The cases exist. The quotations look plausible. The argument still may be wrong. A legal brief can contain real authorities that do not support its propositions, facts without record support, and a procedural request that does not fit the matter. Checking whether a citation resolves is necessary preparation, but it is only one part of reviewing the document.
AI can help organize a draft and surface research questions. The responsible reviewer needs to connect each material assertion to the relevant authority or record and understand the reasoning. This guide offers an editorial review method with a hypothetical brief, a proposition ledger, and defined release criteria. It does not prescribe a universal filing rule or certify that the method satisfies a particular court's requirements.
ABA Formal Opinion 512 interprets the ABA Model Rules and discusses independent verification and professional responsibility. New York Judicial Ethics Opinion 26-37 addresses a specific referee scenario involving AI-generated frivolous papers. Neither is a nationwide sanction schedule or a substitute for the rules governing your actual filing.
Define the filing task and its sources
Begin with the document's purpose. What relief is requested, what procedural stage applies, what issue must the court decide, and what record is available? A draft that sounds persuasive can conceal a mismatch between the argument and the question the tribunal actually needs to answer. Put those requirements in the review record before examining the prose.
Identify the authoritative source set. It can include the docket, operative pleadings, orders, evidence, transcripts, applicable rules, and retrieved authorities. Associate the draft with exact versions and relevant dates. If the assistant was given a prior complaint or an incomplete transcript, a fluent paragraph may accurately summarize the wrong material. The reviewer should know which inputs shaped the draft.
Keep fact development separate from legal research. A legal source can establish a proposition of law; it cannot establish that the client performed an action on a particular date. Likewise, a factual record may show an event without supporting the legal consequence the draft assigns to it. Your ledger should make those different sources visible rather than combine them into one generic citation field.
For the hypothetical brief, counsel is reviewing a motion concerning a disputed contractual notice. The source package includes the agreement, an amendment, correspondence, and an existing procedural order. The assistant received the agreement and correspondence but not the amendment. Its draft treats an old notice provision as operative. Counsel rejects that premise before checking the string of cases built around it.
Describe what the tool may do. It might prepare a draft outline from verified sources, propose a research query, or identify sentences lacking a citation. It should not submit the document, alter the record, or supply facts absent from the source set. The permissions and review method should fit the task. A successful citation-location feature does not establish permission for automated filing or unsupervised legal argument.
Build a proposition ledger before polishing prose
Break the brief into material assertions. Record the proposition, its role in the argument, the supporting authority or factual source, the exact location, and the review disposition. Include statements in introductions and conclusions, where unsupported assumptions often hide because the text looks like advocacy rather than evidence. A final rhetorical sentence can still assert a consequential fact or legal rule.
Distinguish categories. A proposition can state law, characterize a holding, describe a record fact, draw an inference, or recommend a result. Each needs a different check. An inference may be reasonable without appearing verbatim in the source, but the reviewer should understand its premises and how it is presented. A legal characterization should not be labeled a quotation unless the words actually match.
Use the ledger to expose overstatement. If a case supports a narrow proposition concerning particular facts, the draft should not announce a universal rule. If the record shows that an email was sent, the brief should not automatically say it was received by the required recipient. The missing step matters even when the sentence reads naturally.
Assign each material open item to a person. A ledger saying check this next to several propositions does not tell the team who will finish the research or whether the filing depends on it. The item should identify the question, current source, missing evidence, and proposed disposition if the answer cannot be obtained. Counsel can then decide whether to revise the claim, seek additional time where appropriate, use another argument, or take another action under the actual circumstances. The assistant should not make that consequential choice merely to complete the table.
In the hypothetical ledger, one sentence says the client timely delivered notice. The source shows an email date but leaves the recipient and attachment uncertain. Counsel changes the disposition to factual support unresolved and asks for the relevant record. Another sentence states that a court always enforces a notice provision strictly. The cited authority addresses a narrower circumstance. Counsel rewrites the proposition and researches the applicable question without preserving the draft's certainty.
