Artificial-intelligence headlines move faster than a busy docket, and the jargon can feel just as dense as statutory cross-references. Terms like "RAG" or "vectorization" pop up in vendor pitches, bar-association CLEs, and even court opinions, yet few explanations are written with practicing lawyers in mind.
That's why we've compiled this AI Glossary for Attorneys: a plain-English cheat sheet that translates the most common buzzwords into concepts you can immediately relate to client work, discovery, and courtroom strategy.
Think of it as the quick-reference section you'd keep tabbed in a treatise—ready when a partner asks whether a tool has "guardrails," or when opposing counsel touts their "agentic AI" platform.
Each entry starts with a straightforward definition, highlights the legal-practice angle, and flags any pitfalls (like hallucinations or bias) you'll want to watch for. Skim it now, bookmark it for later, and feel free to share it with colleagues who'd rather read headnotes than hash maps.
AI systems that can execute multi-step tasks with some autonomy—for instance, drafting a document based on your guidelines and flagging areas needing review. Current reality: Most systems still require significant human oversight despite marketing claims.
Amazon Web Services' managed platform that lets you plug different large-language-model providers (Anthropic, Meta, Amazon Titan, etc.) into your own app without hosting the heavy AI infrastructure yourself—similar to leasing space in a secure data center rather than building one.
The tendency of AI systems to produce discriminatory results based on skewed training data. Critical for compliance with anti-discrimination laws, especially in hiring, criminal justice applications, or client screening.
The AI's step-by-step explanation of how it reached a conclusion, not just the final answer. Essential when you need to defend the AI's analytical process in court or to clients. Always ask vendors if their tools provide this.
Breaking a long document (e-mail thread, contract, transcript) into smaller segments so the model can process it without hitting its size limits, and so that each chunk stays on point for better answers. Like organizing exhibits by topic rather than presenting one massive binder.
Metadata or inline links that show where the model found each fact. Non-negotiable for legal work—always insist on citation capabilities when evaluating AI tools. Without citations, you're flying blind on verification.
The maximum "working memory" of a model—the total number of tokens it can consider at once. Affects how much of a case file you can analyze in one session. If you exceed the window, older text falls out of scope.
How AI systems handle personally identifiable information and privileged client data. Essential for maintaining attorney-client privilege and complying with GDPR or state privacy laws. Always verify vendor practices before uploading sensitive documents.
A mathematical fingerprint (a long list of numbers) that captures the meaning of a piece of text. Two passages about "tolling a statute of limitations" will have embeddings that sit close together in vector space even if the wording differs. Powers semantic search in modern e-discovery.
Adapting a pre-trained model using hundreds or thousands of your firm's specific examples—e.g., discovery responses—to match your style and expertise. Requires technical expertise and significant investment but potentially valuable for large firms with consistent processes.
The most advanced AI systems currently available, subject to increasing regulatory scrutiny due to their powerful capabilities and potential for misuse. These models face special oversight requirements including third-party testing and disclosure obligations.
Policy or technical limits placed on a model—like blocking certain topics, requiring citations, or preventing privilege waiver—to reduce risk. Always ask vendors about their guardrails and testing procedures. Think firm policies that keep junior associates from sending unvetted client emails.
When a model states something plausible but factually wrong—like citing a non-existent case or misquoting a statute. The #1 risk in legal AI. Always verify outputs before relying on them, treating AI like an overeager junior associate.
The act of running the model on your prompt to get an output. Unlike training (expensive and infrequent), inference happens every time you ask a question. You typically pay for inference, not training.
The pieces—usually words or word-fragments—that get counted for billing and context limits. Understanding token costs helps predict AI tool expenses. Both your prompt (input) and the AI's response (output) consume tokens.
The delay between sending a prompt and receiving an answer. Critical for live applications like courtroom support or client chat. High latency kills real-time workflows.
AI systems trained on vast text databases to predict what comes next in a sequence. More flexible than traditional ML but require careful oversight. Powers most legal AI tools from chatbots to contract review platforms.
Documentation describing what a model was trained on, intended uses, limitations, and ethical considerations. Essential for compliance checklists when evaluating AI vendors. If a vendor can't provide this, proceed with caution.
The underlying AI program (with its parameters or "weights") that turns inputs into outputs—the "brain" behind any AI tool you're evaluating. Different models have different strengths, biases, and costs.
The ability to handle text, images, audio, and video in a single model. Useful for tools that read scanned contracts and answer questions about specific clauses, or analyze video depositions alongside transcripts.
A database technique for quickly finding embeddings that sit closest to your query in vector space—used to surface the most relevant documents before sending them to the model. Powers "find similar documents" features in e-discovery.
Crafting questions to get reliable, useful responses from AI. Specific, detailed prompts yield better results than vague requests. Like asking deposition questions that can't be dodged—precision matters.
Security vulnerabilities where malicious actors manipulate AI systems to bypass guardrails or extract sensitive information. Important when using AI tools with client data or in adversarial legal contexts.
A two-step process: first retrieve relevant documents from your database, then generate an answer citing those specific docs. Enables AI to work with your private documents and data that weren't in its original training. Dramatically reduces hallucinations and makes AI responses more legally defensible. Look for this in any tool handling case law or discovery. In Casefleet, this is RAG is used for Document Intelligence processing.
Caps on how many requests you can send per minute or day—affects both pricing and workflow planning. Like court filing limits, these prevent system overload but can bottleneck urgent projects.
Controls how "creative" vs. consistent the model's outputs are (typically on a 0-1 scale, where 0 = most consistent). Low temperature for contract clauses that need uniformity; higher for brainstorming litigation strategies. For most legal work, keep it low.
Letting the model interact with external systems—databases, calendars, e-discovery platforms—to fetch data or perform actions. Critical for AI tools that need to work with your existing tech stack.
Classic machine learning: narrow, specialized systems requiring structured data. LLMs: broad, adaptable systems working with natural language. LLMs are more flexible but need more oversight—like the difference between a specialist expert witness and a generalist associate.
Turning text into numerical embeddings so computers can measure semantic similarity. Enables searching for all documents about "breach" even if they use terms like "default" or "violation." Foundation of modern legal search.
The millions or billions of numeric parameters inside a model that encode its "learned" patterns. Can contain biases from training data, which is why human oversight remains essential. Think of them as the model's accumulated experience—valuable but not infallible.
Now that you understand the technology, see how Casefleet's AI-powered litigation management tools can transform your practice. Our platform combines RAG-enabled document review, reliable citations, and enterprise-grade security—all designed specifically for legal professionals. No jargon, no hallucinations, just powerful tools that help you win cases.
Start your free trial today and discover why thousands of attorneys trust Casefleet to harness AI safely and effectively.