AI Glossary
This is a curated reference guide that defines key artificial intelligence and generative AI terms in clear, plain language while linking each concept to concrete legal workflows, risks, and best practices. Its purpose is to help lawyers, paralegals, and business professionals quickly understand emerging terminology, apply it to tasks such as research, drafting, due diligence, and litigation, and ensure consistent firmwide usage while promoting responsible, compliant, and effective adoption of AI technologies.
1. Core Concepts
Artificial Intelligence (AI)
Artificial Intelligence refers to the theory and development of computer systems able to perform tasks that normally require human intelligence. It's a broad field encompassing everything from simple rule-based systems to complex models that can learn and adapt. AI is the umbrella term for technologies that can augment or automate cognitive work.
Example:
Microsoft Copilot is a clear example of AI in practice. When a lawyer asks it to "summarize the key points from this discovery document and draft an email to the senior partner about it," the system is using AI to understand the request, read and comprehend the document, extract the most important information, and then compose a professional communication—all tasks that would otherwise require human intellect and time.
Practice Tip:
While a tool like Copilot can perform multiple steps at once, it is best to break down complex requests. In this example, first ask for the summary. Review and verify it for accuracy. Then, in a separate prompt, ask the tool to draft the email based on your verified summary. This sequential approach maintains human oversight and allows more digital processing power to be focused on each individual task, significantly improving the accuracy and quality of the final output.
Ask Yourself:
What is the risk of combining the "summarize" and "draft" tasks in a single prompt to Copilot, and how does breaking them into two separate, sequential steps give me better control over the accuracy and quality of the final work product?
Machine Learning (ML)
Machine Learning is a subset of AI whereby systems are trained on large datasets to recognize patterns and make predictions or decisions without being explicitly programmed. Imagine a vast digital ecosystem where we unleash not one, but thousands of initial AI models, each with a slightly different, randomized configuration. We don't give them answers; we give them a test—a massive dataset—and the single goal of performing a task with the highest possible accuracy. The models that perform best, even by a fraction of a percent, are selected. Their successful characteristics become the blueprint for a new generation of models, which are then copied with slight mutations. This cycle of testing, selecting, and refining is repeated millions of times.
Example:
When an attorney uses Lexis+ AI for legal research, machine learning is working behind the scenes. The system has been trained on millions of legal documents and user search patterns. It uses this training to predict which cases are most likely to be relevant to the attorney's query, ranking the most important results at the top. This predictive ranking happens before the generative AI part even begins to write a summary.
Practice Tip:
Machine learning outputs are probabilistic, not deterministic. An ML model classifying documents for relevance, for instance, provides a likelihood score, not a definitive judgment. In contrast, a generative AI tool creates new content. While both can be wrong, the nature of the error is different—one misclassifies, the other fabricates. Understand which type of AI you're using to properly assess the output.
Ask Yourself:
How might the predictive nature of machine learning influence the information I see first? Am I more likely to click on the top results, and how can I ensure I'm not missing a key piece of information that was ranked lower by the algorithm?
Deep Learning
Deep Learning is a type of machine learning that uses many connected layers of “artificial neurons” to find patterns in big sets of data. Each layer learns something a little more complex than the one before—like recognizing letters, then words, and then full ideas. It’s what powers advanced AI tools such as ChatGPT and image generators.
Example:
The "brains" of Harvey and Lexis+ AI are built using deep learning. This allows them to understand the subtle nuances of legal language, summarize complex documents, and generate coherent, contextually aware text in a way that older technologies cannot.
Practice Tip:
Deep learning models excel at understanding context and nuance, making them powerful for tasks like summarizing depositions or identifying themes in discovery. However, their complexity can make them "black boxes," meaning, you can see the input and the output, but not always how the model got there. Therefore, always treat the output as a highly sophisticated suggestion that requires human verification, not as a definitive answer.
Ask Yourself:
When a deep learning tool provides a summary of witness testimony, what is the key difference between what the AI is doing and what a human paralegal does? What kinds of subtle cues or strategic implications might the AI miss that a human would catch?
Neural Network
A neural network is a computer system that works a bit like the human brain. It has many tiny parts called “neurons” that are connected in layers—each layer looks at information, makes small decisions, and passes results to the next layer. By adjusting these connections as it practices, the network learns to recognize patterns or make predictions, such as spotting key phrases in a legal document or helping an AI write text that sounds natural.
Example:
The ability of Microsoft Copilot to understand your request in a Word document, process it, and then generate a relevant paragraph is orchestrated by a massive neural network working behind the scenes.
