AI-assisted legal notes: trace each claim to its source

An AI-assisted issue note is useful only if a reviewer can follow each material statement back to the supplied record. A polished summary can conceal a missing qualification, an uncertain transcription or a conclusion the documents do not support. Start with a small, attributable note and keep unresolved questions visible.

This is general information about source traceability, with a fictional exercise. It is not legal advice or a determination of privilege, confidentiality, compliance, admissibility or filing requirements. No lawyer has approved this guide, and no legal AI system, redaction process or security configuration was tested. For a real matter, a qualified responsible lawyer must review and authorize the proposed use, with ethics, privacy and security support where needed.

Establish the jurisdiction and permitted use

ABA Formal Opinion 512 discusses generative AI under the ABA Model Rules of Professional Conduct, including competence, confidentiality and review of outputs. It is a U.S. model-rule analysis; the rules adopted in the relevant jurisdiction govern the lawyer's obligations. It does not approve a particular notebook or make every local AI workflow permissible.

For England and Wales, the SRA's warning notice on misuse of AI, published 17 August 2026, addresses firms and individuals it regulates. It emphasizes continuing professional responsibility and appropriate safeguards for client information, whether a tool is free or paid. These two sources do not establish rules for every country, court or practice setting.

Before using real material, the responsible lawyer should resolve applicable professional duties, client instructions, any necessary consent, protective orders, permitted recipients and retention or preservation requirements. Internal analysis, client communication and court filing are different proposed uses. Approval for one does not settle the others. If authority or suitability is unresolved, use invented material for the exercise.

Preserve the source and record transformations

Keep the original in the approved records system. Work from a separately identified derivative when extraction, OCR or minimization is authorized. Record source identifiers, versions and stable locations such as exhibit page and paragraph. A generated heading or an AI-created paragraph number is not an original citation.

An OCR copy should point back to the relevant page image. Mark uncertain characters and missing pages; do not quietly repair a date or quantity because one reading seems plausible. If a derivative uses different pagination, retain a location map. This lets a reviewer distinguish an error in the source from an extraction error or an unsupported summary.

Keep factual extraction and legal analysis in separate fields. A source may support “the witness reported eleven crates” without supporting “eleven crates were delivered.” A reviewer should not have to guess which type of statement the note contains.

Map the processing copies before adding AI

Inventory the original, extracted text, chunks, embeddings, prompts, outputs, logs, caches and backups. Identify where each copy is processed and stored and who can access it. Running inference locally does not establish that every extraction service, plugin, index or backup is local.

An index is not necessarily just keywords. For example, Chroma's collection documentation supports storing documents, embeddings and metadata together. That is a reason to inspect the actual configuration, not a recommendation to put case material into Chroma. Retention and access must be established through the system's controls and applicable arrangements; asking a model not to save a prompt does not configure them.

Redaction also needs its own review. spaCy's named-entity documentation describes statistical recognition of entities. Detection can assist an authorized reviewer, but a limited list of entity types cannot establish that every identifying or sensitive fact is removed. Distinctive events and combinations of details can remain revealing after names are replaced.

Inspect the complete proposed outgoing material and its context. For PDFs, Adobe distinguishes visible redaction from sanitizing hidden data, including metadata and embedded content. A visual spot-check or a black rectangle is not evidence that all underlying information is removed. Preserve required originals; this guide does not authorize altering evidence, deleting records or uploading a derivative.

Treat retrieval as assistance, not validation

Retrieval can supply relevant passages while the generated answer still misstates them or attaches an unsupported citation. The empirical study Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools found errors in the RAG-based products it examined. Its particular evaluation is evidence against a blanket accuracy guarantee, not a current ranking of every product or a measured error rate for your workflow.

Open each cited source. Check that it exists, that the location is correct and that it supports the precise claim, including any negation or qualification. For legal authorities, the responsible lawyer must also assess relevance and applicability. A clickable citation alone completes none of those checks.

Fictional exercise: the crate-count issue note

DEMO-L04 is entirely invented. It contains two records and no contract, receipt or legal authority:

EX-A, interview paragraph 2: “I counted eleven crates at bay two. I did not see the unloading begin.”

EX-B, packing list row 4: “Crates dispatched: twelve.” No receipt signature is supplied.

A proposed source-only extraction request is: “Use only EX-A and EX-B. Attribute each statement and include its location. Preserve qualifications, conflicts and unknowns. Separate suggested questions from observations. Do not invent a receipt, contract clause, legal citation, owner or deadline. Do not send messages or take action.” This is an instruction for a draft, not an enforced access or security boundary.

The answer key should retain both observations: EX-A reports a count of eleven at bay two and says the witness did not see unloading begin; EX-B records twelve dispatched. The actual number delivered is unresolved. The records do not establish when the count occurred relative to unloading or whether crates were moved before the count.

Reject “the supplier delivered one crate short.” It converts two differently situated records into a delivery finding. Also reject an invented breach-of-contract conclusion: no contractual term is supplied. A possible next question, subject to the lawyer's review, is whether an authorized search can locate a receipt or clarify the sequence of unloading and counting. Owner and deadline remain unassigned.

Reusable note and proposed checks

For a new draft, record the matter's approved identifier and purpose; source versions and locations; transformations and OCR uncertainty; attributed observations; supported paraphrases; conflicts and unknowns; separate lawyer analysis; handling classification, privilege assessment and actual permissions; reviewer and review date; retained copies; and the next authorized use. Leave absent information explicitly unresolved.

Try the format with disposable fictional inputs before considering real material. Include a negated statement, conflicting quantities, a missing page and an irrelevant passage. A reviewer should be able to locate every supporting passage and detect a changed count, omitted qualification or invented citation. These are proposed exercises, not completed tests or evidence of suitability for client work.

Access needs a separate technical assessment: a prompt saying “use this matter only” does not prove that another matter cannot be retrieved. Retention, preservation holds and disposal decisions belong with the responsible lawyer and approved records process. Do not substitute an arbitrary deletion schedule or a cache-clearing instruction.

Moyan's AI Tool Lab can help organize tool discovery and questions. Its checking workspaces do not themselves call a model. A discovery result does not approve a processor for client information or establish a local encrypted legal vault. Keep this exercise fictional until the responsible people have approved the actual workflow.

Correction history and limits

Source documentation was checked on 19 October 2026. This correction replaces broad secure-vault assurances with a source-traceability example and explicit review boundaries. The article's original publication and update dates remain unchanged; the source-check date is not a deployment date. The example is original and fictional. No hands-on security test, professional approval or guarantee of confidentiality, privilege, compliance, redaction safety or RAG accuracy is claimed.