The ledger should remain usable. Focus on material assertions rather than creating a burdensome record for every transition phrase. Use it to identify what must be resolved before release and what can be removed if support is unavailable. The goal is an understandable argument with traceable foundations, not a spreadsheet whose existence persuades reviewers that the work must be correct.
Verify that the authority is what the draft claims
Retrieve each relied-on authority through an appropriate source and inspect the document. Confirm the case identity, court, date, procedural setting, relevant text, and relationship to the asserted proposition. A resolving hyperlink or recognizable reporter format does not establish that the authority supports the argument. An incorrect pinpoint can send a reviewer to a real case while concealing the missing support.
Check the jurisdiction and authority level in context. A decision from another jurisdiction may have a role in the argument, but the draft should accurately describe that role. Do not let a generated paragraph call persuasive authority controlling simply because its reasoning sounds favorable. The reviewer must determine the applicable treatment for the actual matter.
Distinguish the court's reasoning from party arguments, background quotations, and descriptions of other decisions. A tool can lift a statement from an opinion without recognizing who asserted it or whether the court adopted it. Read enough context to understand the source. A paragraph copied from a losing argument is not a holding supporting the client's position.
The hypothetical brief cites a real opinion for a categorical statement about notice. The retrieved text shows the statement appears in a description of a party's contention, while the court resolves the case on another basis. Counsel marks the citation unsupported for that proposition and removes it from the argument. The error would survive a check limited to case existence.
Record the result in the ledger: supports as stated, supports a narrower statement, does not support, or requires additional research. If the source cannot be located, do not ask the same assistant to assure you it exists. Retrieve independently, pursue another verified source, or remove the reliance pending resolution. The review method should make lack of support visible rather than convert repeated generation into confidence.
Check subsequent treatment and the current legal question
Source retrieval answers what an authority says. Current validity and treatment are different questions. Use an appropriate research method, including a citator where suitable, to investigate relevant subsequent history and treatment. The precise method depends on the jurisdiction, authority type, and issue. Do not assume that a connected legal AI product completes every necessary check merely because it can retrieve a real opinion.
Read treatment in context. A flag may concern a different issue, while a clean-looking result may not settle the proposition you need. Identify what happened, which reasoning is affected, and whether the draft's reliance remains appropriate. A status label is an entry point into review, not the full legal analysis.
Check the legal question's date and procedural context. A statute, rule, or order may have changed. The matter may concern an earlier version. An AI answer can cite current text without addressing the time relevant to the client's facts, or cite historical text as though it remains operative. Preserve the version and date that counsel has actually assessed.
In the hypothetical brief, the assistant proposes an authority whose later treatment affects the relied-on notice reasoning. Counsel examines the subsequent opinion rather than relying on the initial summary's positive tone. The ledger records the affected proposition and a revised research question. The argument may be rebuilt with narrower support or removed; the correct choice depends on counsel's substantive analysis, not on keeping the draft intact.
Record the review date and any remaining open question. Do not suggest that a check performed months ago proves a filing is current today. The proposed release process should include whatever update counsel determines appropriate for the actual filing. This is an editorial practice recommendation, not a claim that a particular interval satisfies every court or professional rule.
Verify quotations and parentheticals against context
Compare quotations with the retrieved source, including punctuation, omissions, bracketed changes, and surrounding text. Check whether ellipses alter meaning and whether a quoted sentence depends on a condition the brief leaves out. A quotation can be textually accurate and substantively misleading. The review needs both the words and their role in the source.
Treat parentheticals as assertions. A parenthetical saying the court held a proposition may overstate the decision or ignore its procedural posture. An assistant may turn a complex analysis into a short phrase that sounds authoritative. Read the relevant portion and write a characterization the reviewer can defend. Do not assume that a concise parenthetical is merely a formatting issue.
Check repeated quotations independently where context differs. A sentence used in a background section may serve a different purpose when reused in the argument. Ensure each use fits its proposition. An automated search for matching text can assist the comparison but cannot establish that the quotation's meaning fits the draft's reasoning. Preserve a note identifying the context reviewed for material quotations. Another reviewer can then find the passage efficiently without assuming that a copied excerpt tells the entire story. That note supports inspection; it does not replace reading the source when the point matters.
