks rely on training data to further train themselves through the process of gradient descent.Footnote ¹⁹

Node layers are, quite literally, different layers of nodes in a neural network that are arranged into input layers, hidden layers, and output layers.Footnote ²⁰ These nodes are modelled after neurons within a human brain and connect or correspond to other nodes within the neural network in a comparable manner to neural connections in a brain.Footnote ²¹ Each node uses a linear regression model to help it find the relationships or connections between multiple data variables and predict future outcomes based on its identification and analysis of those relationships in the data.Footnote ²² Essentially, the weighting of those connections (using Natural Language Processing algorithms – discussed in depth in a further section of this article) between the nodes in the network, after those connections are found and analysed, determines how much impact each input layer (which identifies and analyses the data put into the neural network from human prompts) will have on the output layer. The data and its many connections flow through the nodes in the network in a feed-forward manner, meaning that the data flows in one direction from the input layers through the hidden layers which identify more data connections to the output layers.

A neural network relies on training data, which means that the data passed through the network trains the network to minimise its cost function – which itself is a mathematic function that evaluates the accuracy of a neural network’s data relationship predictions through instructing the model to adjust its weighting and biases that are inherent in those data predictions to better fit the training data – and to optimise its performance over time in learning from the training data and in executing the tasks set to it. This training is performed through the process of gradient descent,Footnote ²³ which, although it is key to understanding the technical aspects of how neural networks train themselves, is itself beyond the scope of helping legal researchers understand how LRATs work and will not be analysed here, because the mathematical specifics of gradient descent exceed the needs of this article. However, these neural networks, even advanced ones, would not be useful for the purposes of legal research without the technology of transformers.

IIA-D. AI THEORY: HOW AI UNDERSTANDS ITSELF – TRANSFORMERS

Transformers, although they debuted in the 1980s, have experienced an explosion in popularity in the last three years due to an extremely popular AI product – ChatGPT. The massive popularity of ChatGPT has shown that transformers are remarkably effective at conversion/translation of mathematical and numerical values and vectors into cognisable, quasi-intelligent language. To a legal researcher audience, however, given the legal industry’s relatively quick introduction of AI-enabled tools to the wider legal population, a proper understanding of how transformers operate within LRATs will help legal researchers to understand how AI-enabled tools generally and LRATs specifically operate, which in turn will make them more efficient AI and AIALR researchers.

There are three main components of transformers that legal researchers should understand, namely (1) Byte Pair Encoding (BPE, hereinafter), (2) encoders, and (3) decoders. Transformers function in LRATs like translators, converting token sequences into readable output using encoder-decoder architecture and functioning as the LRAT’s ability to take sequences of token values in their neural network’s input layer (again, tokenisation is discussed in a subsequent section of this article) and predict the next word in the output sequence through the process of byte pair encoding and taking all the encodings and using context derived from the training data to generate the transformer’s output sequence – which is what a researcher will see as the answer to their prompts – and iterating it through encoder layers, which generate encodings that define which parts of the input sequence are relevant to each other and passes the encoding onto further encoder layers, which correspond to the hidden layers in a neural network.Footnote ²⁴ A transformer’s main values to a legal researcher are that (1) they translate a researcher’s plain language prompts into numerical values that are readable by the LRAT such that it can respond with an accurate answer to those prompts by translating the numerical values informing its answer to the researcher’s prompt into a readable format for the researcher.

IIB-A: AI FUNCTIONALITY HOW AI UNDERSTANDS HUMAN RESEARCHERS – ALGORITHMS

If human researchers want to grasp how AI-enabled tools understand what humans mean or intend when they issue instructions to those tools, then the fundamental concept they must understand is that of the algorithm, and how algorithms function in the context of AI. Algorithms are, in essence, well-defined step-by-step procedures that are designed to solve a specific problem or perform a specific task or tasks.Footnote ²⁷ In addition, algorithms are almost always mathematical.Footnote ²⁸ What this means in the context of legal research is that algorithms are effectively mathematical or numerical ‘rule sets’ that LRATs must follow when given use case tasks to complete by human researchers such as searching or summarising. This also means that the neural network technology in LRATs, discussed in section IIA-C, follows the instructions of these algorithms (IIB-A) and that the transformer technology in LRATs, (IIA-D), processes, analyses, and translates data using these ML- or DL-based algorithms.

