Wiki · Concept · Last reviewed August 12, 2026

AI in Legal Practice and Courts

AI in legal practice is the use of machine-learning and generative systems inside legal-service and court workflows. The accountable object is the configured workflow—not the model alone—and the central safety question is whether professional judgment, authoritative sources, protected information, due process, and a reviewable human decision remain intact.

Category: Concept Published: June 15, 2026 Modified: August 12, 2026 Last reviewed: August 12, 2026 Tags: Legal AI, Courts, Professional Ethics, Hallucinations, Agentic AI, Evidence, Access to Justice

Definition

AI in legal practice refers to machine-learning or generative systems used to retrieve, classify, summarize, translate, draft, recommend, or act on legal information and matter data. Uses include legal research, contract review, due diligence, discovery, client intake, compliance, litigation support, legal operations, court administration, public legal help, and adjudicative support.

The category includes general-purpose assistants, legal-specific research products, retrieval-augmented systems, e-discovery and contract analytics, court chatbots, internal firm or legal-department systems, and agents that can sequence tasks or use external tools. For governance, the relevant unit is the configured system: model, prompts, retrieval corpus, matter data, connectors, permissions, interface, human reviewers, vendor terms, and deployment setting.

The defining issue is not whether a system can produce legal-sounding text. It is whether the workflow preserves professional judgment, authoritative sourcing, client confidentiality, privilege and work-product protections, due process, evidence reliability, and institutional accountability. AI does not remove legal duties; it redistributes the points at which those duties can fail.

This page is descriptive, not legal advice. Professional obligations and court rules depend on jurisdiction, matter type, tool design, client instructions, local rules, and the facts of use.

Operational Distinctions

Assistive legal AI helps a lawyer or legal worker search, summarize, draft, classify, translate, or compare materials. It should be treated as draft assistance, not authority.

Agentic legal AI plans or sequences work: selecting sources, searching repositories, reading matter files, drafting deliverables, revising files, or using connectors. Its risk is upstream because it can frame the issue, select the record, and alter work product before a human sees a final document.

Public-facing legal AI serves self-represented litigants, tenants, benefit applicants, consumers, or court users. Its risk is not only wrong law; it is the blurring of legal information, procedural help, triage, form assistance, and legal advice for people who may have no alternative source of guidance.

Court-administration AI supports scheduling, document routing, transcription, translation, public information, or other operational work. These uses differ from adjudication, but can still affect access, deadlines, privacy, and equal treatment.

Adjudicative assistance helps research or interpret law and facts, assess evidence, or prepare judicial work. Because it sits closest to public authority, it needs independence, reason-giving, review, appealability, and a record showing that the legally responsible decision-maker exercised judgment.

Machine-generated evidence is output produced by software or AI and offered to prove a fact, such as a forensic score, automated measurement, generated summary, or synthetic-media analysis. Its reliability and admissibility questions are distinct from whether a lawyer used AI to draft a filing.

Grounding is not authority. A system can retrieve a real case and still misstate it, omit controlling law, use the wrong jurisdiction, or attach the source to a proposition it does not support. A citation link is an inspection path, not a validity guarantee.

Legal Uses

Research. AI can summarize cases, statutes, regulations, treatises, and briefs. Legal-specific systems may connect model output to licensed databases, but they still require verification.

Drafting. Lawyers use AI to draft memos, contracts, pleadings, correspondence, deposition outlines, discovery requests, policies, and client-facing explanations.

Review. AI can help classify documents, find clauses, identify conflicts, summarize records, compare versions, and extract facts from large files.

Legal operations. Corporate legal teams use AI for matter triage, billing review, policy management, compliance monitoring, vendor review, and workflow automation.

Courts and public services. Courts and legal-aid organizations can use AI for intake, translation, document routing, plain-language explanations, form completion, and administrative efficiency. These uses require extra care because many users lack lawyers.

Current Landscape

As of August 12, 2026, legal AI is no longer only a public-chatbot problem. Major legal vendors market systems for research, drafting, document analysis, and multi-step workflows. Thomson Reuters launched CoCounsel Legal in August 2025 with Deep Research and guided agentic workflows; LexisNexis announced general availability of Protégé in January 2025 as a personalized assistant grounded in LexisNexis and customer content. These announcements establish product direction, not accuracy, ethical compliance, or fitness for a particular matter.

The practical shift is from answer box to workbench: AI is being connected to legal databases, matter repositories, drafting environments, contract systems, and office tools. That expands both utility and the failure surface. A system that can choose sources, read privileged files, revise a document, or trigger an external action needs controls over permissions, data, verification, and change—not only a warning that generated text may be wrong.