Practice Tip:
Neural networks are nearly as complex and hard to interpret as the human brain. While engineers have been able to create a framework that mimics actual neurons firing, we don’t fully understand what messages are being sent within a neural network to produce the outputs you see in tools like Copilot or PC Chat. These outputs are often highly useful and correct, but not always, and because they act as “black boxes,” engineers typically can’t explain why the outputs are correct at some times but not others. Given this, always verify outputs against authoritative sources—for example, cross-check GenAI-drafted contract language or case summaries in Westlaw/Lexis before relying on them. Treat the outputs from tools built on neural networks as a first pass, not a final authority.
Ask Yourself:
When I use tools like Copilot or ChatGPT, do I take the time to verify the output before relying on it? Are some GenAI tools better set up to help me with this verification process–for example, do they allow me to easily compare the output against my source documents within the platform rather than requiring me to conduct the review manually? Are there times (e.g., when working on creative tasks like brainstorming ideas) when validating the output is not as important?
Natural Language Processing (NLP)
Natural Language Processing is a field of AI focused on enabling computers to understand, interpret, and generate human language in a valuable way. Think of NLP as a translator turning your words into symbols and tokens the computer can work with. It’s the technology that allows you to talk to your AI assistant. NLP covers tasks like text classification, sentiment analysis, and machine translation, and is one of the foundations upon which Large Language Models are built.
Example:
Every AI tool the firm uses that deals with text relies on NLP. When Clearbrief (a tool used by our litigators) scans a brief to find and suggest fact citations from discovery documents, it is using NLP to understand the meaning and context of the sentences in both the brief and the discovery documents.
Practice Tip:
The quality of NLP can vary from tool to tool. If a tool is misinterpreting your request, try rephrasing your text with simpler sentence structures or more direct language to help the NLP model understand your intent.
Ask Yourself:
If an NLP tool fails to find a key document based on my initial search query, what are the ways I could rephrase the query to account for the model's potential misunderstanding of legal jargon or complex sentence structures? Could I use another tool like PC Chat to help me rephrase my query in a more effective way?
Large Language Model (LLM)
A Large Language Model is an expansive NLP model built using deep learning techniques that is trained on vast amounts of text data to understand and generate human-like language. LLMs are the foundation for most modern generative AI tools, enabling them to perform a wide range of tasks like drafting, summarization, and question-answering. If a generative AI tool were a car, the LLM would be its engine.
Example:
The core technology behind tools like Harvey, Lexis+ AI, and PC Chat is their LLMs. When you ask Harvey a question, an LLM processes your text and generates the detailed response you receive.
Practice Tip:
Harvey, Lexis+ AI, and PC Chat all use different LLMs, and some tools let you select from different LLMs within the tool. For example, using PC Chat, try entering the same prompt in the standard mode and again in the “Think Longer” mode (which uses a different LLM). Observe how swapping out one “engine” for another changes the results. Selecting the right LLM for your task can make a big difference. Some LLMs take more time to respond and use more compute power, but produce very detailed results. Others are more creative or efficient at simple tasks. As you become familiar with different LLMs used across GenAI tools, you can start to choose the combination of tool and LLM that is best aligned with the desired output.
Ask Yourself:
For a task requiring a formal research memo with verifiable citations, which LLM-based tool might be a good choice? What tool/LLM might be a better option if I just need help wording an email?
Generative AI (GenAI)
Generative AI is a class of AI models that can create new, original content rather than just analyzing or classifying existing data. This content can include text, images, code, and more. This technology is transformative, enabling efficient legal and client document drafting, deposition and document summarization, and a wide range of other tasks—all initiated from simple prompts. Generative AI is the result of years of advances in machine learning, deep learning, and neural networks.
Example:
Harvey can take a lawyer’s prompt—“Draft a neutral statement of facts for a Rule 12(b)(6) motion in the Northern District of California using the complaint and attached exhibits”—and generate a first-pass draft that includes case-appropriate formatting and inline citations. The lawyer then reviews the draft, checks every citation in Lexis or Westlaw, and applies professional judgment to revise arguments, tone, etc. and make the document their own before filing.
Practice Tip:
In Harvey, treat the AI-generated output as a starting point, not a finished product. Always require human validation of citations, claims, and legal reasoning, and when applicable, save the prompts, outputs, and revised drafts in the matter record for defensibility.
Ask Yourself:
What specific steps do I need to add to my workflow to ensure that every AI-generated citation and factual claim in a brief is validated before it's ever seen by a partner or client or filed with a court? How much time and effort should I expect to spend validating and revising outputs, even if GenAI tools can create the first draft in minutes?