It is important for legal researchers to bear in mind that algorithms are a rule set and not AI-enabled technology itself. AI-enabled technologies (such as ChatGPT) and LRATs (such as Lexis+ AI) generate answers through utilising (1) the data architecture of their neural networks and (2) the mathematic-to-plain-language translation abilities of transformers bestowed on them by ML- or DL-based algorithms. However, the legal researcher must understand that these algorithms are part of how the AI-enabled tools/LRATs themselves formulate their answers and not the tools themselves.Footnote ²⁹ These algorithms rely on teams of data engineers for their creation and power LRAT searches.Footnote ³⁰ In addition, they are often not publicised or shared by the businesses (or teams of data engineers) that created them.Footnote ³¹ Because algorithms are the LRAT ‘rule books’ on how to perform actions and cannot be altered by legal researchers on the back end, these legal researchers have no effective control either over the LRAT’s algorithms themselves or over how LRATs process data using their ML- or DL-based algorithms, unless such information is made public. However, understanding algorithms and how they interact with the data architecture within LRATs and the algorithmic processes that power them will help legal researchers to have a modicum of control over a process that can, by definition, prompt an LRAT to alter or cabin its reasoning to give accurate answers to their questions, which will be discussed in a further section of this article.

Algorithms influenced by ML and DL are, in a manner of speaking, an LRAT’s equivalent to a human researcher’s symbolic logical reasoning or procedural thinking functions in that algorithms provide the steps and instructions on how an LRAT is to process the information retrieved by it from a knowledge database as a part of the research process.Footnote ³² There are four similarities between algorithms and the logical/procedural reasoning of human intelligence, namely (1) that both operate in a step-by-step process (i.e., both are following a rule set), (2) that both are problem-solving processes (i.e., both rule sets are set up to solve problems), (3) that the decision-making inherent in solving problems in both processes are rules-based (i.e., that the decision-making entity does not make up the rules themselves), and (4) that both processes are finite in that they have an ending point (i.e., algorithms and logical reasoning are ‘on’ while they are being used and not when they are not).

Thinking about algorithms using this framework of similarity to a researcher’s logical reasoning allows legal researchers to properly identify the relationship of algorithms within/to LRATs and to properly understand how LRATs use algorithms in the context of their ML or DL processes. This understanding of both these concepts will assist researchers in understanding the tools with which they are working when conducting AIALR. However, as intimated above, algorithms do not run forever unless they are set up to do so, and as also intimated above, algorithms are not LRATs but are rather part of the process of how those LRATs answer questions using Natural Language Processing (NLP) and the tokenisation process. Therefore, for the purposes of legal research, algorithms by themselves cannot answer legal research questions, but how LRATs understand actual human language through Natural Language Processing and the tokenisation process so that they can give an answer to such legal research questions remains to be seen.

IIB-B. AI FUNCTIONALITY: HOW AI UNDERSTANDS HUMAN RESEARCHERS – NLP & TOKENISATION

If ML is the process by which a LRAT ‘learns how to learn’, then any processes that teach LRATs how to do a task are ipso facto based out of ML. Natural Language Processing (NLP, hereinafter) can be described as ML-based algorithms that teach LRATs how to interpret, manipulate and understand human language in such a way as a human can so as to complete a task. This definition is merely a descriptor of what an NLP-influenced algorithm does, and it does not detail how the algorithm itself works, but that is because it relies on a complex mathematical process called tokenisation, which is explained below. Tokenisation itself shapes AI’s understanding of language generally and legal precedent and statutory language specifically by tokenising written language into fragmented units, which can cause distortions in legal interpretation through lexical segmentation, context preservation challenges and sentence fragmentation risks, leading to a high impact on the parsing of legal phrases or terms of art. To be definitionally clear, however, tokenisation is both a step in the NLP process as well as a mathematical process itself.Footnote ³³