Federal court governance remains interim and decentralized. The Administrative Office of the U.S. Courts reported that its AI Task Force developed temporary guidance in 2025 cautioning against delegating core judicial functions, calling for independent review and verification, and asking courts to define approved tasks and consider disclosure. The National Center for State Courts separately recommends beginning with lower-risk tasks, human review, data governance, terms-of-use review, and staff training.

Professional regulation is also moving in stages. ABA Formal Opinion 512 applies existing Model Rules to generative AI; it is ethics guidance, not a nationwide AI statute or a substitute for controlling jurisdictional rules. California approved revised practical guidance addressing agentic AI in May 2026. Separate proposed amendments to six California professional-conduct rules completed a second public-comment period on August 6, 2026; as of this review, they remained proposals rather than operative amendments.

The EU AI Act draws a sharper line between administration and adjudication. Annex III classifies certain systems used by or for judicial authorities and alternative-dispute-resolution bodies to assist with researching and interpreting facts and law or applying law to facts as high-risk, while excluding purely ancillary administrative activities. Regulation (EU) 2026/1744 moved application of the relevant Annex III high-risk requirements to December 2, 2027. The statutory classification is current; those delayed high-risk duties are not yet fully applicable.

Access-to-justice organizations face a different measure of success than commercial legal teams. Public-facing AI improves access only if users can understand its jurisdiction and limits, reach a human or authorized provider when needed, correct errors, and avoid mistaking procedural information for individualized legal advice. A faster answer is not access to justice if it is unusable, inaccessible, stale, or wrong.

The federal evidence-rules process remains unsettled. Proposed Rule of Evidence 707, aimed at machine-generated evidence offered without expert testimony, was published for comment in 2025. In May 2026, the Advisory Committee on Evidence Rules declined to advance its revised proposal for another comment period, planned further expert study, and continued separate work on deepfake-authenticity issues. Rule 707 is therefore a proposal under study, not a Federal Rule of Evidence in force.

Professional Ethics

The American Bar Association's Formal Opinion 512, issued July 29, 2024, explains how existing Model Rules apply when lawyers use generative AI. It addresses competence, confidentiality, communication, supervision, meritorious claims, candor to tribunals, and reasonable fees. The opinion is an important national reference, but lawyers must identify the rules, ethics opinions, court orders, and law that actually govern their jurisdiction and matter.

Competence means understanding a tool's relevant capabilities, data sources, limits, and risks well enough to use it responsibly. Confidentiality requires evaluating third-party tools, connectors, uploads, logs, retention, training terms, subcontractors, and access controls before client information enters the system. Supervision means responsibility does not disappear when work is delegated to another lawyer, a nonlawyer, a vendor, or software.

The State Bar of California approved updated practical guidance on May 14, 2026, replacing its 2023 version and addressing agentic AI. California's later rule-amendment proposals would expressly address competence, communication, confidentiality, candor, managerial policies, and supervision, but proposal language should not be cited as an already-effective rule.

Confidentiality and evidentiary privilege are related but not interchangeable. Before using matter data, counsel should determine what professional rule protects it, whether a third party can access or retain it, whether client consent is required, and what applicable law says about privilege or work-product consequences. A vendor's “private” setting is not itself a legal conclusion.

Client communication is context-specific. AI use may need to be discussed when it materially affects the representation, changes confidentiality assumptions, implicates client instructions, alters outsourcing or vendor arrangements, or changes how fees and expenses are calculated. A silent tool choice can become an ethics issue when the client reasonably needed to know about it.

The ethical baseline is delegation without abdication. A lawyer may use AI to assist research, drafting, review, and administration, but the lawyer still owes independent professional judgment, candor to the tribunal, confidentiality, communication with the client where required, reasonable billing, and compliance with court-specific rules.

Billing is part of the ethics problem. If AI reduces the time required for a task, hourly billing must reflect actual time spent, and separate AI costs require careful treatment under fee agreements and professional-responsibility rules. Efficiency cannot be quietly converted into hidden overbilling.

Courts and Filings

The legal profession's warning case is Mata v. Avianca, where lawyers were sanctioned in 2023 after filing fake cases and quotations generated through ChatGPT and failing to verify them. The lesson is narrower and harsher than "AI can hallucinate": legal professionals cannot outsource their duty of candor to a fluent system.

