AI for Attorney Research: A Citation-Checking Workflow

A lawyer can use AI to generate research questions and candidate authorities, then catch fabricated or misquoted citations by checking each source in a legal database, reading the cited passage in context, running a current citator check and recording a lawyer's sign-off before filing. A linked citation is a lead, not proof that the proposition is right.

Split the job into discovery and verification. The tool may suggest a case, statute or quotation; the lawyer must establish that the authority exists, applies to the right jurisdiction and time, supports the exact sentence and remains good law. That separation is especially important when a fluent answer presents several real cases beside one invented one.

Follow me on Instagram@sagnikteaches

Write the proposition before asking for cases

Begin with a single, narrow research question. “Find all law on dismissal” is too broad to check. Instead, write the proposed proposition, relevant jurisdiction, date, procedural stage and disputed facts. A research memo should say whether the task is to find binding authority, identify persuasive authority, test an opponent's proposition or locate a statutory provision. The better the question, the easier it is to reject a plausible but irrelevant result.

Connect on LinkedInSagnik Bhattacharya

Take a made-up contract dispute. The issue might be: “Does an email exchange satisfy the contractual requirement for a signed variation where the agreement also requires a named officer's approval?” That question has at least two gates: the signature issue and the officer-approval clause. A model may find material on electronic signatures but ignore the separate approval requirement. Record both gates in the research brief so the lawyer can see the omission. Do not paste confidential agreement text into an unapproved tool merely to save a few minutes of summarising.

Subscribe on YouTube@sagnikteaches

If a general assistant such as ChatGPT is used at all, ask it for search concepts and possible counterarguments rather than a final list of supposedly authoritative cases. Its output may help you vary terms such as “variation”, “amendment” and “electronic signature”. It does not replace a subscription database, official legislation source or court record. For tool choice across a wider practice, the attorney tool comparison addresses research alongside drafting and intake; this workflow stays with citation control.

Capture suggestions in a citation ledger

Do not let a research answer flow directly into a filing draft. Copy each proposed authority into a ledger with separate columns for the citation as supplied, source found, jurisdiction, date, proposition, exact passage, citator status, later treatment and reviewer. Give each row a status: unlocated, located, passage checked, treatment checked or approved. An empty cell is not an implicit pass. A spreadsheet or matter-management table works; the key is that the same row follows the authority from suggestion to final sentence.

For example, an illustrative row might contain: “Case A, 2024, paragraph 42; AI says emails varied the contract; source located; paragraph 42 discusses only notice, not variation; reject.” Another might say: “Case B, 2022, paragraph 18; passage supports electronic signature in a different contract type; further analysis required.” These are invented labels for workflow demonstration, not legal citations. They show why a binary “case exists” tick is too weak. The proposition column must be written in the lawyer's own words, not copied from the AI answer.

Require page or paragraph pinpoint references where available. If the model gives a case name without a pinpoint, mark the proposition unverified until a lawyer finds the passage. If it gives only a web summary, find the primary source. Keep the version or date of the decision, statute or regulation. An amendment can make a once-accurate proposition stale, while a later appeal can alter how a case is used. The ledger makes these time questions visible rather than burying them in a footnote.

Locate the primary source and read around the quote

Search the full citation in the firm's trusted legal research service or an official source. Confirm party names, court, year and document. Search for the exact quotation, then read the paragraphs before and after it. Ask whether the cited words are the court's holding, a party's submission, a quotation from an earlier case or a description of an argument that the court rejected. This is the point at which many polished AI answers fail: they attach the right citation to the wrong legal point.

A practical rule is to record one sentence of context beside every quotation. If a draft says “the court held that oral modifications always bind”, the context note may reveal “the court held that this particular waiver was effective despite a no-oral-modification clause”. The second statement is narrower and may depend on facts absent from the current case. Change the draft proposition, find a better authority or drop the citation. Do not keep the broad sentence because the linked decision is real.

When a quotation is material, compare punctuation and omitted words against the source, including any ellipsis. For a statutory provision, check the version effective at the relevant time and read definitions, exceptions and cross-references. For a procedural rule, check current text and any local application before using an AI summary. These are legal judgments, not text-matching chores. A paralegal can perform the initial locating and comparison, but the responsible lawyer must approve the final proposition.

Thomson Reuters says CoCounsel Legal is grounded in Westlaw and Practical Law content and returns linked citations for verification. LexisNexis says Lexis+ with Protégé (the product formerly sold as Lexis+ AI) builds in Shepard's citation validation, and its newer Shepard's Verify checks citations in AI responses, documents and drafts and flags unsupported statements. Both can reduce the effort needed to get to a source, but neither removes the need to inspect the passage and the matter-specific fit. A general assistant's link may point to an unofficial summary or an unrelated page. Treat any broken or ambiguous link as an unresolved ledger item.

Run a citator check after the passage check

A citator tracks how later authorities treat a case or provision. Use the firm's current citator, such as KeyCite within Westlaw or Shepard's within LexisNexis, after verifying that the quoted passage supports the sentence. Record the date and result. A case can be correctly quoted yet no longer be good authority for that point. Conversely, a caution signal may concern a different issue; the lawyer needs to read the later decision rather than treating every signal as an automatic rejection.