Tokenisation is the process of breaking down input text into small or smaller word units called “tokens,”Footnote ³⁴ which can be words, or even sub-words or characters. The process of tokenisation may be divided into three steps of its own, those steps being (1) Pre-Tokenisation, (2) Vocabulary Construction, and (3) Segmentation. These steps, while vital for understanding the ins and outs of how NLP works, are beyond the scope of this article, which focuses on a legal researcher’s understanding of the processes that make up AI to better use LRATs. It should suffice legal researchers to know that once words are tokenised, each token is converted to a vector, which is a numerical string representing a point in a multi-dimensional space – like an X/Y axis but with more dimensions than two – which is then embedded mathematically and numerically at a certain point on that multi-dimensional axis and given a value.Footnote ³⁵ A vector itself is a quantity that cannot be expressed in a single number; therefore, a token’s vector value on the aforementioned multi-dimensional axis includes its value on the X, Y, and Z axes (length, height, width, etc.Footnote ³⁶). The value of these vectors on those axes are then weightedFootnote ³⁷ by the neural network through comparing the value of the vectorised and tokenised words of the research prompt to the value of the vectorised words of the LRAT’s knowledge database to determine how much attention the LRAT should pay to each token when generating the next token in the sequence of its answer to the prompt.Footnote ³⁸ This weighting process, discussed earlier in this article and entitled linear regression via gradient descent, takes the yhat – or the values that the AI-enabled tool predicts will be the answers in a regression model – and weighs those yhats against what was found in the database searched by the LRAT to determine the accuracy of the answer.

What this means, in non-scientific phraseology, is essentially that words that have been vectorised by an LRAT into certain values on the aforementioned multi-dimensional axis can have similar weights as other vectorised words, and such similar weighting can lead to an internal conclusion by the LRAT that such a word possesses some kind of similar relationship to another word, which can inform both the tool’s reasoning and answer to the prompt. This tracking of similarly weighted vectorised words also allows the LRAT’s DL-influenced algorithms discussed earlier in this article to learn from the relationships between the similar weights of the vectorised words such that the tool itself can answer the legal research prompts it receives by answering the prompt with tokenised words with weights that correspond to the weights of the tokenised words in the prompt, and then the NLP algorithms convert those numerical vectors on that multi-dimensional axis into human language answers.

However, even though LRATs possess the ability to learn through applying ML-influenced algorithms, even though they possess the somewhat cabined ability to reason through applying DL-influenced algorithms, and even though they possess the ability to analyse the meanings of human words given to them in prompts such that they can give intelligible answers to those prompts through the processes of NLP and tokenisation, LRATs cannot become truly effective without the human researcher using the LRAT learning how to ask the proper questions to the tool such that any answer given by the tool is accurate and on-point, which is absolutely necessary in the paradigm of legal research. This process, in which researchers do have control over unlike algorithms or AI theoretical concepts, is prompting, and it will be discussed further in the next section of this article. The use of the crafting of prompts by legal researchers as a part of the AIALR method of conducting legal research is prompt engineering, and this will also be discussed in a later section of this article.

IIB-C. AI FUNCTIONALITY: HOW AI UNDERSTANDS HUMAN RESEARCHERS – PROMPTING

AI Prompting is the process by which human researchers write prompts to an LRAT, which by itself is not a very sensible definition without next defining prompts themselves. AI Prompts are (1) written in plain language of various types and (2) contain specific instructions and/or queries to shape the answers given by an AI-enabled tool/LRAT.Footnote ³⁹ The various types of plain language that researchers can prompt with are instructive language, Q&A language, contextual language or mixed language.Footnote ⁴⁰ These definitions are largely defined by common sense: where researchers instruct an LRAT using commanding language (Instructive Prompting), where researchers question an LRAT using structured and specific queries (Q&A Prompting), where researchers question an LRAT using additional researcher-given context (Contextual Prompting), and where the researcher combines these methods (Mixed Prompting).Footnote ⁴¹