A published Ninth Circuit sanctions order, Lnu v. Blanche, added a current appellate statement in June 2026. The court imposed monetary sanctions and six-month suspensions after briefs contained nonexistent cases, misattributed quotations, and gross misrepresentations, followed by repeated failures of candor. The order emphasized that it was not sanctioning the use of generative AI itself: the governing duties attach when counsel signs and files material without reading and verifying the authorities, and candor requires prompt, transparent correction when an error is discovered.

Stanford RegLab and HAI researchers tested leading AI-powered legal research products in 2024 and found that legal-specific retrieval systems reduced hallucinations compared with the general-purpose system tested but did not eliminate them. In the published study, Lexis+ AI and Ask Practical Law AI produced incorrect or misgrounded information in more than 17 percent of benchmark responses, while Westlaw AI-Assisted Research did so in more than 34 percent. Those figures describe named product versions, a preregistered query set, and the study's definition of hallucination; they are not current scores for every legal-AI product.

The deeper issue is source discipline. A case can exist and still be cited for the wrong proposition. A quotation can be real and still omit the limiting context. A RAG system can retrieve an authoritative source and attach it to a false synthesis. For legal use, verification must check existence, citation, quotation, holding, jurisdiction, procedural posture, current validity, and fit between source and claim.

Source discipline applies to the factual record as well as legal authority. An AI-generated deposition summary, medical chronology, discovery digest, administrative-record timeline, or contract-exception list must be traceable to record citations, page and line references, or source documents. The model can help locate the passage, but it should not become the file's only memory.

Court approaches remain nonuniform. Some courts and judges use standing orders, disclosure requirements, certification rules, or AI-use guidance; others rely on existing duties of candor, signature obligations, professional discipline, and sanctions. A filer must check the governing court, judge, local rule, standing order, and filing date rather than infer a nationwide rule from a single order. Disclosure can support accountability, but it does not replace verification.

Courts face a second problem when AI enters chambers, court administration, evidence review, translation, public help desks, or draft orders. A lawyer's bad filing can be sanctioned after the fact. A court's bad AI-assisted order, chatbot answer, translation, or evidence ruling can damage public legitimacy. Court AI therefore requires stricter attention to independence, review, appealability, records, security, accessibility, and the boundary between administrative assistance and adjudication.

Risk Pattern

Fabricated authority. AI can produce plausible-looking case names, citations, quotations, holdings, statutes, or procedural histories that do not exist or do not say what the output claims.

Misgrounded authority. Legal RAG systems can cite real cases, statutes, regulations, or practice materials that fail to support the proposition attached to them.

Jurisdiction and freshness error. A legally accurate sentence can still be unsafe if it reflects superseded law, a different forum, an inapplicable procedural posture, or a database that was not current when the answer was produced.

Confidentiality leakage. Client facts, privileged communications, draft strategy, or settlement material can be exposed through unsafe tools, connectors, prompts, logs, vendors, subcontractors, retention settings, or training pipelines.

Agentic workflow risk. Systems that plan tasks, search sources, use connectors, revise documents, or draft filings can shape legal work before a lawyer sees the final output.

Record contamination. A generated summary can introduce invented facts, merge witnesses, flatten uncertainty, or detach a statement from its source and then propagate that error into chronologies, briefs, advice, or later retrieval.

Overreliance. Legal users may accept fluent analysis because it sounds like legal writing, especially under deadline pressure.

Unauthorized practice of law. Tools that give legal guidance directly to non-lawyers can cross legal and ethical boundaries if they substitute for licensed counsel without appropriate safeguards.

Bias and access gaps. AI legal tools can encode unequal data, misread marginalized users, or make premium legal assistance even more powerful for those who can pay.

Public self-help ambiguity. A tool built for legal information can slide into individualized legal advice if it asks for facts, predicts outcomes, recommends strategy, or drafts filings without appropriate boundaries.

Billing distortion. If AI reduces time spent, lawyers still must charge reasonable fees and communicate appropriately about AI use where duties require it.

Apprenticeship erosion. If AI absorbs first-pass research, cite checking, document review, chronology building, and drafting without replacement training, junior legal workers may lose the work that teaches source discipline.

Evidence fragility. If prompts, retrieved sources, model versions, and outputs are not preserved, it becomes difficult to reconstruct how a legal document or decision was produced.

Court legitimacy risk. AI errors in judicial drafts, public-facing court tools, translations, evidence screening, or administrative routing can undermine trust because courts act with public authority.

Governance Requirements

Controls should follow the consequence of the use, not the prestige of the product. Brainstorming from public material, searching a confidential matter repository, advising a client, filing in court, guiding a self-represented person, and assisting adjudication require different approval, evidence, and review gates.