Multimodality
Multimodality refers to the ability of an AI model to process and understand information from multiple types, or "modes," of data simultaneously. This could include text, images, audio, and video. A multimodal AI model can, for example, look at a diagram (image) and answer a question about it (text), providing a more holistic understanding of information.
Example:
Imagine feeding an AI a photograph of a car accident scene along with the text from a police report. A multimodal AI could analyze both the visual evidence (such as vehicle damage and position) and the written report to generate a comprehensive summary of the event, identifying potential inconsistencies between the two sources.
Practice Tip:
When using a multimodal tool, ensure your inputs are clearly related. If you provide both an image and a text document, the AI may assume they are connected. Providing a chart of financial data alongside an unrelated email will likely confuse the AI and produce a nonsensical or irrelevant output. The quality of the AI's synthesis depends on the logical connection between the different data types you provide.
Ask Yourself:
What kind of routine tasks could I streamline by combining different types of information? For instance, could I combine a project timeline (image) with a series of status update emails (text) to get a quick summary of potential delays and risks?
Inference
Inference is the process of using a trained AI model to make a prediction or generate an output based on new, previously unseen input. It's the "live" or "runtime" phase, as opposed to the "training" phase. Every time you submit a prompt to a generative AI tool, you are running an inference task. The speed and efficiency of inference (see Latency & Throughput) are critical for a good user experience.
Example:
When you ask PC Chat a question, the time it takes to give you an answer is the inference time. The model is analyzing your prompt and based on its training, inferring the most probable sequence of words to form a relevant response.
Practice Tip:
Complex prompts require more complex inference, which can take longer and cost more. If you need a quick result, starting with a simpler prompt can be more efficient.
Ask Yourself:
If a complex query in PC Chat is taking a long time to generate a response, what are two ways I could simplify the prompt to reduce the inference time and get a faster, more focused answer?
Context Window
The context window is the maximum amount of information, measured in tokens (roughly, words), that a model can consider at once. This includes your prompt, any documents provided, and the conversation history. A larger context window allows for more complex queries, but also increases processing cost and time.
Example:
When you use an AI tool to summarize a document, the length of the document is constrained by the tool's context window. These limits are usually in the tool's documentation. Harvey’s context window is about 80,000 tokens (~240 pages of text). This includes: (1) The prompt or question you enter, (2) Any documents you attach, and (3) the Conversation history in the thread (your previous questions + responses). When the limit is reached, older parts of the conversation or documents will be dropped automatically. If the initial prompt is larger than the context window, you’ll notice it gets cut off.
Practice Tip:
If a document exceeds the context window, provide only the most relevant sections or use summarization techniques first. Information outside the context window is "forgotten" by the model for that query.
Ask Yourself:
If I need to analyze a 500-page document that exceeds a tool's context window, what is my first step? Should I summarize it first, or should I break it into sections, and what are the risks of each approach?
2. Training & Adoption
Fine-Tuning
Fine-tuning is the process of taking a pretrained AI model and training it further on a smaller, specialized dataset. This adapts the model to excel at a specific task or to adopt a particular style or domain knowledge, such as aligning an AI model so that it reasons and writes more like a lawyer. It is a much more complex and resource-intensive process than simply connecting the model to a document database (see Grounding) or providing a system prompt that always runs in the background. Fine-tuning actually modifies the core model itself.
Example:
Tools like Harvey and Lexis+ AI are fine-tuned on legal data to produce results that mimic how a lawyer might respond. Even though Harvey and Lexis+ AI are powered by general-purpose LLMs like GPT-5, Gemini 2.5, and Claude Sonnet 4, fine-tuning can cause the results you get from Harvey and Lexis+ AI to differ noticeably from the outputs you get if you put the same prompts into platforms like ChatGPT, Gemini, and Claude.
Practice Tip:
Even though two tools may be powered by the same LLMs, consider whether those tools use fine-tuning that will produce different types of results.
Ask Yourself:
Which tool is best tuned to provide the output that will get me as far as possible in my project before I have to take over? If I need to draft a memo, will I get a better result from a tool like PC Chat that isn’t fine-tuned, or from a tool like Harvey or Lexis+ AI that is fine-tuned to mimic the reasoning and speech of a lawyer?
Instruction Tuning / RLHF
Instruction Tuning is a refinement process for pretrained LLMs to make them better at following user instructions in a conversational, chatbot-like manner. This is often combined with Reinforcement Learning from Human Feedback (RLHF), whereby human reviewers rank the quality of different model responses to a prompt. This feedback trains the model to produce outputs that are more helpful, harmless, and aligned with user preferences.