As LRATs follow algorithmic rule sets when answering questions, these LRATs also follow prompting from the researcher as to how/when to run a task, how to alter its process of answering, or where to look for its answer. As can be seen by the fact that LRATs’ algorithms do not continually run unless the researcher instructs them to do so, LRATs will only ever perform what legal researchers instruct them to perform. Therefore, when a researcher is prompting an LRAT, the instructions given in their prompt are executed by the LRAT exactly as they were written, subject to the weights of the tokens that the LRAT examines. The dangerous side of this, however, is that if researchers do not prompt the LRAT in an easy-for-it-to-understand way, that task will not be performed reliably. This danger inherent in prompting seems to suggest that, as far as currently existing AI-enabled tools/LRATs are concerned, the methods by which legal researchers have conducted research in databases and in physical libraries might not always lead to reliable research without the researcher themselves developing additional skills. To juxtapose this danger against the traditional reliability of database research for a moment, if a researcher searches using Boolean/Terms & Connectors, the Boolean/T&C search is reliable so long as the researcher knows the language of the database, which can be different between databases.Footnote ⁴² Likewise, an LRAT search can be reliable so long as the researcher knows the language of the LRAT, or, more properly, knows exactly what they are saying/typing.

The cognitive tension inherent in prompting an LRAT as Artificial Intelligence requires researchers to think not in the terms of a computer-mathematic language as they would for searching in databases using traditional Computer-Assisted Legal Research (CALR, hereinafter), but rather in the terms of the LRAT itself and how/what it considers in answering prompts. To do this, researchers must always be aware of the words they are using so that they can ask questions that will guide them to their answers effectively, efficiently, and reliably. Given this tension between the LRATs only doing what it is told to do and the researchers not always knowing what words or phrases they might be missing in their question that the LRAT would normally perform if instructed to do so, the question remains of how a legal researcher can use LRATs to perform AIALR, or AI Legal Research. The key to understanding AIALR lies in two methods: The Five Ps Research Plan and Prompt Engineering.

IV-A. AIALR: HOW TO DO AI RESEARCH – THE FIVE PS RESEARCH PLAN METHOD

As was hinted at in the prior section of this article, any new system of legal research dealing with LRATs will require two things: (1) an equivalent to the CALR research plan method to help law students make sense of the subtle differences that exist between AI research plans and CALR research plans and (2) an equivalent to the Boolean search language to empower law students with a tool in their tool belt to find legal resources in the AI paradigm. This article recommends that the first of these requirements of an AI research plan can be met through utilising what is called The Five Ps research plan method. The Five Ps research plan, at its core, is a method of epistemological framing of research to allow the LRAT to cover the pre-research elements of research, i.e., thinking of words to use in searches, thinking of jurisdictions, and thinking of the issues in the case, before responding. It shares several similarities to the Five Ws research plan commonly used when conducting CALR-based research. These Five Ws may be found in footnote 52 of this article, but to reiterate them, they are (1) who, (2) what, (3) when, (4) where, and (5) why.

These Five Ws are sensational language, meaning that they refer to human senses, and they aid legal researchers using CALR to figure out what would be useful to put into a Boolean search string. The ‘who’ often refers to the parties in the case, the ‘what’ usually refers to the thing that happened that caused the case to be a case, the ‘when’ often refers to timing of the case (think of statues of limitations), the ‘where’ usually refers to the jurisdiction in which the events took place, and the ‘why’ generally refers to the mental state of the parties, or mens rea. The usage of sensational language helps each researcher to put themselves into the case that they are looking into such that they could produce words to translate into Boolean search language to find materials that would help them answer the question they are researching.

On the other hand, the Five Ps research plan involves five distinct concepts/words starting with the letter “P.” These Ps are not sensational language in the same manner as the Ws referred to above but instead help the human researchers to orient their thinking around the fact that there is an artificial intelligence helping them to conduct research. These Five Ps are (1) Prime, (2) Persona, (3) Prompt, (4) Product, and (5) Polish. Each of these Ps corresponds to a step in the traditional CALR research plan except for the fourth P, which will be examined below in its own paragraph.

The first of these Ps is Prime P, and this refers to the big, named issues and facts in the inquiry. Once a researcher thinks of their Prime Ps, they will help themselves to keep the facts of the matter straight, which will help them prepare for usage of the LRAT.Footnote ⁵³ This P corresponds to the first, second, and third steps in a traditional research plan.