Use legal AI under a written policy that distinguishes public tools, approved tools, legal-specific tools, agentic tools, confidential matters, client-restricted matters, and prohibited uses.
Maintain an AI tool register that names each approved product, vendor, owner, allowed use, prohibited use, data-retention setting, connector scope, review requirement, and renewal date.
Classify the task and matter before use: public information, internal administrative work, ordinary client work, privileged strategy, regulated data, court filing, client advice, public legal help, or judicial work.
Assign an accountable human owner with the competence, time, evidence, and authority to reject, correct, narrow, or stop each consequential AI-assisted workflow.
Verify every legal citation, quotation, rule statement, factual assertion, procedural claim, and jurisdictional claim against authoritative sources before filing, advising, or relying on it.
For filings and formal advice, preserve a source packet or cite-check trail showing the authorities and record materials that support the submitted claims.
Separate generation from verification. The same model output that generated a claim should not be treated as proof that the claim is true.
Use least privilege for agentic tools. Read-only access, matter-scoped repositories, disabled write-back, approval gates, and revocation should be the default for legal agents unless stronger access is justified.
Protect client information through vendor and subcontractor review, data minimization, retention controls, encryption, access controls, connector limits, matter-level permissions, and clear rules on what may be pasted, uploaded, stored, or transmitted.
Require human review proportionate to risk, with stronger controls for filings, client advice, settlement communications, privileged investigations, regulatory submissions, public legal help, and court work. Do not permit an automated draft-to-file path.
Define disclosure and certification duties by forum and client, including local court rules, standing orders, engagement terms, client instructions, and internal matter policy.
Train lawyers, judges, clerks, staff, and legal operations teams on hallucination, misgrounded citations, privilege, supervision, billing, disclosure, bias, evidence reliability, and local court rules.
Evaluate the configured workflow on representative tasks, jurisdictions, adversarial prompts, false premises, stale-law traps, record citations, access controls, and failure escalation before consequential use and after material changes.
Track product, model, prompt, retrieval, connector, permission, and policy changes. Revalidate the workflow when a change could alter sources, behavior, data exposure, or reviewer assumptions.
Preserve prompts or prompt references, outputs, retrieved sources, model or product versions, tool calls, approvals, and review notes for high-stakes matters where reconstruction may be needed, subject to privilege, privacy, security, records, and deletion rules.
Design court and access-to-justice tools to clearly distinguish legal information, procedural help, document assistance, triage, and legal advice.
Give public-facing tools jurisdiction limits, source and freshness notices, plain-language boundaries, escalation paths, accessibility and language testing, complaint routes, and records sufficient to investigate harmful guidance. A disclaimer should not carry the whole safety case.
For court systems, separate administrative from adjudicative use, protect decisional independence, document the human decision path, and preserve meaningful notice, correction, challenge, and appeal where rights may be affected.
Map legal AI governance to a broader risk-management framework, such as the NIST AI RMF and the NIST Generative AI Profile, while adding legal-specific controls for privilege, candor, court rules, and source verification.
Maintain an incident process for hallucinated authority, confidential-data exposure, erroneous filings, unsafe public guidance, vendor failures, and AI-generated work that reaches a client or court before adequate review.

Minimum Matter Record

For consequential legal work, keep a matter-level record sufficient for a qualified reviewer to reconstruct the AI-assisted step without treating the model transcript as the official file. The minimum record should identify:

the matter or workflow, intended task, risk classification, accountable owner, and applicable client, court, or internal restrictions;
the approved product, model or service version where available, material configuration, prompt-template version, connectors, permissions, and data-retention setting;
the input documents or stable references, retrieval sources, database or law-current-through date, jurisdiction, and search scope;
the generated output, material tool actions, human edits, authorities and record cites checked, verifier, approval, and final disposition;
any required client or forum disclosure, error or incident, correction, notice, escalation, preservation hold, and later model or source change relevant to the work.

Retention should be purpose-limited. Use restricted evidence stores, stable references, hashes, redactions, or matter-system links when they support reconstruction without duplicating privileged material, personal data, credentials, or sealed records into a broad AI log.

Source Discipline

Legal AI claims need strict source discipline because legal text can become legal action. Identify the source's authority and status: enacted law and rules; controlling opinions and filed orders; regulator or judiciary policy; professional-responsibility guidance; empirical research; or vendor documentation. These sources answer different questions. A vendor launch post can show that a feature was announced; it does not prove reliability, ethical compliance, admissibility, or fitness for a matter.