The hypothetical brief quotes a source stating that a requirement applies unless the agreement provides otherwise. The assistant omits the exception. Because the client's amendment may bear on that exception, the omission matters. Counsel restores the context and revises the analysis rather than keeping the favorable fragment. The ledger identifies both the textual defect and the factual question the full statement creates.
Set a proposed acceptance criterion that no relied-on quotation remains unverified and no parenthetical claims a holding the reviewer cannot locate and explain. The criterion concerns this review method. It does not promise that all legal reasoning becomes correct through quotation checking. Review still needs to assess the argument, contrary material, facts, and the requested relief.
Inspect introductory characterizations of the authorities too. A phrase such as every court agrees can overstate a collection of narrower decisions even when all quoted passages are accurate. A heading can announce a result broader than the section proves. Ask the reviewer to compare those statements with the ledger and remove unsupported certainty. In the notice example, counsel can explain the particular principle the verified authorities support without claiming unanimity the research has not established. The correction often makes the argument easier to follow because the reader can see its actual premises. Avoid replacing an overstatement with a vague hedge that conceals the issue. State the supported proposition clearly and preserve the distinction or open question that matters to the tribunal's analysis.
Trace factual assertions to the actual record
Give each material factual assertion an inspectable source. Identify the document, page, paragraph, exhibit, or other appropriate location. Distinguish an allegation, witness statement, admission, correspondence, and established finding. A generated summary can collapse these categories and present a disputed claim as a fact agreed by everyone.
Review identity, chronology, amounts, and attribution. A correspondence thread can contain forwarded text with dates that differ from the event described. A summary may attribute a statement to the person who forwarded it instead of the original sender. An attachment can change the meaning of an email. Inspect the full relationship rather than treating a snippet as the entire record.
Ask whether the source supports the degree of certainty. The fact that the client says notice was sent may support a description of the client's account. It may not support an unqualified statement that the required notice was delivered. The governing procedural treatment of disputed facts requires counsel's assessment in the actual matter. The workflow's job is to reveal the source and uncertainty.
In the hypothetical motion, the assistant states that all contractual documents were supplied to the other party. The record contains an email referencing attachments, but the preserved copy does not show which files were attached. Counsel requests the transmission record and changes the draft pending confirmation. The reviewer does not fill the gap with what probably happened to make the narrative flow.
Keep factual corrections visible to the person reviewing the legal argument. If the recipient, date, or document changes, the analysis built on that fact may need revision. A separate proofreader correcting a factual sentence is insufficient if counsel continues relying on the original premise elsewhere. The ledger should connect the fact to affected propositions and notify the argument owner.
Review summaries sent to the client through the same lens. A status email can imply that counsel confirmed a favorable fact when the litigation draft still marks it unresolved. Use the actual evidence status in the explanation. For the notice example, counsel can state that the transmission record has been requested and that the argument will be assessed after its receipt. This language lets the client understand why a polished draft is not ready for reliance. It also prevents the client from repeating an unsupported assertion in a separate communication. Keep the correction limited to what the team knows rather than announcing that the entire case is stronger or weaker before the missing facts have been examined.
Look for omitted issues and contrary material
A generated draft tends to answer the question it was asked. If the prompt assumes a favorable conclusion or omits an important issue, the draft may develop a polished argument around that narrow frame. Ask the reviewer to identify the strongest objections, missing legal questions, and facts that could change the result. This is substantive review, not a request for decorative counterarguments.
Use independent research and professional judgment rather than treating the assistant's second answer as a complete adversarial check. Asking the same tool for contrary cases may surface useful leads, but the resulting list still needs verification and contextual analysis. The absence of a generated objection is not evidence that none exists.
Review the duty concerning contrary authority under the actual governing rules and circumstances. ABA Formal Opinion 512 discusses professional responsibilities including candor and meritorious contentions. This article does not supply a universal test for which authority must be disclosed or how a particular filing must address it. Counsel must assess that question directly.
