Research assist patterns

This lesson is educational and is not legal advice. Legal research requirements vary by jurisdiction, forum, date, and matter. A qualified legal professional should supervise research used for an actual decision.

“Research electronic notice” is a topic, not a research question. A useful question names the jurisdiction, legal issue, key facts, time frame, forum, and desired decision. It also identifies what remains unknown.

A structured question might be: “Under the law governing this agreement, and using authority current through the review date, what rules affect whether notice sent to the specified email address satisfies section 14 when the recipient disputes receipt?” That question still requires refinement, but it exposes the contract section, factual dispute, and currency requirement.

AI can help turn an initial problem description into a question tree:

• What source of law might control?
• Which facts trigger or defeat the rule?
• Are definitions, exceptions, or procedural rules relevant?
• What jurisdiction and date limits apply?
• What terms and synonyms should a researcher search?
• What contrary authority should be tested?

The output is a research plan, not an answer.

Legal research depends on authority, not merely relevance. A plausible blog post and a controlling appellate decision are not interchangeable. Define the source hierarchy before asking for synthesis. Depending on the task, it may include constitutions, statutes, regulations, binding cases, court rules, agency materials, contractual text, official guidance, and carefully selected secondary sources.

Ask AI to work only from sources you can identify and retrieve. If the system has a verified research database, keep the retrieval record. If you provide documents, require the output to point to page, paragraph, section, or other stable location. Do not treat a generated citation as evidence that a source exists.

A useful evidence table contains:

1. proposition;
2. source name and stable identifier;
3. authority level and jurisdiction;
4. exact supporting quotation;
5. treatment or currency status checked by a human;
6. facts that connect the source to the question;
7. contrary or limiting authority;
8. unresolved issue.

AI may draft this structure. The researcher verifies every field in the authoritative system.

In the discovery pass, use AI to suggest vocabulary, issue branches, source types, and candidate authorities. This pass values coverage and can tolerate false leads because no candidate is relied upon yet.

In the validation pass, retrieve each candidate independently. Confirm that the source exists, says what the draft claims, applies in the relevant jurisdiction, remains good law or current guidance, and supports the proposition in context. Read enough surrounding material to detect qualifications. A quotation can be exact yet misleading if an exception appears in the next paragraph.

In the synthesis pass, provide only validated materials and ask for a comparison organized by proposition, authority, factual fit, and conflict. Require citations back to supplied sources. Then compare the synthesis against the originals. This separation prevents the model’s memory-like generation from being mistaken for retrieval.

Avoid “Find the law and give me the answer.” A better instruction is:

Create a research plan for the issue below. Do not state a legal conclusion.
Identify sub-issues, missing facts, search terms, likely primary-source types,
jurisdiction and date questions, possible exceptions, and contrary positions.
For any candidate authority, label it UNVERIFIED until a human retrieves it.
End with stopping conditions and questions for supervising counsel.

Later, after sources are verified:

Using only the supplied sources, map each proposition to an exact quotation
and location. Distinguish binding, persuasive, and secondary authority.
Identify conflicts, factual differences, exceptions, and unsupported gaps.
Do not add sources or infer current validity.

These prompts make limitations visible. They do not make the model a legal researcher of record.

Save the research question, source list, search date, databases used, search terms, filters, and reviewer. Record whether the model saw full documents or excerpts. If an excerpt omits footnotes or attachments, the synthesis should say so.

Scope matters. “No contrary authority found” is only meaningful when the search process is documented. It should never become “there is no contrary authority.” Likewise, a model’s statement that a search is comprehensive cannot prove completeness.

Use explicit stopping rules. Stop and escalate when jurisdictions are unclear, authority conflicts, sources cannot be retrieved, currency cannot be checked, the issue requires specialized expertise, or the answer will support a consequential filing or decision. The appropriate output may be a list of open questions.