Why AI Terminology Matters for Legal Professionals

Last reviewed: July 9, 2026. This glossary belongs in the site’s glossary category and is maintained as a source-cited reference, not a vendor-style list of fashionable AI words. The point is narrower and more practical: when a lawyer sees terms such as “RAG,” “fine-tuning,” “agentic AI,” or “hallucination” in an ethics opinion, procurement questionnaire, court order, or workflow memo, the vocabulary is already touching duties of competence, confidentiality, communication, and supervision.

ABA Formal Opinion 512, issued in July 2024, frames generative AI use through existing professional responsibility rules rather than treating it as a novelty exception. Its practical message is familiar: lawyers must understand enough about the technology they use to evaluate its risks, protect client information, communicate when necessary, verify work product, and supervise people and systems involved in legal work.[1] Model Rule 1.1 Comment 8 has been saying the same thing in broader form for years: competence includes keeping abreast of changes in relevant technology.

That duty is no longer confined to a few legal-innovation committees. A vendor-sponsored 8am report, discussed in ABA Law Practice Magazine, reported that 69% of surveyed legal professionals use generative AI for work, based on more than 1,300 respondents.[2] The sponsorship matters; this is adoption context, not independent proof that the profession has mastered the tools. Still, it explains why AI vocabulary in the legal profession now shows up in ordinary practice management conversations.

How to Use This Glossary

The entries are arranged by legal usefulness, not by computer-science taxonomy. A technically elegant definition is not enough here. A useful definition should help a legal professional decide whether an output needs source verification, whether client data may be processed by a third party, whether a human must review a workflow step, or whether a vendor’s capability claim is narrower than it sounds.

Foundational AI Categories

Artificial Intelligence

Artificial intelligence is the broad umbrella term for computer systems that perform tasks associated with human cognitive work, such as classifying documents, finding patterns, ranking likely outcomes, summarizing text, or generating drafts. In legal practice, “AI” by itself is usually too vague to support a procurement decision or an ethics analysis.

The legal implication is simple: never stop at the label. A contract-analysis tool that extracts indemnity clauses, a litigation analytics platform that estimates motion outcomes from historical data, and a chatbot that drafts a memo may all be called AI, but they create different verification, confidentiality, and supervision questions.

Machine Learning

Machine learning is a subset of AI in which a system improves its performance on a task by learning patterns from data rather than by following only hand-coded rules. In legal settings, machine learning appears in document classification, predictive coding, legal analytics, contract review, and risk scoring.

The professional responsibility issue is not whether the system “thinks.” It does not need to think to affect a legal judgment. If a model ranks documents as likely responsive, marks clauses as unusual, or predicts likely case outcomes, a lawyer still needs to understand the task, the data source, the likely error pattern, and the appropriate level of human review.

Generative AI

Generative AI refers to systems that produce new content, such as text, images, code, summaries, emails, deposition outlines, contract language, or legal research drafts. A large language model predicts likely sequences of words based on patterns learned from training data and, in some systems, additional retrieved material.

For legal professionals, generative AI is useful precisely where it is also dangerous: it can create plausible, fluent work product faster than a human reviewer can easily inspect it. ABA Formal Opinion 512 specifically addresses generative AI and connects its use to duties including competence, confidentiality, communication, candor, fees, and supervision.[1]

Generative AI vs. Extractive or Analytical AI

Generative AI creates new text. Extractive AI pulls or labels existing text. Analytical AI identifies patterns, rankings, or probabilities from data. The distinction matters because the review burden changes.

• An extractive contract tool may identify every governing-law clause in a set of agreements. The review question is whether it missed clauses or mislabeled them.
• An analytical litigation tool may show how often a judge granted a certain motion in historical cases. The review question is whether the data set and comparison are appropriate.
• A generative AI tool may draft an argument or summarize a case. The review question includes whether the cited authority exists, whether the propositions are accurate, and whether the generated text fits the client’s facts.