Dates and procedural posture are part of the claim. Guidance is not a rule, a proposed amendment is not an operative duty, a public-comment draft is not an enacted evidence rule, and an order from one court is not a nationwide filing requirement. Current-law checks should record the jurisdiction, source, effective or decision date, later history, and the date through which the research was validated.

For legal research, verify at the claim level. Check that the authority exists, that the citation is correct, that the quotation is exact, that the proposition matches the holding or rule, that the case remains good law, that the jurisdiction and procedural posture fit, and that contrary authority has not been erased by the system's synthesis.

For court and evidence claims, prefer primary records: the filed order, standing order, advisory-committee report, rule text, docket entry, regulator page, or official judiciary guidance. News reports and vendor summaries can identify issues, but the article should not treat them as substitutes for the source that creates the legal obligation or records the court action.

For empirical claims, preserve the test context. The Stanford legal-RAG study tested specific products, query sets, definitions of hallucination, and time windows. Its durable lesson is that legal-specific retrieval systems can reduce some hallucination rates without eliminating misgrounded or false legal claims; it should not be stretched into a universal score for every later product version.

Spiralist Reading

Legal AI is the Mirror speaking in the voice of authority.

Law is a language that changes reality: a filed motion, a signed contract, a citation, a court order, a waiver, a confession, a settlement demand. When AI speaks legal language fluently, it does not merely imitate style. It enters a ritual system where words have institutional force.

For Spiralism, legal AI shows why fluency is not authority. The machine can sound like precedent while inventing precedent. It can sound like counsel while lacking duty. It can sound like certainty while concealing probabilistic assembly. The safeguard is not awe. It is verification, responsibility, and a human professional who remains answerable for the words.

Open Questions

Which legal AI uses should require client disclosure, and which should be treated like ordinary legal technology?
Should courts require AI-use disclosure in filings, or rely on existing duties of candor, Rule 11-style certification, and sanctions?
How should public-facing legal tools distinguish legal information, procedural assistance, triage, and individualized legal advice?
What evidence should a lawyer, court, or vendor preserve when an agentic legal workflow searches sources, drafts text, or uses matter documents?
How should legal organizations preserve enough evidence for audit, discovery, correction, and appeal without building a second, overbroad archive of privileged or personal information?
Which judicial tasks are genuinely administrative, and which influence the interpretation of facts or law enough to require adjudicative safeguards?
How should law firms preserve training and apprenticeship when AI absorbs first-pass research, cite checking, document review, and drafting?

Sources

American Bar Association, Formal Opinion 512: Generative Artificial Intelligence Tools, July 29, 2024.
State Bar of California, Ethics & Technology Resources, noting May 14, 2026 approval of updated generative AI guidance.
State Bar of California, Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, 2026 update.
State Bar of California, Proposed Amendments to the Rules of Professional Conduct Related to Artificial Intelligence, second public-comment deadline August 6, 2026.
United States District Court, Southern District of New York, Mata v. Avianca, Inc., Opinion and Order on Sanctions, June 22, 2023.
U.S. Court of Appeals for the Ninth Circuit, Lnu v. Blanche, published sanctions order, June 3, 2026.
Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning, and Daniel E. Ho, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies 22, no. 2 (2025): 216-242.
Administrative Office of the U.S. Courts, Court Operations - Annual Report 2025, section on developing artificial intelligence policies.
Advisory Committee on Evidence Rules, Report of the Advisory Committee on Evidence Rules, May 17, 2026.
European Parliament and Council, Regulation (EU) 2024/1689 (Artificial Intelligence Act), including Article 6 and Annex III.
European Parliament and Council, Regulation (EU) 2026/1744, amending the AI Act implementation timetable, July 24, 2026.
National Center for State Courts, AI and the Courts: Interim Guidance - Getting Started, March 2024.
National Center for State Courts, AI and the Courts: Platform Considerations, March 2024.
National Center for State Courts, Artificial intelligence resources for courts, reviewed August 12, 2026.
Legal Services Corporation, Technology Summit Report 2024-2025, reviewed August 12, 2026.
NIST, AI Risk Management Framework, reviewed August 12, 2026.
NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, July 26, 2024.
Thomson Reuters, Thomson Reuters Launches CoCounsel Legal, August 5, 2025.
LexisNexis, LexisNexis Introduces Protégé Personalized AI Assistant with Agentic AI, January 27, 2025.
U.S. District Court, Western District of North Carolina, Standing Order - In Re: Use of Artificial Intelligence, June 27, 2024.

Return to Wiki