Example:
The reason you can have a natural, back-and-forth conversation with PC Chat or Microsoft Copilot is that the underlying LLMs have undergone extensive instruction tuning. This makes them act less like a simple text predictor and more like a helpful assistant.
Practice Tip:
Models that are well instruction-tuned are better at complex, multi-step prompts. You can give them a persona to adopt ("Act as a senior associate...") and they will generally adhere to it throughout the conversation.
Ask Yourself:
When a general prompt to Copilot gives a generic answer, how can I prompt it again using a specific persona and a more detailed, multistep instruction to leverage its instruction tuning and get a more useful, work product-ready response?
Model Feedback (or “User Feedback Loop”)
Model Feedback is the process of users reviewing and rating AI outputs—identifying what’s accurate, useful, or needs correction—so that the system and its future responses improve over time. Every time a user refines an AI draft, flags a hallucination, or marks a correct citation, that feedback contributes to more reliable performance.
Example:
After using Harvey, Copilot, or Lexis+ AI to draft a client update, a lawyer edits and rephrases unclear sections while a paralegal corrects citations. These changes can be logged or shared with the Library and then passed along to the vendor, thus strengthening the tool’s future outputs and fine-tuning.
Practice Tip:
Treat every interaction with an AI tool as a mini training opportunity—note inaccurate results, suggest improved prompts, and share examples of successful outputs for various use cases with the Library or Knowledge & Innovation team. Consistent feedback builds institutional knowledge and better aligns the AI with firm standards.
Ask Yourself:
How can I build a simple, consistent habit of providing feedback on AI outputs? What is the most efficient way to share my findings—both good and bad?
3. Grounding & Retrieval
Grounding
Grounding is the concept of anchoring an AI model's responses in specific, verifiable sources of truth to avoid hallucinations and improve the quality of responses. The goal is to prevent the model from relying solely on the generalized, and sometimes incorrect or incomplete, information it learned during training, and instead force it to base its answers on a provided, authoritative context. Grounding can happen in a variety of ways. The simplest is by uploading a document during your chat session and asking questions about the specific file. Alternatively, specialized tools are often connected to databases of information that can be searched before responding (see Retrieval-Augmented Generation). Still other tools have “web search” modes, which allow them to search the internet and parse the results before generating a response.
Example:
When you use Harvey Vault to ask questions about a set of discovery documents, the model is grounded in that specific dataset. Its answers are based on the content of those documents, not only on the LLM's general knowledge gained during training.
Practice Tip:
The most critical step in using AI responsibly is selecting a tool grounded in the right information for your task. For legal research, use a tool grounded in a legal database (like Lexis+ AI). For questions about firm knowledge, use a tool grounded in our internal systems. Using a general-purpose, ungrounded tool for a specialized task dramatically increases the risk of receiving a hallucinated or irrelevant answer.
Ask Yourself:
Before I start my task, am I defaulting to a general chatbot out of habit when a specialized, grounded tool would provide a more reliable and accurate answer?
Model Context Protocol (MCP)
Model Context Protocol (MCP) is a technique that preloads structured, scoped knowledge into an AI tool before the user types their prompt. It’s like giving the chatbot a preapproved cheat sheet it knows how to use. This differs from Retrieval-Augmented Generation (RAG), which functions as a search that looks for relevant documents after you submit your prompt. MCP gives the AI model what it should already know before it begins to answer, ensuring its responses are consistently framed within a specific context.
Example:
Consider an AI assistant for a project team. Using MCP, the project's key documents—the official project plan, team roles, and communication guidelines—are preloaded into the AI. A team member may ask, "What are my top priorities this week?" The AI then uses this preloaded context to provide a personalized, accurate answer based on the official plan, rather than a generic guess.
Practice Tip:
Recognize when a tool has this "superpower." If you know an AI has a preloaded guide, prompt it to use that guide explicitly. Your prompt may read, for example, "According to our team's communication plan, draft a status update for the client." This makes the AI a true extension of your team's established processes, ensuring consistency and saving you from having to provide the context every time.
Ask Yourself:
If I could give an AI assistant a "cheat sheet" for my most common tasks, what information would it include? How would having an AI that already knows our project goals, key client preferences, or standard procedures change the way I delegate tasks to it?
Retrieval-Augmented Generation (RAG)
RAG is a technique to improve LLM accuracy by connecting the LLM to external knowledge sources. When you ask a question, the RAG system first searches the database for relevant information. It then provides this retrieved information to the LLM as part of the prompt, instructing the LLM to use these sources to formulate its answer. This grounds the response in verifiable facts and significantly reduces (but does not eliminate) hallucinations.
