Adversarial Testing
The practice of intentionally probing legal AI tools with edge cases, contradictory prompts, or malicious inputs to identify failure points, hallucinations, or security vulnerabilities before deployment.
Agentic AI
AI systems capable of carrying out multi-step tasks with some autonomy, such as researching, drafting, and revising a document in sequence without step-by-step human instruction.
AI Governance
An organisation's internal policies, approval processes, and oversight structures for how AI tools may be used, by whom, and with what safeguards.
AI Literacy
A practical understanding of how AI tools work, their limitations, and how to use them appropriately.
AI Slop
Low-quality, generic content produced by AI with little human oversight, often mass-produced and lacking accuracy, originality, or editorial value.
AI Washing
Marketing a product or service as AI-powered when its actual use of AI is minimal or exaggerated.
Algorithm Bias
Systematic and repeatable errors in an AI model's outputs that create unfair outcomes or skew legal analysis due to unrepresentative training data or flawed feature weighting.
Application Programming Interface (API)
A defined way for one piece of software to send requests to and receive responses from another, commonly how organisations connect their own systems to an AI provider's model.
Chat Model
An AI model designed to respond to a prompt in a single, near-instantaneous pass, as distinct from a reasoning model that works through intermediate steps.
Cite-Checking
The practice of verifying that citations produced by an AI tool actually exist and accurately reflect the source they are attributed to.
Closed Model
An AI model whose internal workings are not publicly available, typically accessed only through a company's own paid API or product.
Confidence Score
A numerical metric generated by an AI model indicating its statistical certainty in a given output, which can be misleading in legal contexts where fluent, confident delivery frequently masks factual fabrication.
Context Window
The amount of text an AI model can process at once. A larger context window allows longer documents or more conversation history to be considered in a single request.
Copilot
An AI assistant embedded within existing software, such as a word processor or practice management system, designed to help with tasks like drafting or summarising within that tool.
Custom Instructions
Standing preferences a user sets within an AI tool, such as preferred tone or format, which are applied automatically across chat sessions without being repeated in every prompt.
Data Residency
The geographic location in which an AI provider stores or processes a user's data, relevant to firms with confidentiality or jurisdictional obligations.
Digital Twin
A continuously updated digital model of a real-world entity — a company, contract portfolio, or piece of infrastructure — that mirrors its current state and can be used to test scenarios or answer questions without touching the underlying system itself.
E-Discovery
The process of identifying, collecting, and reviewing electronic documents for use as evidence in litigation, increasingly assisted by AI-based document review tools.
Frontier Model
One of the most advanced AI models available at a given time, developed by a small number of large AI companies. Most AI-branded products are built on top of a frontier model rather than a model built from scratch.
Generative AI
AI systems that create new content, such as text, images, or code, rather than simply analysing or classifying existing material. The category that includes tools like ChatGPT and Claude.
Grounding
Tying an AI system's output to specific source material, such as a document set or knowledge base, rather than letting it answer purely from general training data.
Hallucination
When an AI system generates information, such as a fact, quote, or citation, that sounds plausible but is factually incorrect or entirely invented.
Human In The Loop
A workflow design where a person reviews, checks, or approves AI-generated output before it is used or relied upon, rather than the AI acting unsupervised.
Jailbreaking
Deliberately phrasing or structuring a prompt to get an AI model to bypass its built-in restrictions and produce output it would normally refuse to generate.
Knowledge Cutoff
The date beyond which an AI model has no training data, meaning it has no awareness of events or information after that point.
Large Language Model (LLM)
An AI model trained on very large volumes of text to predict and generate language. The technology underlying tools such as ChatGPT, Claude, and Gemini.
Model Context Protocol (MCP)
A standard that allows AI models to connect to external tools and data sources, such as a document management system, in a consistent way.
Model Drift
The gradual degradation of an AI system's performance and accuracy over time due to changes in underlying legal standards, statutory updates, or shifting patterns in user input.
Multimodal
Describes an AI model that can process or generate more than one type of content, such as text, images, or audio, rather than text alone.
Open-Source Model
An AI model whose underlying code or weights are published for anyone to use, inspect, or modify, as distinct from closed, proprietary models accessed only via a paid service.
Parametric Memory
The internal knowledge stored within an AI model's weighted neural network during its initial training phase, distinct from external data retrieved dynamically via search or document repositories.
Persistent Memory
A setting in some AI tools that allows the model to retain and reuse information from earlier chat sessions, rather than treating each session as fully separate.
Prompt Engineering
The practice of crafting instructions given to an AI model to produce more accurate, relevant, or consistently formatted output.
Reasoning Model
An AI model that breaks a prompt and its response into multiple intermediate steps before producing a final answer, rather than generating a response in one pass.
Retrieval-Augmented Generation (RAG)
A method where an AI model retrieves relevant documents from a database before generating an answer, so responses are grounded in specific source material rather than general training data.
Shadow AI
The unauthorised use of consumer-grade AI applications or unapproved tools by fee-earners or staff without the knowledge, vetting, or oversight of firm IT and risk management teams.
Synthetic Data
Artificially generated data used to train or test AI systems, sometimes used to avoid exposing real personal or confidential information during development.
System Prompt
A set of instructions built into an AI tool by its provider, sitting between the user's prompt and the model, that shapes or restricts how the model can respond before the output reaches the user.
Technology-Assisted Review (TAR)
A document review method where AI ranks or classifies documents by likely relevance, based on decisions made by human reviewers on a sample set. Also known as predictive coding.
Temperature
A setting that controls how random or deterministic an AI model's output is, with lower values producing more consistent, predictable responses.
Token
A unit of text, roughly a word or part of a word, used by AI models to process language and by AI providers to measure usage and cost.
Training Data
The text, images, or other material used to teach an AI model to generate language or make predictions. Its quality and scope shape what the model can and cannot do well.
Vibe Coding
Writing software by describing what you want in plain language and letting an AI tool generate the code, with little or no manual coding by the person directing it.
Zero Data Retention
A contractual or technical arrangement in which an AI provider does not store or reuse data submitted to it beyond processing the request.
