NotebookLM for Lawyers: Synthesis, Not Free Jazz (Now Gemini Notebook)

Ask ChatGPT which of two witnesses is lying about the date of a phone call and it answers from a vague recollection of every deposition ever posted online. Ask NotebookLM the same question and it knows nothing except the two transcripts you uploaded, answers from those, and pins a clickable citation to every sentence. That difference is the whole case for NotebookLM for lawyers, and it is why Ernie Svenson’s line has become the standard description: “NotebookLM is built for synthesis, not free jazz” (Attorney at Work).

Three things have changed since most lawyers last looked. Google renamed the product Gemini Notebook on 16 July 2026. The source caps are now 50, 300 and 500 to 600 per notebook. And Google has published, in plain words, what happens to your uploads on a personal account versus a Workspace account, which settles the confidentiality question most earlier write-ups hedged.

Why a closed universe suits lawyers

Most hallucinations come from a model answering out of its parameters, the “lossy zip file of internet knowledge” in Andrej Karpathy’s phrase. Grounding the model in supplied documents is the cheapest fix, and the benchmarks agree: on Vals AI’s Document Q&A task the best tool scored 94.8% against a lawyer baseline of 70.1%. NotebookLM is that idea sold as a product. It will not cite a case you did not upload, because it has nothing else to draw on.

Benjamin de Seingalt, a one-person legal department in Ohio who uses Gemini, NotebookLM and GC AI, describes the role precisely: “It’s a gut check” (Bloomberg Law).

Two cautions. Grounding reduces invention; it does not abolish error. Stanford’s 2024 study found retrieval-grounded legal tools still hallucinating on more than 17% (Lexis+ AI) to more than 34% (Westlaw AI-Assisted Research) of queries, often because retrieval pulled the wrong passage (Stanford HAI). A citation that exists but does not support the sentence is the failure mode to watch, which is why the workflow below ends with spot-checking. And more sources are not better: Chroma’s context-rot research found accuracy falls as input grows. One notebook per matter, holding only the documents that matter.

The rename to Gemini Notebook (16 July 2026): what changed and what did not

“Today, we’re renaming NotebookLM to Gemini Notebook,” Google wrote on 16 July 2026 (Google). The notebooklm.google domain still resolves and nothing about the source caps or citation mechanics changed. What the rename signals is where the product sits: inside the Gemini family, so a Workspace user’s notebooks fall under Google Cloud terms and the Cloud Data Processing Addendum, a personal user’s under consumer Google terms. The Gemini for lawyers guide covers the rest of the family.

Source caps: 50, 300, 500 to 600, and what fits in them

Google’s help page gives the limits as of September 2026 (Google Help); older lawyer-facing guides still quote 200, which is stale.

A source is a document, not a page, so a 400-page deposition and a two-page letter each count once. The constraint is not the cap but the discipline: a notebook stuffed with every document in the DMS answers worse than one holding the thirty that matter, for the reasons in the context window guide.

Is NotebookLM confidential? Personal accounts, Workspace accounts and the compliance gap

Google’s current pages answer this precisely, and the answer turns on the account you are signed into (Gemini Notebook privacy page; Workspace privacy hub).

The compliance gap decides the edge cases. Gemini Notebook “does not support ISO, SOC, or FedRAMP compliance and is not covered by the Google Business Associate Agreement”, so medical records, protective-order material and matters whose outside-counsel guidelines demand SOC 2 belong on a platform with the paperwork. For an ordinary commercial record on a Workspace account, NotebookLM is on stronger contractual ground than a consumer chatbot, as the ChatGPT confidentiality guide shows by comparison.

The trial-prep workflow: upload everything, one-page brief, timeline, cross-exam, spot-check

Svenson’s workflow is the one to copy, because a practising lawyer wrote it. Create a notebook for the matter and upload every source: pleadings, depositions, exhibits, expert reports, the authorities you have verified. Ask for a one-page brief “that lays out the key issues, the players and the disputed facts, each one nailed to a specific passage”. Build the timeline. Generate the cross-examination outlines and issue maps. Then, in his words, “the most important step: you spot-check.”

Spot-checking in NotebookLM means clicking. Every sentence carries a numbered citation; click it and the source passage opens beside the answer. If the passage says what the sentence says, tick it; if not, you have found the misgrounding a citation checker would have missed.