Example:
Lexis+ AI is a prime example of RAG. When it answers a legal research query, it actively retrieves information from legal databases of primary and secondary law and use that content to generate the output, complete with citations.
Practice Tip:
For any task requiring high factual accuracy, always prioritize a tool that employs RAG. Even then, review the cited sources to confirm they support the generated text.
Ask Yourself:
After receiving an output from Lexis+ AI that uses RAG to provide a summary and five supporting case citations, what is the first verification step I must take, and why is it ethically required even though the tool is designed to be accurate?
4. Infrastructure & Deployment
Data Centers & Cloud GPUs/TPUs
The immense computational power required to train and run large AI models necessitates specialized hardware. Most of this work happens in large data centers operated by cloud providers (like Microsoft Azure, Google Cloud, and AWS). These data centers are equipped with tens to hundreds of thousands of high-performance servers with Graphics Processing Units (GPUs) or Tensor Processing Units (TPUs) that are optimized for AI calculations. This affects the speed, cost, and energy consumption of the AI services we use.
Example:
When you use cloud-based tools like PC Chat, Harvey or RelativityOne, your requests are sent to and processed in one of these secure data centers. The firm does not need to own or maintain this complex hardware; we access it as a service.
Practice Tip:
Be aware that using these powerful tools involves sending data to a third-party cloud provider. Different tools use different data center providers. Firm-approved AI tools have been vetted by our Information Governance team to ensure that prompts from those tools are being routed to secure data centers and that the necessary protections are in place to preserve the security of our firm’s and clients’ data in transit. Always adhere to the firm's data handling policies to ensure client confidentiality and data security are maintained.
Ask Yourself:
Before I upload sensitive client documents to a cloud-based AI tool, have I confirmed that the tool is approved for use on firm and client matters? (A full list of approved tools is available on the firm’s GenAI Connections page.)
Latency & Throughput
These two terms describe the performance of an AI system. Latency is the time delay between when you send a prompt and when you start receiving a response (user-visible speed). Throughput is the total number of requests the system can handle simultaneously (system capacity). Balancing low latency for a responsive user experience with high throughput to accommodate many users is a key challenge in deploying AI.
Example:
If you find that PC Chat is responding slowly during peak business hours, this could be a latency issue caused by high demand on the system (affecting its throughput). Different models and tools are optimized differently for these metrics.
Practice Tip:
For interactive tasks like drafting, low latency is crucial. For large, offline tasks like analyzing an entire Harvey Vault, throughput is more important. Plan ahead for projects that will require significant processing time and for which you have limited options (e.g., uploading thousands of files to Harvey Vault) so that you can allow the process to run in the background while you work on other tasks.
Ask Yourself:
If I am facing a tight deadline for an interactive drafting task and find PC Chat is responding slowly (high latency), should I wait, or would it be more efficient to switch to a different workflow that doesn't rely on immediate AI feedback? Or could I get drafting assistance from another GenAI tool that has lower latency, such as Copilot? How does this choice differ from a large, overnight analysis task in Harvey Vault, where overall processing speed (throughput) is more critical than initial response time?
5. Workflows & Agents
Tool Use / Function Calling
Tool use, or function calling, is a capability whereby an LLM can go beyond generating text and interact with external tools, websites, or other Application Programming Interfaces (APIs). The model can determine that to answer a user's request, it needs to perform an action, such as searching a database, running a calculation, or checking a calendar. It can then formulate a "call" to the appropriate tool and integrate the tool's output back into the response.
Example:
This is a core component of Agentic AI. When Lexis+ AI with Protégé searches the LexisNexis database to answer your question, the LLM is using a "search" tool. It recognizes the need for information outside of the LLM’s training data and calls the function to retrieve it.
Practice Tip:
As models gain greater capabilities to use more tools, prompts can become more goal-oriented. Instead of providing all the information, you can ask the model to "find the latest quarterly report for Company X and summarize the key financial risks," trusting it to use a search tool to find the report.
Ask Yourself:
My initial prompt to Lexis+ AI was, "Summarize the attached case." How can I rephrase this into a more goal-oriented prompt that leverages the AI's "search" tool capability to find relevant context or opposing arguments, rather than just analyzing the single document I provided?
Agentic AI (AI Agents)
Agentic AI refers to a system that can autonomously pursue a goal by breaking it down into steps, selecting and using tools, and iterating on its approach. Unlike a simple chatbot, an agent can perform a series of actions to accomplish a complex task with limited human intervention, moving from instruction-following to goal-oriented problem-solving.
