Legal & AI Definitions (Glossary)

Law

A system of enforceable rules created by recognized authority to regulate behavior, allocate power, and resolve disputes.

Not morality. Not fairness. Enforceability is the key property.

Rule of Law

The principle that decisions are governed by pre-existing rules, not individual discretion.

Without this, law becomes power without predictability.

Source of Law

The recognized origin of legal rules (statutes, regulations, case law, custom).

AI must know which sources count and which don't.

Jurisdiction

The legal authority of a system to govern a situation. Always defined by place + subject matter + time. Our Jurisdiction Finder tool applies exactly this test.

Wrong jurisdiction = irrelevant answer.

Hierarchy of Norms

The ranking of legal rules by authority (constitution > statutes > regulations > contracts).

Lower rules cannot override higher ones. Ever.

Procedure

The formal steps required to activate legal rights.

Procedure is not decoration. It is the gatekeeper of substance.

Procedural Nullification

The extinction of a legal claim or right due to failure to follow required procedure, regardless of substantive merit.

Miss a deadline, file wrong venue, use wrong form: your claim dies, not weakens.

Temporal Validity of Law

The principle that a legal rule or interpretation is only correct at a specific point in time.

Legal answers without a date are incomplete by definition.

Liability

Legal responsibility for damage or wrongdoing.

Does not require intent. Only causation and attribution.

Risk

The probability and impact of a legal consequence - the core subject of Legal Risk & Strategy.

Lawyers don't eliminate risk. They price and position it.

Privilege

A legal protection preventing disclosure of certain communications, covered in full in Privilege & Confidentiality with AI.

Fragile. Easy to lose. AI misuse can waive it.

Admissibility

The legal criteria determining whether evidence may be considered by an authority.

Truth without admissibility has zero legal value.

Regulatory Discretion

The authority of regulators to decide how, when, and whether to enforce rules.

Explains why identical facts can produce different outcomes.

Artificial Intelligence (AI)

Systems that perform tasks requiring pattern recognition, inference, or decision support using statistical models.

Not intelligence. Not consciousness. Automation of cognition-like tasks.

Large Language Model (LLM)

A probabilistic model trained to predict the next token in a sequence based on massive text data.

Optimized for plausibility, not truth.

Hallucination

A confident but false output produced when a model fills gaps. See Correctness Over Confidence for how to test for it.

In law, hallucinations are not amusing. They are liability events.

Prompt

A structured instruction defining task, context, constraints, and output expectations for an AI system.

Bad prompt = uncontrolled assumptions.

Context Window

The amount of information an AI can consider at once - the whole subject of Context Engineering for Lawyers.

If key facts fall outside it, outputs degrade silently.

Fine-tuning

Training a model further on specific data to shape behavior.

Useful, but not a substitute for jurisdiction control or reasoning.

Retrieval-Augmented Generation (RAG)

A method where AI retrieves external sources before generating output.

Critical for law. Without retrieval, AI guesses.

Legal AI

An AI system designed to operate under legal constraints: jurisdiction, sources, hierarchy of norms, and accountability.

If it can't explain why an answer is valid, it's not legal AI.

Legal Reasoning

Applying legal rules to facts under procedural and jurisdictional constraints to reach a defensible conclusion.

Not explanation. Not summary. Application under constraint.

Legal Coherence

The internal consistency of a legal system as it exists, including tolerated contradictions and exceptions.

Not the same as logical or mathematical consistency.

Interpretive Latitude

The bounded freedom authorities have when applying legal rules to facts.

Creates outcome ranges rather than binary answers.

Traceability

The ability to link an AI output to: assumptions, sources, versions, and reasoning steps.

If you can't trace it, you can't defend it.

Auditability

The ability to review how an output was produced after the fact.

Mandatory for regulated environments.

Human-in-the-Loop

A design where a human reviews, validates, or overrides AI output before action.

Not optional in law. Ethically required.

Accountability Chain

The documented path from AI output to human decision-maker.

Courts follow this chain. Break it and responsibility still lands on you.

Scope

The explicit boundaries of what the AI is allowed to consider or produce.

Undefined scope = hallucination invitation.

Assumption

A condition treated as true for the purpose of reasoning.

Good AI systems surface assumptions instead of hiding them.

Iterative Drafting

A multi-pass workflow where outputs are refined through review and correction.

One-shot drafting is amateur behavior.

SLA-Style Prompting

Prompting that specifies limits, structure, and performance expectations.

Reduces verbosity. Increases reliability.

Preventive Law

Legal work done to avoid disputes or sanctions before they occur.

Cheaper. Faster. Smarter.

Reactive Law

Legal work responding to an existing conflict.

Always more expensive.

Litigation Leverage

Using legal process as pressure, not necessarily to reach judgment.

Most cases settle because of leverage, not justice.

Settlement

A negotiated resolution avoiding formal adjudication, usually documented in a Settlement Agreement.

The most common legal outcome.

Professional Responsibility

The non-transferable duty of a legal professional to act competently, confidentially, and ethically.

AI does not dilute this. It intensifies it.

Confidentiality

The obligation to protect client or sensitive information, typically enforced through a confidentiality clause.

AI systems do not create exceptions.

Waiver

The loss of a legal right through action or omission - the exact failure mode a waiver clause is written to prevent.

Careless AI usage can trigger it instantly.