Suppose an illustrative case supports a notice proposition in 2019, but a later appellate decision limits it in 2025. The ledger row should state exactly what was limited and whether the draft still fits. A red or warning indicator alone cannot answer that. The reviewer should open the relevant citing decision, identify the affected proposition and adjust the draft. If the research tool generated the answer before the later decision or searched a narrower collection, this step is where that gap appears.

Check statutory currency separately. A citator signal on a case does not prove that the statutory text it interprets remains unchanged. In a filing with ten authorities, give every case its own treatment result and every statute its effective-date check. Do not use “all citations checked” as a single checkbox without a row-level record. The tax research citation tutorial uses a similar source discipline for a different body of authority, but court filings need their own procedural sign-off.

Try the workflow on a ten-citation draft

Imagine a two-lawyer firm preparing a short motion with ten AI-suggested authorities. These figures are illustrative. An assistant spends 25 minutes turning the issue outline into search terms and a candidate list. A paralegal spends 70 minutes locating sources and comparing passages. Three citations fail: one case cannot be found, one real case discusses the opponent's argument rather than the court's conclusion, and one is from the wrong jurisdiction. The lawyer spends 50 minutes on treatment and fit, and a further 20 minutes checking the final draft after edits. Total review time is 165 minutes, or two hours and 45 minutes.

That does not establish a saving against conventional research. It establishes a control. If the firm previously spent three hours researching without a ledger, the apparent time difference is only 15 minutes and may disappear on a harder matter. The benefit to test is whether the team can identify every unsupported proposition before the document leaves the firm. If three of ten suggestions are rejected, sending the draft after a superficial link check would be plainly unsafe. Record how often the model proposes a usable authority, not merely how quickly it generates ten names.

The three failed rows lead to different actions. For the unlocated case, search names and citation variants once in a trusted database, then reject it if still absent. For the misquoted case, rewrite the sentence or find another authority. For the wrong-jurisdiction case, decide whether it is persuasive and whether it should appear at all. None should be silently replaced by a new AI citation. Replacement starts a fresh ledger row and repeats all checks. This prevents a last-minute substitution from bypassing the process.

Use prompts that expose uncertainty rather than conceal it

Ask the research tool to separate primary authority, secondary commentary and hypotheses. Require it to say when it cannot identify a source. A prompt cannot guarantee truth, but it can reduce the temptation to treat a continuous narrative as a checked memo. In a specialist service, request links and pinpoints; in a general assistant, request search terms and competing interpretations. Keep confidential material out unless the firm's vendor and client-data review allows the use.

Research issue: [narrow proposition and relevant dates].
List candidate primary authorities separately from commentary.
For each candidate, give the exact proposition it might support,
a pinpoint if available, and a reason it might not apply.
Do not invent citations or fill gaps from memory.
Mark anything you cannot verify as UNVERIFIED.
I will check every source and its later treatment independently.

A realistic illustrative output might say, “Case C may support the signature point, but I have not confirmed whether the officer-approval clause was present.” Keep that caveat. A less helpful output says, “Case C conclusively establishes that emails are sufficient.” The lawyer should not merely soften the adverb. Open the case, inspect the clause and decide whether the analogy is sound. The prompt is a way to surface uncertainty; the ledger and source review decide what may be cited.

For a search with no directly applicable case, record that outcome rather than forcing an authority into the draft. An honest memo may explain the distinction between available cases and the client's facts. AI often fills that gap with an attractive analogy. The lawyer decides whether an analogy is persuasive and how candidly to describe its limits. The marketing content workflow is a different setting, but the shared habit is to separate drafting from approval.

Test for omissions as well as false citations

A verification process that only catches invented cases can still miss the authority most damaging to the proposed argument. After the initial search, ask a lawyer to run a separate adverse-authority search using the opposite proposition and the opponent's likely terminology. Check whether a binding case, statutory exception or later amendment changes the conclusion. Record the search terms and databases used. This is distinct from asking the model to “find counterarguments”, because it gives the team an independent route to material the model may have overlooked.

Return to the email-variation question. Suppose an AI answer identifies four decisions about electronic signatures. An independent search for “approval by designated officer” finds a fifth decision where the form of signature was accepted but the internal-authority condition was not met. The fifth case is the most useful for deciding how to draft the motion, even though it is less favourable. The team should put it in the ledger, check its passage and treatment, and decide whether the argument needs a factual distinction or a narrower claim. A research result with only favourable cases is a warning sign, not a sign of efficiency.

Also test source coverage. A platform may search its own content, uploaded documents or the web, and those collections have different strengths. Record which collection answered each query. If the question turns on an enacted rule, read the official text; if it turns on court interpretation, use the relevant case database; if it turns on a fact in the matter, verify the underlying document. A single summary blending these layers can appear complete while one layer is absent. The lawyer should decide when enough independent search has been done for the stakes of the filing.