Calling all three “AI” hides the most important part of the analysis: what the system is doing to the material and what kind of human verification must follow.

Legal Workflow Terms

Technology-Assisted Review and Predictive Coding

Technology-assisted review, often called TAR, uses software to help identify, prioritize, or classify documents in discovery. Predictive coding is a TAR method in which a model learns from human-coded examples and applies those patterns to larger document sets.

This is not new merely because current generative AI is new. Research by Maura R. Grossman and Gordon V. Cormack helped establish the evidence base for technology-assisted review, and Da Silva Moore v. Publicis Groupe in 2012 is widely cited as an early federal decision approving computer-assisted review in discovery.[5] The legal lesson from TAR remains useful: validation, sampling, process documentation, and defensibility matter more than whether the software is described in impressive terms.

Legal Analytics

Legal analytics uses structured or semi-structured legal data to identify patterns, such as motion outcomes, judge behavior, opposing counsel history, settlement ranges, or timing trends. It is often predictive in the everyday sense, but a prediction based on historical data is not a legal conclusion.

The practice implication is one of communication and judgment. A lawyer using analytics should be able to explain what the metric measures, what data it includes, what it leaves out, and why it does or does not affect advice to the client. A percentage derived from historical cases may inform strategy; it does not relieve counsel of analyzing facts, law, venue, procedural posture, and client objectives.

Contract Analysis

Contract analysis tools review agreements to extract clauses, compare language, flag deviations from playbooks, or summarize obligations. Some tools are largely extractive and classification-based; others add generative summaries or suggested revisions.

That distinction should shape review. If a tool extracts change-of-control clauses, the main concern may be missing a clause or placing it in the wrong category. If it drafts revised language, counsel must review the legal effect of the new wording, not merely whether it reads smoothly.

General-Purpose AI vs. Purpose-Built Legal AI

A general-purpose AI tool is designed for broad tasks across many domains. A purpose-built legal AI tool is designed for legal workflows, often with legal databases, matter-management features, permission controls, audit logs, or workflow constraints. The label “legal AI” still needs testing; it is not a professional responsibility safe harbor.

For tool selection, compare the tool’s design to the actual use case. A brainstorming prompt, a privileged client memo, an e-discovery classification workflow, and a filed brief do not belong in the same risk bucket. For more practical comparisons, see the site’s AI legal software comparison guide and AI legal research tools comparison.

RAG, Fine-Tuning, and Grounding

Retrieval-Augmented Generation

Retrieval-augmented generation, or RAG, is an architecture in which a system retrieves relevant source material and uses it to generate an answer. In a legal research setting, that source material might include cases, statutes, regulations, secondary sources, internal knowledge documents, or matter files.

The legal value of RAG is not that it magically prevents error. Its value is that it can make an answer more inspectable. If the tool shows the authorities it retrieved, a lawyer can check whether the cited sources exist, whether they support the proposition, and whether the system ignored better authority. RAG reduces one class of risk only when retrieval quality, source coverage, citation display, and human verification are adequate.

Fine-Tuning

Fine-tuning means further training a model on a more specific data set so it performs better for a domain, style, task, or organization. A model might be fine-tuned to produce contract summaries in a preferred format or to recognize certain document types more reliably.

Fine-tuning is not the same thing as source grounding. A fine-tuned model may sound more legally fluent or follow a firm’s preferred format, but that does not mean the answer is tied to current authority in the moment of use. If a vendor says a model is “trained on legal data,” the next question is whether the tool retrieves and displays sources at inference time, how current those sources are, and whether the user can verify them.

Grounding

Grounding means connecting an AI output to identified source material or a defined factual context. In legal work, grounding may come from retrieved case law, uploaded matter documents, a contract repository, a regulation database, or a user-provided record.