Example:
Microsoft Copilot Researcher conducts extensive research using internet sources and/or your Microsoft 365 files and data to provide a response, and will self-check its preliminary responses until it’s “satisfied” that the answer is internally consistent. If new information surfaces during the first round of its research, it may choose to conduct additional research to further validate the information it uncovered and make sure it is consistent with its earlier findings. The processing time required for an agent to produce a response will vary (since the agent may revise its response several times before actually sharing the final output with the user) but could range from a few seconds to several minutes, and in some complex cases maybe an hour or more.
Practice Tip:
Agents are best for complex, goal-oriented tasks, not simple questions. Your role shifts from writing a perfect prompt to defining a clear objective and ensuring the agent has access to the right information. Since an agent performs multiple steps, be patient with its processing time—it's doing the work of a research assistant in the background.
Ask Yourself:
What complex research questions, which would normally take me hours of manual work, could I delegate to an agent? How does my planning change when I can set a high-level goal and let the AI figure out the steps, instead of me having to prompt it for each individual action?
Human-in-the-Loop (HITL)
Human-in-the-Loop is a critical process by which a human expert reviews, verifies, and corrects the output of an AI system. This combines the speed of AI with the judgment and ethical oversight of a professional. In the legal context, HITL is an ethical necessity to ensure the accuracy and quality of any AI-generated work product.
Example:
When using Solve Intelligence to help draft a patent application, a lawyer is always in the loop. The AI generates claims, but the lawyer must review every word, make corrections, and ultimately approve the final submission.
Practice Tip:
When reviewing AI-generated patent claims from Solve Intelligence, your focus should be on nuances the AI might miss: Is the claim scope broad enough to cover future product iterations but narrow enough to be patentable? Does the language introduce any ambiguities that could be exploited in litigation? The AI handles the structure; you provide the strategic legal judgment.
Ask Yourself:
Beyond checking for factual accuracy, what strategic questions should I focus on when reviewing AI-generated work product to ensure it equals or exceeds what I would otherwise create on my own?
6. Prompting & Controls
Prompt Engineering
Prompt engineering is the art and science of designing effective inputs (prompts) to guide a generative AI model toward a desired output. A well-crafted prompt is clear, specific, and provides sufficient context, constraints, and format parameters. It is the primary method users have at their disposal to control the model's behavior, and it significantly impacts the quality of the response.
Example:
A weak prompt like "summarize this document" is vague. A strong, engineered prompt is specific: "Summarize the attached 10-page project proposal for an executive audience. Focus on the key objectives, the proposed budget, and the timeline. The summary should be a three-paragraph memo." This provides the AI with a clear role, context, constraints, and format.
Practice Tip:
Think of prompting as delegating a task to a new team member. The more specific your instructions, the better the first draft will be. A simple formula for a strong prompt is Persona + Task + Context + Format. For example, an associate might prompt: "Act as a senior associate (Persona). Draft a two-paragraph summary (Task) of the attached plaintiff's deposition transcript (Context). The summary should be in a neutral tone and formatted for an internal case strategy memo (Format)."
Ask Yourself:
Before I write a prompt, have I clearly defined what a successful output looks like? Am I providing the AI with all the information it needs to succeed, or am I making it guess my intent? How can I break down my request into clearer, more specific instructions to get a better first draft?
System Prompts / Meta-Prompting
A system prompt is a high-level instruction that is automatically prepended to a user's prompt to set a persistent context, persona, or set of rules for the AI's behavior. It works behind the scenes to guide the model's tone, scope, and adherence to safety guidelines, ensuring more consistent and reliable outputs across many different user interactions.
Example:
A tool such as PC Chat has a system prompt that tells the underlying model something similar to the following: "You are a helpful AI assistant for the top-tier law firm Perkins Coie. Always answer questions in a professional manner and use the firm's branding and colors in your outputs. Cite your sources." You don't see this system prompt, but it guides every response.
Practice Tip:
While you can't change the system prompt, you can often override or reinforce parts of it in your own prompt. For example, explicitly asking for a specific format or tone can further guide a model that has already been given general instructions by a system prompt.
Ask Yourself:
Knowing an AI tool has its own "secret instructions," how can I write my prompts to better align with its built-in-persona (e.g., a helpful legal assistant) to get more relevant and useful responses?
Few-Shot / Patterned Prompts
A few-shot prompt is a technique that involves providing the model with a few examples of the task you want it to perform directly within the prompt. This helps the model understand the desired pattern, structure, and content of the output much more clearly than just an instruction alone. It's a powerful way to guide the model's output for structured or stylistic tasks.