In the hypothetical notice dispute, the assistant focuses on the master agreement and favorable decisions. The reviewer identifies an amendment and an unfavorable factual distinction. Counsel revises the issue framing and evaluates the consequence. The stronger argument may be narrower than the first draft. Keeping a broad unsupported claim because it reads persuasively would defeat the review's purpose.
Create a final open-issues list. It should identify unresolved contrary material, missing facts, uncertain procedural assumptions, and points removed for lack of support. Assign each remaining issue or remove its associated claim. A document should not be labeled ready simply because the citations in its retained paragraphs look clean while the central question remains unanswered.
Review the reasoning after the source checks
Source checks supply the premises. They do not establish the inference. Ask whether the argument's conclusion follows from the verified propositions and facts, whether material distinctions are addressed, and whether the requested relief fits the reasoning. A brief can pass every citation check and still ask the tribunal to make a leap the authorities do not support.
Read an outline stripped of rhetorical language. State the legal standard as counsel understands it, the facts relevant to that standard, the contested point, and the requested result. If the logic becomes difficult to explain without the draft's persuasive phrases, investigate why. The assistant may have produced continuity of prose rather than continuity of reasoning.
Check internal consistency. An introduction might characterize notice as undisputed while the argument acknowledges a contested recipient. A conclusion might request relief broader than the motion described. A section heading may announce a categorical rule the body qualifies. These inconsistencies often emerge when several generated drafts are combined. Assign someone to review the whole document, not only isolated sections.
The hypothetical brief initially argues both that the amendment controls and that the master agreement independently resolves every issue. Counsel explains the relationship carefully and removes the inconsistent shortcut. The accepted argument identifies which source supports each premise and what unresolved fact remains. The review result is clearer and more qualified than the first draft, without pretending the case outcome can be guaranteed.
Use reader comprehension as a practical check. A second reviewer should be able to explain the proposition, source, reasoning, and requested action. Disagreement should lead to examination rather than be resolved by asking which draft sounds stronger. This check can reveal unclear logic, but it is an editorial practice, not proof of legal sufficiency or an assurance that the tribunal will agree.
Inspect the full release package and local requirements
Review the version intended for submission and its attachments. Confirm that edits have not broken references, exhibit labels, pinpoints, document identity, or the relationship between the brief and supporting materials. An earlier approved draft is not necessarily the current filing package. The release record should identify the exact version and responsible reviewer.
Check the requirements governing the actual tribunal and matter through authoritative sources. Those can concern form, timing, signatures, certifications, AI disclosure, evidence, or other procedural matters. Requirements can differ and change. This article does not announce a universal AI disclosure requirement for court filings or a rule that AI assistance is always prohibited or always permitted.
The New York courts' AI advisory committee report discusses varied court approaches and recommendations. A report or recommendation is not itself an adopted rule for your filing. Use it as context, then verify the source that actually governs the matter. Do not copy a practice from a different court because a research summary describes it as common.
For the hypothetical package, the last edit changes the factual paragraph but leaves an old exhibit reference in the conclusion. The release reviewer catches it and returns the document for correction. The team then checks the actual procedural sources relevant to the filing. The accepted package has verified source relationships, resolved open issues, and the appropriate release decision, without claiming that this workflow guarantees acceptance by the court.
Limit automation at the release step according to the team's chosen controls and applicable requirements. A tool that drafts and checks a table need not have permission to submit documents or email them externally. Keep the consequential action assigned to the authorized person, with the reviewed version visible. Review records should reflect what was actually checked rather than imply a comprehensive approval from a checkbox.
Train associates with errors that require legal judgment
Give associates a bounded task and a source package, then ask them to explain their review. One useful training example includes a real authority supporting a narrower proposition, a quotation missing a condition, a factual assertion based on an incomplete attachment, and an inconsistent conclusion. The exercise should require source reading and reasoning, not merely recognizing malformed citations.
Ask the associate to create the proposition ledger and mark each disposition. Have the supervising lawyer discuss why a source supports or fails to support the statement. Teach the difference between retrieving a case, understanding its role, checking its treatment, and deciding how it belongs in the argument. Those are separate skills even when one interface offers all the buttons.