The word should be handled carefully. A grounded answer can still be wrong. The system may retrieve the wrong case, misread a holding, overstate a quotation, omit contrary authority, or combine correct fragments into an unsupported conclusion. Grounding is a reason to verify more efficiently, not a reason to stop verifying.

Hallucination and Source Verification

Hallucination

A hallucination is an AI output that presents false or unsupported information as if it were true. In legal writing, the most visible form is a fabricated citation, but the risk is broader: a real case can be cited for a proposition it does not support, a quotation can be altered, or a rule can be stated without an essential limitation.

The term can sound almost whimsical. Court sanctions are not. In February 2025, a Wyoming federal court fined Morgan & Morgan $3,000 in connection with AI-hallucinated citations, and Goody Law Group was fined $1,000 in a separate incident involving fabricated AI citations.[4] Those amounts are not the full measure of the harm. The larger costs are credibility, client trust, judicial patience, and the professional burden of explaining how a false authority reached a filed document.

Citation Verification

Citation verification means confirming that each authority exists, that the citation is accurate, that quoted language appears in the source, and that the cited material supports the proposition for which it is used. With AI-assisted legal research, verification also includes checking whether the tool missed controlling authority or relied on outdated, overruled, distinguishable, or nonbinding material.

A sensible workflow does not treat citation checking as clerical cleanup after the legal analysis is complete. The verification step is part of the analysis. If a generated paragraph depends on three cases, those cases must be read before the paragraph becomes legal work product.

Source Coverage

Source coverage describes what materials a tool can search or retrieve from. A legal AI system may cover federal cases but not state trial orders, U.S. law but not foreign law, public law but not a client’s private contract archive, or one publisher’s database but not another’s.

This matters because a confident answer may reflect only the universe the tool can see. A lawyer evaluating an AI research result should ask what jurisdiction, date range, document type, and source collection were available to the system. The absence of a result is not the same as the absence of authority.

Agentic AI and Supervision

Agentic AI

Agentic AI is an evolving term for systems that can pursue a goal through multiple steps with some degree of autonomy. Instead of merely answering a prompt, an agentic system might plan tasks, call tools, search databases, draft messages, update a matter record, or trigger workflow actions.

The definition is still stabilizing, so it should not be treated as a settled product category. Thomson Reuters reported in its 2026 AI in Professional Services Report that 15% of professional services organizations had adopted agentic AI and 53% were planning adoption.[3] Those figures say something about market direction, not about whether any particular legal workflow is safe to automate.

The legal issue is authorization. A tool that drafts a research memo creates one type of review obligation. A tool that sends a client email, files a ticket, changes a deadline field, routes a contract for signature, or instructs another system creates a different one. Rules 5.1 and 5.3 supervision duties become more concrete when the “assistant” is not a person but still performs steps that affect legal work.

Human-in-the-Loop

Human-in-the-loop means a person reviews or approves an AI system’s work at a defined point. The phrase is useful only if the loop is specific. “A lawyer reviews it” is too vague for a policy, a client explanation, or a court-facing quality-control process.

• Who reviews the output?
• What exactly is reviewed: citations, facts, privilege, tone, legal reasoning, client instructions, or all of these?
• At what point does review occur: before internal circulation, before client delivery, before filing, or before an automated action?
• What happens when the reviewer disagrees with the system?

A review step that occurs after a document is filed is not a meaningful loop. Neither is a review step assigned to someone without the training, time, or authority to challenge the output.

Workflow Automation

Workflow automation uses software to move tasks through a defined process. AI may be added to classify incoming material, draft responses, summarize records, suggest next steps, or route work to the right person. The more a system can act without waiting, the more important the permission structure becomes.

For implementation details, the relevant operational question is not whether the workflow is modern. It is whether the firm can identify each automated step, each human approval point, each data transfer, and each person responsible for the final work product. See the site’s AI law firm workflow framework for a more detailed implementation lens.