The second P is the Persona or Party P, and the Persona/Party P asks which persona, or party, you – as the researcher – are representing in the litigation. Once a researcher identifies who it is that they are representing and in what legal context, it is much easier to identify some party-specific legal elements, such as affirmative defences or burden-shifting, in one’s AI research.Footnote ⁵⁴ This P corresponds to the first and third steps of the traditional CALR research plan.

The Third P is the Prompt P – this P seeks the researcher to look to the call of the question they are trying to answer. Prompt P is intended for legal researchers to look at the call of the question or issue they are seeking to answer so that they can see any terms in the document(s) they have been given to work with.Footnote ⁵⁵ This P corresponds with steps two and three of traditional legal research plans.

The Fifth P is the Polish P, and the point of researchers considering the Polish P in their AI research plans is for it to function as a catch-all step where a researcher considers what they have not considered yet. This step can lead to narrowing of research, expanding of research, or the pivoting of research, depending on what is found in the results by the researcher.Footnote ⁵⁶ This P corresponds to step 3 of the traditional CALR research plan.

Product P is the fourth step of the Five Ps, and it is the only step that is unique to AIALR.Footnote ⁵⁷ This is because in CALR-based researching, the research plan that researchers do is intended to help them to think of terms for the purposes of making a physical notation of what was searched; in CALR, the researcher has no ability to augment what the tool they were using is doing. However, in AIALR, because LRATs are AI-enabled tools, researchers practicing AIALR do possess the ability to augment how their LRATs perform the tasks they are set out to do.

The essential reason for researchers to consider Product P is that it makes them think of the format the answer to the question they asked should take. Because legal researchers can augment what the LRAT they are using is doing, those researchers can alter the product that the LRAT delivers. For instance, using ChatGPT, a prompt was used to generate the American Pledge of Allegiance using the phraseology of Bill & Ted’s Excellent Adventure through including my thoughts on what the finished work product should look like vis-à-vis including it in my prompting to the LRAT.Footnote ⁵⁸ An example of where a research deliverable can be altered exists, here for the purposes of legal drafting, would be where a researcher might ask an LRAT to draft a client letter in different tones and styles. Thinking of Product P helps the researcher develop a habit of checking what format their deliverable should be in, which will help that researcher to always think of the changeable format their work product should be in.

IV-B. AIALR: HOW TO DO AI RESEARCH – PROMPT ENGINEERING

Because AIALR is so novel, AIALR research requires not only an AI-based version of the Research Plan but also requires a new search language or syntax to make AIALR searching easier. Thankfully, during the last few years, the burgeoning prompt engineering field has taken AI-enabled tools and LRATs by storm. Prompt engineering itself is performed by legal researchers through instructing an LRAT’s algorithm to perform a task. The researchers, using their knowledge of how AI thinks about itself and of how AI thinks about human researchers, can practise prompt engineering informed by that knowledge to reliably retrieve legal information effectively and efficiently. It is important to keep in mind here that what a legal researcher is doing while doing CALR-based research and what they are doing while performing AIALR are substantively different from each other.

In CALR, a researcher plans their research and then executes it using a specialty language called Boolean search language/Terms & Connectors searching and converting the terms they conceived of from their research plan into a mathematical Boolean search string. This search string is then readable by the database as search instructions, which will be executed by the database using the instructions given in the string exactly as they are written. However, in AIALR, a researcher plans their research and executes it using plain language searching, but this plain language searching is more powerful when certain strategies of AI-based plain language searching (read: Prompt Engineering) are utilised.

There are four common prompt engineering methods that researchers can use to prompt an LRAT, these being Retrieval Augmented Generation, or RAG, Few Shot Prompting, or FSP, the Chain-of-Thought or Chain-of-Logic method, or COT/COL, and Prompt Chaining, or PC. Each of these methods of prompt engineering is useful in different contexts, which will be discussed below.