Use a manual explanation as a learning check. The associate should be able to state the legal issue and verified reasoning without reciting the generated text. If they cannot explain why an authority matters, further review is needed. The purpose is an independently understood argument, not confidence based on the tool's fluency or the supervisor's reputation.
Include client communication. An associate may need to explain that a requested argument lacks support, that a factual premise needs confirmation, or that the team revised a broad claim. Practice that explanation with concrete sources and respectful language. Clients sometimes bring AI-generated material of their own; the lawyer should investigate the useful question without accepting unsupported conclusions.
Supervision also needs time. A policy telling juniors to verify everything is ineffective if assignments leave no review capacity or if supervisors accept neat drafts without questions. Define the expected checks, escalation route, and release owner. Document how errors changed the training or process. Do not claim that one exercise qualifies a reviewer for every subject matter or future filing.
Use a 30-day review trial to settle the method
In the first week, choose a document family and create an independently reviewed reference package from material you are authorized to use. Define the ledger, source checks, reasoning review, and release record. Confirm the approved tool environment and exclusions. The trial should not begin with an urgent filing whose necessary review the team cannot staff.
In the second week, test draft preparation and source-location assistance on known examples. Record unsupported propositions, omitted facts, quotation defects, and review effort. Examine real citations that fail to support the claim, not only nonexistent cases. Have reviewers explain disagreements and refine the reference answer when necessary.
In the third week, run a supervised trial on approved work. Keep the responsible lawyer's release decision visible. Track the source package and current draft. Pause the workflow if reviewers cannot trace material claims, if the tool substitutes unapproved sources, or if a draft circulates as final without the required review. The pause criteria are the team's operating choices for this task.
In the fourth week, decide which assistance is useful. The tool may help with outlines or locating missing citations while remaining unsuitable for unsupervised argument. Record the boundary, observed errors, reviewer effort, and next review date. A narrower method can be a successful decision when the evidence supports it.
If you know the next review step but keep delaying it, identify why. An uncertain legal issue needs research; reluctance to ask a colleague for help needs a conversation. Choose the action that fits the actual obstacle. AI and the Andes Retreat is a five-day seminar in Peru's Sacred Valley focused on examining behavior and changing action. It does not promise litigation training, legal advice, or a guaranteed professional outcome.
Compare your options
| Work item | Evidence to inspect | Decision or escalation |
|---|---|---|
| Legal proposition | Full authority and relevant context | Supports, narrower support, unsupported, or research pending |
| Current treatment | Relevant history and later reasoning | Assess effect on the actual proposition |
| Quotation | Exact text, omitted context, and attribution | Correct words and meaning before reliance |
| Fact | Record source and degree of certainty | Resolve gaps or qualify the assertion |
| Reasoning | Verified premises and requested relief | Explain the inference independently |
| Release | Current brief, exhibits, and authoritative requirements | Authorized reviewer accepts exact package |
Work through one ai legal brief review decision
Use the worksheet on a historical or approved draft. It is an editorial method for organizing review, not a certification of legal sufficiency or a substitute for governing rules.
A reflection tool for planning your next step, rather than a predictive score.
- Choose one substantive paragraph and list each material assertion.
- Retrieve the relied-on authority independently and identify what it supports.
- Check relevant treatment and quotation context using appropriate research methods.
- Trace factual statements to the record and distinguish allegations from established facts.
- Explain the reasoning without relying on the generated wording.
- Inspect the full release package and actual governing filing requirements.
A hypothetical notice brief cites a real opinion, but the favorable sentence describes a party argument rather than the court's holding. The draft also omits an amendment and quotes a rule without its exception. Counsel marks the proposition unsupported, retrieves the full agreement package, corrects the quotation, and requests a missing transmission record. The argument is rebuilt from verified premises. Acceptance requires a source-backed proposition ledger, resolved material factual gaps, independently understood reasoning, and review of the exact release version. The method makes no prediction about the tribunal's decision.
Your working notes
Use this page while you think. Your notes stay in this browser tab; download them before you close it.
What does this question change for you?