Evaluation Terms That Change Legal Risk

Recall

Recall measures how many relevant items a system successfully finds. In e-discovery, high recall means the review process captures a large share of responsive documents. Low recall means relevant documents are being missed.

The risk of low recall is often severe because the error is invisible at the moment it happens. A missed privileged document, a missed hot document, or a missed responsive category may not announce itself until an opponent, court, or later investigation finds the gap.

Precision

Precision measures how many items identified by the system are actually relevant. In document review, low precision means reviewers spend time on many false positives. In contract analysis, it may mean the system flags ordinary clauses as risky or classifies irrelevant language as responsive.

Precision problems can waste money and attention, but the legal risk depends on the task. In privilege review, a false negative may be more dangerous than a false positive. In client-alert systems, too many false positives can cause warning fatigue. No single metric answers the professional question by itself.

False Positive and False Negative

A false positive is an item incorrectly flagged as belonging in a category. A false negative is an item incorrectly missed. Legal professionals should translate those terms into consequences: unnecessary review time, missed production material, missed privilege, inaccurate risk scoring, or a bad client communication.

Benchmarking and Validation

Benchmarking tests a tool against a defined set of tasks or examples. Validation checks whether a workflow performs adequately for its intended use. A benchmark may be useful, but a vendor’s benchmark does not automatically validate a firm’s matter-specific workflow.

For legal use, validation should be tied to the task and risk level. A tool used to summarize public articles may need a lighter process than a tool used to classify privilege, draft filed documents, or analyze confidential merger agreements. The more consequential the use, the more important it is to document how the tool was tested and who approved the workflow.

Confidentiality and Data Processing Terms

Prompt

A prompt is the instruction or input a user gives an AI system. In legal practice, a prompt may contain client facts, strategy, privileged communications, personal information, contract text, deposition excerpts, or draft legal analysis.

Because prompts can contain client information, they raise confidentiality questions under Model Rule 1.6. Before putting matter-specific information into any AI system, a lawyer or legal team should know who operates the system, whether data is retained, whether it is used for training, whether it is reviewed by humans, where it is stored, and what contractual protections apply.

Input, Output, and Training Data

Input is what the user or connected system provides to the tool. Output is what the tool returns. Training data is material used to train or further tune the model. These categories are often blurred in casual discussion, but they matter in confidentiality review.

A vendor may say that customer prompts are not used to train foundation models. That does not answer every relevant question. The firm still needs to know whether prompts are logged, retained, analyzed for abuse monitoring, stored by subprocessors, exposed to human review, or used to improve a separate product feature. A procurement answer should be read for what it says and what it does not say.

Third-Party Processing

Third-party processing occurs when client or firm information is handled by an outside provider or its subprocessors. In AI tools, this may include the application vendor, model provider, cloud host, analytics provider, support vendor, or logging service.

The legal issue is not solved by saying “the data is encrypted” or “the vendor is reputable.” Confidentiality review should identify the data being sent, the purpose of processing, retention periods, access controls, training-use restrictions, audit rights, incident notice obligations, and deletion procedures. ABA Formal Opinion 512 treats confidentiality as one of the core duties implicated by generative AI use.[1]

Data Retention

Data retention describes how long a provider keeps prompts, uploaded files, outputs, logs, metadata, or derived records. Retention can serve legitimate purposes, such as security monitoring or account administration, but it should be disclosed and bounded.

For legal teams, retention terms should be matched to matter sensitivity. A public marketing draft and a confidential acquisition agreement do not create the same retention risk. If a provider cannot explain what is retained and for how long, that is not a terminology problem; it is a procurement problem.

Ethics and Governance Vocabulary

Technological Competence

Technological competence is the obligation to understand relevant technology well enough to provide competent legal representation. It does not require lawyers to become engineers. It does require enough understanding to choose tools responsibly, evaluate outputs, protect client information, and avoid delegating legal judgment to software.