Retrieval Augmented Generation is extremely useful in the use case of legal drafting. When a researcher practices RAG, they instruct the LRAT first to retrieve relevant information from a database or document, and then to answer or generate only based on the retrieved information. Prompting in such a way will help researchers prevent LRAT hallucinations/confabulations and will help enhance the relevance of the text that is generated by the LRAT.Footnote ⁵⁹

Few-Shot Prompting is especially useful in contexts where an LRAT has never had to do a particular task before and needs education on how to run a search. In FSP prompting, a researcher uses examples of how they want the task assigned to be completed by the LRAT. Prompting in this way ensures that the task is performed accurately by the LRAT per the parameters of the researcher because this style of prompting requires the LRAT to follow the examples given by the researcher.Footnote ⁶⁰

COT/COL Prompting is useful in situations where a researcher uses an LRAT to conduct a chain of research that requires complex or abstract legal reasoning. When a researcher uses COT/COL, that researcher requires the LRAT to logically reason in a step-by-step manner in its response. This style of prompting allows researchers to break down logical leaps made by the LRAT, ensuring that the logic of the LRAT is sound. In addition, this style of prompting allows LRATs to think through more complex or abstract reasoning.

Lastly, Prompt Chaining is useful in situations where a researcher is working on a line of research that requires multiple iterations. In prompt chaining, a researcher uses the output of their prior prompt and puts it into the prompt of the next question to the LRAT. Essentially, this style of chaining is useful when there are multiple levels of abstractions required or logical leaps being made by the LRAT because the researcher can use the answer to their question once the abstracted question is answered, and this is done through plugging the output of the last prompt into the input of the next prompt. Prompting this way allows researchers to break down complex tasks or tasks with multiple abstractions into simpler steps, which opens the way for more complex research. In addition, prompts such as the above tend to be repeatable given how the context is carried over from one prompt to the other. When using prompt chaining to research in the law, the first prompt is generalised and focuses on the broad legal issue, whereas the second prompt narrows the jurisdiction, precedent date range, and legal doctrines at play. The third prompt clarifies that the LRAT’s interpretation matches the legal intent, and the fourth prompt cross-references the output with primary legal sources.

However, one aspect of prompt chaining to keep in mind is that researchers must deal with the lost middle paradox, which is best understood by a legal researcher through thinking of a conversation between two parties. When a party is chatting with another party, it is more likely that that party will remember the context at the beginning and at the end of their conversation with the other party than the context in the middle of their conversation. In a similar way, when researching using LRATs, due to the tokenisation process (which is limited for each LRAT), if one is prompt chaining their prompts to work on an abstract research question, the LRAT will also remember less context from the middle of the query. This is because LRATs are limited by tokenisation limits, wherein if the usage of the LRAT has not breached the tokenisation limit and the LRAT is reminded by the legal researcher of the context in the prompt chain, then this would, at least theoretically, sidestep the lost middle and allow for less confabulations. However, if the usage of the LRAT breached the tokenisation limit, the LRAT will begin to need to be reminded of the context that was lost in the lost middle. The lost middle is important to keep in mind, especially in deeply involved research over time, because confabulation/hallucination, or the forgetting of epistemological context, occurs in the lost middle alongside tokenisation limits, not at the input or output boundaries.

All five of these prompt engineering methods – RAG, FSP, COT/COL, and/or PC – are useful in AIALR because a researcher has agency and power over their prompting where they do not have such control vis-à-vis how the LRAT thinks about their question or which algorithms it will use to get to the answer to that question. Understanding how to prompt on the part of the legal researcher will be akin to knowing Boolean searching, which is a much more efficient and exact method of research as compared to other popular methods. However, at the stage of initial legal research, where a researcher is using an LRAT to conduct legal research, those other popular methods might not be as useful.

When legal researchers have done enough brainstorming in their research plans for terms such that they feel comfortable initiating CALR-based research, the methods by which they do so are multifarious and useful within the context of CALR research, but these methods are not as useful when practicing AIALR. An example of this method, which is utilised when a legal researcher has found a resource (for this example, a case) but does not know where to start researching from that point, is the ‘one good case’ (OGC) method. This method exists so that legal researchers can broaden their sources to find other cases with similar outcomes by examining certain metadata about the case and using that metadata to find the other cases. The three main methods by which legal researchers use the one good case method are (1) looking backwards, (2) looking forward, and (3) looking around. When looking backwards, certain citator tabs exist in the legal research databases that one might see containing information about the Table of Authorities (the authorities cited to in the one good case). When looking forwards, case headnotes and digests are useful because they can send researchers to other cases that share the same headnote since they are similar issues.