Example:
To summarize case facts in PC Chat, you could provide examples: "Extract key dates and events from the text below. Format it as: [Date] - [Event]. For example: 'January 5, 2023 - Plaintiff filed the initial complaint.' Now, apply this pattern to the following document..."
Practice Tip:
Use few-shot prompting for any task that requires a consistent, repeatable format, such as creating timelines, filling out checklists, or extracting specific types of data from unstructured text.
Ask Yourself:
What is a recurring task I perform that requires a specific format? Could I save time by creating a "template" prompt that includes a few examples, which I can reuse whenever I need to perform that task?
7. Reliability, Risk & Governance
Hallucination
A hallucination is an AI-generated output that is factually incorrect, nonsensical, or entirely fabricated, yet presented with the same confidence as a factual statement. This is one of the most significant risks in using generative AI for legal work, as a hallucinated case citation or fact could be difficult to catch and have serious consequences.
Example:
If you ask a non-grounded public chatbot for cases supporting a legal doctrine, it might invent a realistic-looking case name and citation that doesn't actually exist. A RAG-based tool like Lexis+ AI is designed to prevent this by grounding responses in real legal sources. Even when using RAG-based sources, though, be careful to check that cited cases actually stand for the proposition they are being used to support. RAG-based tools reduce, but never fully eliminate, hallucination risk.
Practice Tip:
The single most important rule is to never trust, always verify. Any factual assertion, legal citation, or data point generated by an AI must be independently verified using a reliable source.
Ask Yourself:
My workflow includes a final 'never trust, always verify' step for every AI output. If a grounded tool like Lexis+ AI provides a case citation, what is my specific, step-by-step process for independently verifying that citation's accuracy, relevance, and current legal standing before including it in a client-facing document?
Bias & Fairness
Bias in AI refers to systematic skew in the model's outputs that can result in unfair or prejudicial outcomes. Since AI models learn from vast amounts of human-generated text, they can inherit and amplify existing societal biases related to race, gender, age, and other characteristics. This is a critical concern in legal applications like risk scoring, investigations, or human resources.
Example:
In early 2024, Google’s Gemini image generator was found to be producing historically inaccurate images. In an attempt to counteract biases favoring lighter skin tones in its training data, the model had been programmed to overcompensate by prioritizing diversity in its outputs. This led to it generating images of ethnically diverse people in historically inappropriate contexts, such as portraying Vikings or America's Founding Fathers as people of color. This demonstrated that even well-intentioned efforts to correct bias can lead to skewed and inaccurate results if not implemented carefully.
Practice Tip:
When using AI to analyze datasets (e.g., for hiring, internal investigations, or evidence review), be aware that the AI may have hidden biases or perform overcorrections. An AI might flag communication patterns as "problematic" based on biased training data, or it could overcorrect and fail to flag genuinely problematic behavior from underrepresented groups. The AI can be a powerful tool for identifying patterns, but human oversight is essential to interpret those patterns correctly and ensure a fair, legally defensible outcome. Also note that bias can often be addressed with precise prompting. If you ask a tool to provide pros and cons of a potential decision, or to assume a particular role or viewpoint before providing feedback, the precise instructions can help overcome latent bias in the model.
Ask Yourself:
If I am using an AI tool to review employee communications for an investigation, what steps can I take to ensure the AI's findings are not unfairly skewed by demographic factors? How can I validate that the patterns it identifies represent actual issues rather than reflections of bias in its training data?
Confidentiality & Data Handling
This refers to the policies and technical safeguards that govern how our data, and especially client data, is handled when using AI tools. It addresses what information is sent to the model, how it is encrypted in transit and at rest, where it is stored, and whether it is used to train future versions of the model. Maintaining confidentiality is our paramount ethical and contractual duty.
Example:
A tool like SimplyAgree is a prime example of why data handling is critical. When managing a deal closing, firms upload signature pages and final agreements containing personally identifiable information (PII) and confidential deal terms. The platform must ensure this data is encrypted, isolated, and not used for any purpose other than serving the client's specific matter, making its data handling policies a core feature of the product.
Practice Tip:
Before uploading documents to any cloud platform, you are responsible for understanding the data you are sharing and whether the platform is secure. Anonymize or redact sensitive PII that is not essential, and always confirm that using the platform complies with the firm’s guidelines as well as your client's specific outside counsel guidelines regarding third-party data hosting.
Ask Yourself:
Is the tool I’m using approved for client use and has the client authorized it as well? You can see a list of clients' AI use restrictions as well as a list of firm-approved AI tools on the firm’s GenAI Connections page.