The case chronology guide and the deposition preparation guide take the same workflow further; the deal-side equivalent, a notebook per data room, is in the due diligence workflow.

Five ready prompts for litigation

These expand Attorney at Work’s published set: the short original opens each prompt in quotation marks; the additions enforce the citation and no-invention rules a partner would insist on.

"Create a one-page case brief: parties, posture, issues, key facts, and relief sought." Use only the sources in this notebook. Every sentence must include a linked citation to a specific passage; omit any sentence you cannot cite. Where sources disagree on a fact, give both versions with citations and label it DISPUTED. Where no source covers an element, write NOT IN SOURCES. Neutral tone, maximum 400 words.

"Build a dated timeline of events with people, documents, and significant notes. Flag contradictions and missing links." Format as a table: Date | Event | People | Source and page | Note. Mark inferred dates INFERRED and give the basis. After the table, list (a) every point where two sources give different dates or accounts of the same event, with both citations, and (b) every gap of more than [30] days in which the sources record nothing.

"List the top three themes supporting liability and the top three against it, with the strongest citations for each." For each theme give a one-sentence statement, the three best supporting passages quoted verbatim with citations, the single passage that most undermines it, and the witness or document I would need to shore it up. Rank by how much a neutral reader would be moved, not by frequency. Use only the notebook sources.

"Compare Expert A and Expert B on [topic]. Summarize agreements, conflicts, and methodological weaknesses with page cites." Structure: (1) points on which both agree, with a page cite to each report; (2) direct conflicts, in a table with each position quoted; (3) each expert's stated assumptions and any the other does not share; (4) methodological weaknesses visible from the reports, cited to the page. Do not say who is right; use nothing outside the two reports.

"Draft a cross-exam outline for [Witness] limited to inconsistencies across [Docs A/B/C], grouped by topic, each point with a citation." For each point give the earlier statement (quoted, cited), the later statement (quoted, cited), one closed, leading question that puts the inconsistency to the witness, and the exhibit I need on screen. Include no question for which you cannot cite both passages. If there are fewer than [five] genuine inconsistencies, say so rather than padding.

The last prompt earns its restrictions. In September 2026 a Santa Fe lawyer fed a transcript to ChatGPT expecting “a bulletproof summary” and filed a murder-appeal brief containing “false testimony from wholly fabricated witnesses”; $5,000, contempt and a disciplinary referral (The Guardian). NotebookLM’s citations are the safeguard, and only if you open them. More variations are in the prompt library.

Audio Overviews: the appellate record on the commute

NotebookLM’s Audio Overview turns a notebook into a podcast-style audio summary of the sources, reported at ten to thirty minutes. Three sensible uses: absorbing an appellate record or a long judgement while driving, briefing yourself on an unfamiliar area from a stack of client alerts, and turning CLE materials into something that can be listened to twice. It is a comprehension tool, not work product: nothing the hosts say should be quoted, and anything that surprises you should be checked against the written sources. Generate the audio after you have built the one-page brief; listening with the brief in mind makes the simplifications audible.

What it cannot do: it is not Westlaw

NotebookLM has no legal corpus, no citator and no way to know that the case you uploaded was reversed last month. It cannot find authority; it can only read the authority you found. The Divisional Court’s warning in Ayinde applies: “Freely available generative artificial intelligence tools, trained on a large language model such as ChatGPT are not capable of conducting reliable legal research” (Ayinde v Haringey). NotebookLM’s virtue is that it does not pretend otherwise: asked for the leading case on a point that is not in the notebook, it has nothing to cite, and the absence of a citation is your answer.

Nor is it a drafting tool like Claude for Word, or a review platform for thousands of documents like Harvey Vault or Legora’s Tabular Review. It is the reading room, not the library and not the typing pool.

NotebookLM vs Claude Projects vs ChatGPT Projects

All three let you load documents and ask questions across them; they differ on what happens when the documents run out.

NotebookLM has only your sources; Claude and ChatGPT can fall back on what the model remembers, which is useful for drafting and dangerous for facts. For a playbook to apply or a clause to draft, see the Claude Projects and custom GPTs guide; the tools map places all three in the wider stack.

Where to go next: the rest of the tools cluster has the head-to-heads. NotebookLM is one of the tools taught live in AI Lab for Lawyers: in session two you build a matter notebook on anonymised documents and run the brief, timeline and spot-check yourself.