Liability Concentration

The phenomenon where automation increases speed and scale, making errors more impactful and responsibility more focused on decision-makers.

Legal AI Twin

An AI system designed to think with a specific lawyer or team, reflecting their jurisdiction, style, risk tolerance, and accountability.

Not autonomous. Augmented judgment.

Co-Reasoning

A workflow where human and AI collaboratively structure legal thinking.

The human decides. The AI accelerates.

System of Record

A unified environment where documents, decisions, reasoning, and audit trails live together.

Fragmentation kills accountability.

Token

The unit a model actually reads and writes: a chunk of text, usually a few characters and shorter than a word. Your documents are cut into tokens before the model sees any of them, and every limit, price and window you meet is counted in tokens, not pages. Context Engineering for Lawyers does the conversion in front of you.

A page count tells you nothing about whether a bundle fits. Convert first, and remember the same clause in Arabic costs more tokens than it does in English.

Temperature

The setting that decides how much randomness the model is allowed when it picks each next token. Low temperature takes the most likely continuation almost every time; higher temperature spreads the choice across less likely ones. It is why asking for the same indemnity clause twice gives you two different drafts, and why neither draft is more correct than the other. The sampler is on show in Legal AI Fundamentals.

Need the same output twice? Lower the temperature and save the prompt. Repeatable is not the same as right: a low temperature reproduces a wrong clause perfectly.

Knowledge Cutoff

The date a model's training data stops. Everything after it is absent from the weights, and nothing in the model marks the edge. A statute amended after the cutoff, or a judgment overturned after it, leaves the model stating the superseded position in fluent, confident prose.

The model does not know what it does not know. Anything with a date on it needs retrieval, or a human holding a current source.

Context Rot

The loss of reliability as the context fills up, even while everything still fits inside the window. What thins first is the detail in the middle of a long input, which on a litigation bundle is the schedules, the side letters and the amendments. Measured against your own documents in Context Engineering for Lawyers.

Fitting is not reading. Put the clause you are actually asking about at the end, or cut the bundle down until it is the only thing in there.

Harness

Everything wrapped around the model: the standing instructions, the tools it may call, the permissions on those tools, what it remembers, and whether it is allowed to keep going. Two products built on the same model behave nothing alike, because the harness is different. That distinction carries all of Asking a Question vs Delegating a Task.

Most of what impresses you in a demo is harness, not model. That is the part the vendor actually built, so that is the part to interrogate.

Agent

The same model, given tools it may call and permission to keep going. It decides a step, takes it, reads the result and decides again, stopping when the goal is met, when it is stuck, or when it needs you. Nobody scripts the steps in advance. "Summarise this clause" is a chat; "turn this bundle into a chronology and flag anything that contradicts the witness statement" is an agent.

Delegation moves the steps, not the duty. Ask for the trail: what it opened, what it changed, what it was unsure about. Read that before you rely on the output.

Skill

Your method written down where the assistant can load it: a name, one line saying when to use it, then the steps, the checks and the shape of the answer. The assistant holds the name and that one line at all times and opens the rest only when a request matches it, which is why installing thirty skills does not cost thirty skills' worth of context. Skills, MCP and Plugins takes the mechanism apart.

A skill nobody opened did nothing. If the one line does not match how your people actually ask, the firm's position never gets applied and nothing tells you it was skipped.

MCP (Model Context Protocol)

The open standard for joining an assistant to a system it does not own. Without a shared protocol, every assistant needs a bespoke link to every system: three assistants and four legal systems need twelve of them, each one somebody's job to keep working. With it, each system needs one.

The letters almost never appear in the product. You are usually buying this under the word connector, or app.

Connector

A live link that lets the assistant read from and write to a system you already use: the document management system, the matter database, a filing portal, a contract repository. Without one, the workflow is download, copy, paste, which is where most confidentiality mistakes happen.

A connector is a permission, not a feature. Ask what it can write and send, not only what it can read.

Plugin

The bundle a team installs as one thing: a skill for each part of a task, the connectors those steps need, a command that starts the work, and the defaults. It is how a new joiner gets the firm's method on day one rather than in year three.

You are installing somebody else's instructions and running them with your own permissions. Read what is in the bundle before it becomes the way your firm works.

Prompt Injection

An instruction hidden inside something the model reads: a document, an email, a web page, a file name, white text in a PDF. The model has no way to tell your instruction apart from one it found in the material, so an exhibit filed by the other side can steer the tool that is reviewing it. Handled in Privilege & Confidentiality with AI.

The risk grows the moment an agent gets tools. Treat nothing it reads as trusted input, and contain it with permissions, not with a better prompt.

Benchmark

A scored run of a fixed set of tasks, used to compare models or products against a reference. The score means nothing until you know who wrote the tasks, who graded them and what the reference was, which is the whole argument of Practitioner-Led Benchmarks.

If the vendor wrote the tasks, graded them and published the result, you are reading marketing. Ask whether your jurisdiction and your document types are in the set.

Open Source

Published so that anyone can inspect, run and modify it. For a model, what gets published is normally the weights: enough to run it on your own hardware, which is not the same as publishing what it was trained on. A product can also be open source around a closed model, which is a different claim entirely.

Ask what exactly is open: the weights, the code around them, or the licence on the output. The three answers have different consequences for confidentiality and for lock-in.