The OGC method, while useful especially in CALR research where the researcher has control over the parameters of research, is less useful in AIALR because the LRAT would be made to do other tasks besides extractive legal research to explain to the researcher the cognitive importance of what was analysed – i.e., whether a case is an OGC. To explain, because AI-enabled tools and LRATs produce text via a limited but substantial number of tokens during the tokenisation process, discussed in section IIB-B, they are somewhat limited in their current architecture regarding running the other steps of the OGC Method (looking back, forward, and around) in order to analyse and determine if a case is an OGC, which would make them less effective due to their not focusing on how many tokens they have with which to produce their response.

In addition, the OGC method works well precisely because CALR-based researching uses citator tabs (looking backward) and case headnotes/digests (looking forward) to analyse which other cases would be useful in research for the question at hand. Because an LRAT finds legal materials, links to legal materials, and moderately explains those materials, depending on what is asked, it does not analyse other good cases and use up tokens. What this means is that those other good cases that are found using the OGC method of searching are covered and highlighted in the first sweep of the LRAT. Essentially, the LRAT cuts out the ‘middleman’ of the research process, which is finding other cases using metadata citators. Because LRATs essentially perform the OGC method as a matter of course in their initial sweep, using the OGC method for AI research to uncover further cases would lead to a lower return on the research investment. This does not diminish the usefulness of the OGC method in traditional CALR research, but it does show the limits of its usefulness as applied to the age of generative artificial intelligence. In addition, this exhibition of the limits of traditional CALR methods of research in AI simultaneously displays the necessity of a fundamental epistemological shift on the part of legal researchers – it requires legal researchers to develop their skills in AIALR by recognising how AI works inside the black box, how it understands its own internal working environment and how it understands those it is interacting with, and how they themselves as researchers must alter their research style when using AI-assisted research tools generally or LRATs specifically.

IV-C. CONCLUSION – WHERE WILL AI LEAD LEGAL RESEARCH?

In this article, I have traced the contours of a current iteration of a Generative/Extractive Artificial Intelligence – ChatGPT – for legal researchers such that They have a better understanding of Artificial Intelligence, thereby improving their own general skills of legal research as well as preparing themselves for the next iterations of AI research. This was done through briefly sketching the outlines of AI’s history and discussing the lack of a common definition for the field of AI, connecting this history and lack of a common definition for the field to users’ lack of understanding of how AI works. Then, I explained several AI concepts and how they work, connecting Machine Learning, Deep Learning, Neural Networks, Linear Regression via Gradient Descent, and Transformers to Algorithms, Natural Language Processing, Tokenisation, and Prompting. Each of these concepts plays a role in how AI-enabled tools/LRATs function, and my explanation of these connections shows how legal researchers, even after understanding these concepts, should (1) utilise a new species of AI research plans (The Five Ps Research Plan) and (2) utilise a new search grammar for AI (Prompt Engineering).

Earlier, I mentioned that legal researchers must make a cognitive shift from working in CALR to working in AIALR. This cognitive shift must reflect that AI-enabled tools/LRATs are not merely machines or informational architectures; they are, or in the context of legal research can be, research quasi-agents that, although not ontologically human, can operate in human-like, recursive ways. Although this is beyond the ambit of this article, a subsequent article will attempt to define this recursive AI and its attendant architecture. This cognitive shift, from viewing AI as a tool to viewing AI as an epistemic quasi-agent, will not relieve the legal researcher from the burden of accurate legal research. Even with such recursive AI being able to function as research quasi-agents, the responsibility of epistemic interpretation remains with the human legal researcher, especially given this article’s definition of AI in legal research terms.

In the new recursive research paradigm that I will endeavour to articulate in my next article, the law librarian cannot be merely a navigator of databases; rather, the law librarian must become a curator of epistemic clarity, wherein they guide researchers through recursive inquiry architectures. In other words, if CALR taught legal researchers to search databases using specific languages and syntaxes, AIALR demands that legal researchers prompt with heightened care, that they reflect recursively on their own lines of inquiry, and that they never forget that behind every legal inquiry is a human judgment still waiting to be made; these demands require that law librarians become the curators of inquiry through developing their prompting skills in concert with their research quasi-agent LRATs.