8. Legal-Specific Applications
eDiscovery / Predictive Coding
Predictive coding (also known as technology-assisted review or TAR) is an ML application used in eDiscovery to classify large volumes of documents. An expert lawyer reviews the documents in a sample set, coding them for relevance or privilege. The system learns from this and applies the logic to the entire document set, allowing teams to focus on the most important documents first.
Example:
Both DISCO Cecilia and RelativityOne aiR feature predictive coding. You can "train" the system on what constitutes a responsive document, and it will then help you quickly identify similar documents across millions of files.
Practice Tip:
The success of predictive coding relies on a sound workflow. Ensure you have a knowledgeable expert creating the initial training set, and regularly sample the AI's predictions to validate its accuracy.
Ask Yourself:
When might predictive coding help on my matters? Who would need to sign off on this approach? If overseeing a predictive coding review, what is my workflow for sampling the AI's classifications? How do I use that sampling data to decide whether the model is accurate enough to rely on or if it needs further training from our expert reviewer?
Semantic Search
Semantic Search is an advanced search technology that goes beyond the literal matching of keywords to understand the intent, context, and conceptual meaning behind a user's query. Instead of just finding documents that contain the exact words you typed, it finds results that are conceptually related and most likely to answer your question, even if they don't use the same terminology.
Example:
Instead of searching a witness deposition in Copilot for the exact phrase "remembered seeing the red car," a lawyer can use semantic search to ask a conceptual question such as, "What did the witness say about the vehicle involved in the collision?" The AI understands the meaning and can find relevant testimony where the witness describes "a crimson sedan," "the car that ran the light," or "that maroon vehicle," even if the exact search terms are never used.
Practice Tip:
When you get poor results from a keyword search, switch your thinking from words to concepts. Rephrase your query as a direct question about the information you're seeking. Trust that the AI can understand the meaning behind your question and find relevant information, even if the phrasing is different. This is especially powerful when reviewing documents you didn't draft yourself.
Ask Yourself:
What is a complex topic I frequently research? How can I rephrase my typical keyword searches into conceptual questions to uncover more relevant information and different perspectives?
Legal Entity Recognition (LER)
Legal Entity Recognition is an AI technique that automatically identifies and classifies key elements within legal text—such as party names, statutes, case citations, monetary amounts, or dates—so that they can be searched, linked, or analyzed efficiently.
Example:
When an AI tool like Harvey Vault or Kira reviews a contract, it can automatically highlight “Effective Date,” “Indemnifying Party,” or “Governing Law,” turning unstructured text into structured data that lawyers and paralegals can filter, compare, or export to a database.
Practice Tip:
Use LER to quickly create summaries or checklists from long documents. For example, after an LER tool processes a set of contracts, you can instantly filter for all agreements governed by California law or those with an expiration date in the next quarter. This turns a manual review task into a quick data query.
Ask Yourself:
What repetitive information-finding tasks in my daily workflow could be automated if key pieces of data were automatically identified and tagged for me?
9. Capability Horizon
Artificial General Intelligence (AGI)
Artificial General Intelligence is a hypothetical future form of AI that possesses the ability to understand, learn, and apply its intelligence to solve any problem that a human being can. Unlike the narrow AI we have today, which is designed for specific tasks, an AGI would have broad, flexible, and adaptive cognitive abilities comparable or superior to humans. It does not currently exist.
Example:
A lawyer with access to AGI would not need to use multiple specialized AI tools for a new case. Instead, they could give a hypothetical AGI a single high-level goal, such as: "Assess a potential trade secret theft case against our client and outline a strategy." The AGI would autonomously understand the legal concepts, conduct a multifaceted investigation by analyzing contracts and emails, research public information on the competitor, identify relevant case law, and then reason and synthesize all its findings. It would then produce a single comprehensive strategy memo outlining the case's strengths, weaknesses, and a proposed plan of action, fully sourced with accurate citations. This ability to independently plan, reason, and execute a complex, multidomain goal is the core difference between today's narrow AI and a true AGI.
Practice Tip:
Treat today's powerful generative AI tools like Harvey and Lexis+ AI as specialized assistants, not autonomous strategists. Your role is to break down a complex legal problem into a series of discrete tasks (e.g., legal research, document analysis, draft creation) and assign each task to the appropriate tool. Unlike a hypothetical AGI, current AI requires you to be the human general contractor who orchestrates the entire project.
Ask Yourself:
When I'm faced with a complex legal task like assessing a new case, how do I deconstruct that high-level goal into a series of specific, concrete prompts that can be executed by different specialized AI tools, such as Lexis+ AI for research and Harvey for analysis, while ensuring that I remain the strategic "human in the loop" throughout the process?