Set a stopping rule for the pilot: no filing use if the team cannot open primary sources, cannot determine current treatment, or cannot reproduce why a proposition passed. A specialist AI service may make these steps easier through linked material, but a trial should be judged by whether reviewers actually complete them. If reviewers routinely skip the ledger because the answer looks polished, reduce the number of candidate citations and force a smaller, clearer output. Better five checked authorities than twenty attractive references with uncertain status.

Check the final document, not only the research memo

After the lawyer edits the argument, verify that each citation still supports the sentence that now precedes it. A correctly checked authority can become misplaced when paragraphs are rearranged. Compare the final filing against the ledger: every cited item should have a matching approved row; every quotation should match the source; every pinpoint should take the reader to the right passage. Run the citator again if a significant period has elapsed or a new development is known. Have the responsible lawyer sign off on the actual version to be filed.

Use a final reading order that catches different failures: first read only the propositions and footnotes, then only quotations, then the whole argument. The first pass asks whether citations attach to the right claims. The second catches text errors. The third tests logic and fair presentation. A second reviewer is valuable for a high-stakes filing because the drafter may recognise what they intended rather than what the sentence says. A tool-generated verification report can assist the review, but it is not the sign-off.

An illustrative final-document failure is a footnote shifted during editing: a case checked for notice now follows a sentence about damages. The ledger shows the approved proposition and exposes the mismatch in seconds. Another is a direct quotation with “may” changed to “must”; a text comparison reveals the change, but only the lawyer can decide its significance. A third is a new authority pasted in from an email after the ledger was closed. The release rule is simple: no approved row, no filing citation.

Keep the ledger, source links, review notes and final version with the matter record under the firm's retention policy. Restrict access to the matter team. If the platform retains prompts or uploads, assess that before including confidential facts. For the choice between a general assistant and a legal research product, see general models compared with CoCounsel; for this workflow, judge the tool by verified authorities and reviewer effort rather than the fluency of its answer.

A monthly quality review can use a small sample of closed matters. Count candidate authorities, those located, those supporting the claimed proposition, those surviving the citator check and those ultimately cited. Compare the number of corrections found before filing with those found at final review. If the same misquotation pattern appears twice, update the research prompt and reviewer training. Keep the rejected examples in an internal teaching set with client details removed. That gives new staff concrete examples of how a real citation can still be a bad citation.

Questions about checking AI-suggested authorities

Does a clickable citation prove that the quoted proposition is correct?

No. A link may lead to a real decision that says something narrower, addresses a different procedural posture or has a quotation taken out of context. Open the primary source, find the passage, read surrounding reasoning and compare the exact proposition. Then run a current citator check. Save a record of those steps before using the authority.

Can a small firm use ChatGPT for the first research pass?

It can help frame questions, generate search terms and organise a non-confidential research plan. Do not treat its citations or legal conclusions as verified, and do not submit client material until the firm's confidentiality and tool terms have been reviewed. Every proposed authority still needs primary-source and citator checks in a trusted legal research service.

What if an AI answer cites a case that cannot be found?

Mark it unverified, search the citation and party names in an authoritative database, and do not include it in a draft unless a lawyer independently finds the source. Ask the tool to identify the passage it meant, but do not let a replacement citation inherit the original claim without a fresh check. Keep the failed item in the research log so another drafter does not reuse it.

Further reads

• CoCounsel or Lexis+ With Protégé? Legal Research AI for a Small Firm — Compare the two research environments before a trial.
• How to Check Sources and Citations in AI Research — Apply source checks to non-legal research too.
• AI App for Lawyers: Score Accuracy, Privacy and Cost in a Short Trial — Score accuracy and privacy during a tool pilot.
• AI and Attorney Fees: Can You Bill Clients for AI-Assisted Work? — Consider how review time is recorded and billed.
• AI Forensic Accounting for Lawyers: Can It Trace Transfers? — See the separate checks needed for financial evidence.
• CoCounsel Pricing: What a Small Firm Should Ask Before Buying — What Thomson Reuters publishes about CoCounsel pricing for small firms, what it doesn't, and eleven questions plus an email to get a real quote.
• Lexis+ With Protégé Pricing: Online Plans for Firms of One to Three — What a one-to-three-lawyer firm can learn about Lexis+ with Protégé pricing online, plus the renewal terms and extra charges to check.
• AI Discovery Review for Defence Firms: Index, Search and Cross-Check — Index discovery, map each statement to the disputed issues, build a contradiction table with page cites, then verify every cite by hand before filing.
• AI Legal Brief Writing: A Draft and Cite-Check Routine — A routine for small litigation firms: build an authority table, draft only from it, cite-check each source line by line, and meet court AI-disclosure rules.
• AI for Barristers: A Chambers Checklist for Casework and BSB Rules — Turn BSB guidance into a practical chambers policy and a per-matter check before any AI-assisted work is used.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Thomson Reuters CoCounsel Legal and Westlaw Edge product pages; LexisNexis Lexis+ with Protégé product page (Shepard's citations and Shepard's Verify); ABA Model Rules on competence and confidentiality. Confirm current procedural and court requirements for the matter.