A practical answer can change how you work. It can also leave a personal decision unresolved: what to learn, what to stop postponing, or which responsibility to take next. Name that decision before collecting another prediction.
Work through your next decision. Your notes stay in the browser tab, and the exercise gives you no predictive score.
If a recurring pattern keeps preventing action, AI and the Andes Retreat offers five days of practical inquiry in Peru's Sacred Valley, followed by support and accountability. Read about the seminar and assess whether that work fits your question.
Frequently asked questions
Search research and audience discussions informed these questions. Some wording has been adapted to the task covered by this guide.
Is AI reliable for citations?
Do not infer reliability from a plausible citation or resolving link. Retrieve the authority independently, inspect what it supports, and evaluate relevant history or treatment. A real case can still be mischaracterized. The useful question is whether the particular proposition has support and whether the reviewer understands its role in the actual matter.
Can I use AI for legal research?
AI can assist with leads, outlines, and source organization when the use is appropriate. Verify relied-on sources and reasoning through a competent review method. Assess confidentiality and governing requirements separately. ABA Formal Opinion 512 discusses these responsibilities under the ABA Model Rules; it does not certify a particular research product or replace the rules governing your representation.
Can a lawyer use AI to write a brief?
AI assistance does not settle whether the particular use and filing are appropriate. Counsel needs to assess governing rules, information handling, source verification, factual support, and the actual tribunal's requirements. The workflow here treats generated language as preparation and keeps the understood argument and exact release package with the responsible reviewer.
Can AI verify an AI-assisted brief for me?
It may help locate uncited sentences or organize a source table, but a second generated answer is not independent assurance. Counsel should inspect the authority, record, quotation context, and reasoning. A verification feature needs its own task-specific evaluation. Do not treat a completed checklist as proof that the argument is legally sufficient.
How should I look for authority an AI research answer omitted?
Identify the legal issue and important distinctions independently. Research relevant contrary material and omissions through an appropriate method, then inspect any leads the assistant supplies. Its silence is not evidence that no contrary authority exists. Counsel must assess the applicable duty concerning contrary authority in the actual matter; the page does not supply a universal disclosure rule.
How should I compare AI research with my existing search method?
Use the same defined issue and compare verified, accepted results. Record relevant authorities located, omissions, unsupported propositions, source retrieval, and review effort. A convenient conversational interface can be useful without proving better research coverage. Keep the issue's distinctions visible in both methods and avoid comparing a difficult search with an easy demonstration.
How do I review an AI memo in a subject I do not know well?
Recognize the expertise gap and obtain suitable help or research before relying on the conclusion. Source existence does not establish that you understand the legal question or relevant omissions. Give the specialist the sources, facts, proposed propositions, and open issues. Do not use the tool's confidence to substitute for the knowledge needed to assess the work.
Why can a fluent AI brief still fail as legal writing?
Fluency can connect sentences without connecting supported premises. Inspect whether the legal standard, relevant facts, distinctions, and requested result fit together. A draft can quote accurately yet draw an unsupported inference. Read an outline stripped of rhetoric and ask a reviewer to explain the logic. Source checks and reasoning review are different tasks.
How should counsel handle a client's AI-generated legal argument?
Ask for the question, facts, and underlying sources. Identify any useful issue, then show concretely where the output assumes an unsupported fact or misstates authority. Explain what counsel can responsibly pursue and what evidence is missing. Respectful source-based discussion is more useful than debating the model's reputation or accepting its confidence.
How can I make AI brief review more deliberate?
Assign specific objects of review: propositions, quotations, record facts, reasoning, and the release package. Use the ledger to mark unresolved items and their owners. A single instruction to read carefully can leave important checks unassigned. Give reviewers time, source access, and an escalation path so the process supports the attention it requires.
Who should check current treatment in a team-reviewed brief?
Assign the check to a person with the appropriate capability and give the responsible lawyer an inspectable record. Role labels alone do not establish competence or completion. The release reviewer should understand the result and unresolved questions. The actual supervision duties depend on governing rules and circumstances; this workflow recommends explicit assignment rather than assuming somebody else performed it.