In AI work, technological competence includes knowing the difference between a tool that retrieves sources and one that merely generates text, between a system that stores prompts and one that contractually limits retention, and between a workflow that assists review and one that acts without meaningful supervision.

Client Communication and Consent

Client communication under Model Rule 1.4 may be implicated when AI use affects the representation in a way the client should understand, approve, or discuss. Not every use of spelling correction or formatting software requires a client conference. But AI use involving confidential data, material strategic judgment, outsourced processing, unusual billing implications, or client-specific restrictions may require communication or consent depending on the circumstances.

The useful vocabulary question is: what exactly will the tool do with client information or legal work product? A client cannot meaningfully consent to “AI” in the abstract. The conversation must describe the workflow closely enough for the client to understand the material risks and benefits.

AI Governance Policy

An AI governance policy defines how a firm or legal department approves tools, classifies use cases, protects data, verifies outputs, trains users, monitors vendors, and assigns responsibility. A policy that only says “do not enter confidential information” usually fails the moment real work begins, because real work is full of mixed materials, client exceptions, urgency, and unclear ownership.

• Approved tools and prohibited tools
• Permitted, restricted, and prohibited use cases
• Rules for client data, privileged material, and personal information
• Citation, quotation, and factual verification standards
• Supervision responsibilities for lawyers, staff, and vendors
• Incident response when an AI output creates a filing, confidentiality, or client-service problem

State Bar AI Guidance

State bar AI guidance is developing quickly, and counts of how many jurisdictions have issued guidance should be treated as maintenance items rather than permanent facts. Secondary trackers collected substantial state-level activity by 2025, but the landscape changes as new opinions, reports, and ethics materials are released.[6]

The common practical themes are more stable than the count: competence, confidentiality, communication, candor, supervision, billing, and verification. A lawyer evaluating AI use should check the current guidance in the relevant jurisdiction and should not assume that a national article, vendor resource, or internal policy captures all local expectations.

Terms That Deserve Careful Skepticism

AI-Powered

AI-powered is a marketing phrase, not a technical explanation. It may describe anything from a simple classifier to a generative AI system connected to multiple data sources. The correct follow-up is: powered how, by what model or method, using what data, with what controls, and for what task?

Professional-Grade

Professional-grade suggests that a tool is suitable for serious work, but the term has no fixed meaning. In legal technology, it should point to concrete features: source-linked answers, legal database coverage, confidentiality controls, administrative permissions, audit logs, contract terms, support obligations, and clear limits on training or retention.

Safe

Safe is always task-specific. A tool may be safe enough for brainstorming a CLE outline and unsafe for drafting a dispositive motion without close review. It may be acceptable for public information and unacceptable for privileged client facts. Legal teams should avoid global approvals that ignore task, data, user, jurisdiction, and output destination.

Maintenance Note

This glossary should be reviewed as AI tools, ethics guidance, and court responses evolve. The most likely entries to need updates are agentic AI, state bar guidance, vendor data-processing terms, and legal research architectures. Internal cross-links should connect this glossary to tool comparisons, workflow explainers, regulatory trackers, and risk records so readers can move from definition to implementation without losing the professional responsibility thread.

Legal professionals do not need to become AI engineers. They do need enough vocabulary to ask competent questions, verify outputs, supervise use, protect confidentiality, and recognize when a tool’s marketing language hides a legal risk. The accountable human does not disappear because the interface has become easier to use.

References

1. ABA issues first ethics guidance on a lawyer’s use of AI tools, American Bar Association, July 2024.
2. 8am Legal Industry Report, ABA Law Practice Magazine, March-April 2026.
3. AI in Professional Services Report 2026, Thomson Reuters, 2026.
4. AI Legal Issues and Concerns for Legal Practitioners, American Bar Association, 2025.
5. Colorado Technology Law Journal overview, Colorado Technology Law Journal.
6. 2025 State Bar Guidance on Legal AI, Paxton, 2025.