Does having several lawyers on a brief ensure it was verified?
No. Several names do not show which source, factual, or reasoning checks occurred. Assign review responsibilities and inspect the resulting record. The responsible release decision should identify the current package and material open issues. A shared assumption that another lawyer checked the citations can leave the task undone.
Can AI used to locate exhibits trigger a court disclosure question?
The answer depends on the actual tribunal's authoritative rule and its scope. Verify the current requirement and the activity it covers. Do not extend a requirement from another court or treat a advisory report as the governing rule. This page supplies a release-check method, not a nationwide AI disclosure rule for every software feature.
What are AI legal hallucinations?
The term commonly describes generated content with no adequate basis, such as a nonexistent authority, invented quotation, or unsupported claim. For review, identify the actual defect rather than rely on the label. A real authority attached to the wrong proposition is also a material problem, even if it differs from fabrication.
Should I assume current legal AI has eliminated hallucinations?
No such assurance follows from this guide or a product's connection to source databases. Evaluate the actual task and tool configuration, and inspect relied-on output. The year in a query does not prove that a known failure mode disappeared. Avoid claiming a current universal error percentage from an older study or an audience comment.
How should I choose a legal research AI platform?
Test a defined issue set and compare verified source support, relevant coverage, omissions, current-treatment workflow, information handling, and reviewer effort. This guide does not rank current vendors. A platform may be useful for finding leads while requiring a different method for checking the relied-on authority and argument.
How should I organize a brief before asking AI to draft?
Define requested relief, procedural context, legal issues, verified facts, and the authoritative source package. Create an outline showing the propositions the argument must support. Mark missing evidence. That preparation reduces the chance that the assistant writes persuasively around an incorrect premise, although the resulting draft still needs source and reasoning review.
Which rules govern my AI-assisted brief?
Identify the professional rules, tribunal requirements, orders, client instructions, and other obligations applicable to the actual matter. Verify them through current authoritative sources. ABA Formal Opinion 512 interprets Model Rules; it is not binding nationwide law. A recommendation or proposal must remain labeled with its status.
Should I still use a citator for an AI-assisted brief?
Use an appropriate method to investigate relevant history and treatment; a connected AI answer does not automatically complete that work. Read the treatment in context and determine its effect on the proposition. A flag is an entry point to analysis, not the entire conclusion. Record the check and remaining issue for the responsible reviewer.
Does a legal AI tool connected to legal databases remove the need to verify its output?
Source access does not establish that the system selected the right passage or correctly characterized it. Retrieve the relied-on source, read relevant context, and inspect the proposition. The worked example uses a real opinion whose favorable passage described a party argument rather than the court's holding. A case-existence check would miss that defect.
Sources and evidence
These guides were drafted with AI assistance and checked against the linked sources. Examples labeled hypothetical illustrate a decision; they are not participant stories or measured results.
Studies answer specific questions about particular people, tasks, and tools. Their dates and limits matter. A result about task assistance is not a forecast for an entire profession.
- ABA Formal Opinion 512: Generative Artificial Intelligence Tools
- New York Advisory Committee on Judicial Ethics Opinion 26-37
- Advisory Committee on Artificial Intelligence and the Courts: 2025 Annual Report
Where the audience questions came from
These linked discussions and comments supplied questions and wording. They do not establish the truth of the replies, verified professional identities, or how common a concern is.
- Public audience question: How should I look for authority an AI research answer omitted?
- Public audience question: How do I review an AI memo in a subject I do not know well?
- Public audience question: Why can a fluent AI brief still fail as legal writing?
- Public audience question: How should counsel handle a client's AI-generated legal argument?
- Public audience question: How can I make AI brief review more deliberate?
- Public audience question: Who should check current treatment in a team-reviewed brief?
- Public audience question: Does having several lawyers on a brief ensure it was verified?
- Public audience question: Can AI used to locate exhibits trigger a court disclosure question?
- Public audience question: Should I still use a citator for an AI-assisted brief?
- Public audience question: Does a legal AI tool connected to legal databases remove the need to verify its output?