Choreography: a legal opinion anchored in sources (without hallucinating)
The choreography for producing a legal opinion or memo with AI without falling into the deadly trap of fabricated case law. AI drafts fast, but the truth of the sources and the strength of the argument stay your signature.
Friday, 4pm. The client needs an opinion on termination for cause by Monday. You open AI and type: give me the five leading precedents on termination for job abandonment, with case numbers. In fifteen seconds a beautiful list comes back: five rulings, full case numbers, well-written summaries, all formatted. You almost paste it into the opinion. Except three of those cases don't exist. AI made them up. The numbers are plausible, the summaries sound right, but if you search the court's website, you find nothing. And your name was going on the signature of that opinion.
You need to build the quarterly performance review matrix and ask AI: bring me the turnover, engagement, and absenteeism numbers for my sales team. In seconds a flawless table comes back: round percentages, a comparison with last quarter, even a market benchmark. You almost send it to the people committee. Except AI has no access to your HRIS, it invented the numbers from the pattern. Think with me: the speed of building the matrix is AI's, but the truth of the employee data is yours. Whoever presents to the committee, answers for it. Anchor AI in the real HR spreadsheet before asking for any read, and check every number before signing off.
Friday at the end of the day, you're closing out the PRD and ask AI: list the adoption metrics for the last feature, with D7 and D30 retention. Everything comes back formatted, pretty percentages, even a cohort curve description. You almost paste it into the roadmap to justify the next bet. Except AI isn't plugged into your product dashboard, it filled in a plausible pattern. And that quarter's prioritization was about to rest on those numbers. AI speeds up the PRD structure and the writing, that's real. But the truth of the metric comes from your analytics, and the decision to prioritize is yours. Paste the real data first, check every metric, and only then stand behind the bet.
You're putting together the proposal for a big account and ask AI: bring me case studies from clients in our industry who closed above 200 thousand, with name and outcome. In seconds a convincing list comes back: three logos, ROI numbers, short testimonials. You almost drop it straight into the meeting deck. Except two of those cases don't exist the way it described them, AI filled in the pattern of a typical case. Fine, the proposal came together fast, and that's worth gold at month end. But if the prospect asks for the reference, it's your name in the pipeline that takes the hit. AI speeds up the proposal writing; the truth of the case comes from your CRM and the strength of the negotiation is yours. Anchor it in real sales data, check every case, and only then sign it.
You need to review the SLA with a vendor and ask AI: show me their on-time delivery history for the last six months, with the incidents. A beautiful table comes back: compliance percentage, delay dates, root causes. You almost bring it to the renegotiation table. Except AI isn't connected to your vendor management system, it built the pattern of a plausible-looking history. Think with me: the speed of organizing the report is AI's, but the truth of the operational data is yours. Whoever sits at the renegotiation, answers for the number. Anchor AI in the real SLA log, check every incident against the record, and only then stand behind your position at the table.
With an audit knocking at the door, you ask AI: list the data-privacy law articles that apply to this data flow and the regulator's rulings on the topic. In seconds a tidy compliance opinion comes back: articles cited, administrative case numbers, reference fines. You almost attach it to the internal controls report. Except part of those regulator decisions AI invented, it filled in the pattern of a regulatory precedent. And your risk assessment was about to rest on that. AI speeds up the structure of the compliance opinion, that helps. But the truth of the rule and the decision is yours, and whoever signs the control answers for it. Paste the real text of the regulation, check every citation against the official source, and only then stand behind the risk analysis.
Incident in production, you ask AI: show me how our authentication function handles token refresh and which version of the library we're using. A confident answer comes back: a well-written code snippet, an exact version number, even the changelog. You almost apply the fix directly and ship the deploy. Except that snippet isn't your real code, AI generated a plausible pattern, and it made up the library version. Fine, AI speeds up reading the architecture and drafting the fix, that's real and worth time in the middle of an incident. But the truth of the code comes from your repository and the decision to ship is yours. Paste the real file, check the version in the real package, and only then deploy with your name on the commit.
Friday afternoon, you're closing out the research report and ask AI: summarize what users said about the checkout flow in the latest interviews, with the main pain points. A lovely summary comes back: three pain points named, user quotes in quotation marks, even a percentage of who got stuck at the payment step. You almost bring it to the product team as an insight. Except AI never read your interviews, it filled in the pattern of a typical checkout pain point, and those quotes are made up. Think with me: the speed of structuring the report is AI's, but the truth of the user's voice is yours. Whoever presents the insight, answers for it. Anchor AI in the real research transcripts, check every user quote against what was actually said, and only then stand behind the flow recommendation.
You're closing out the quarter's board deck and ask AI: bring me the market share of the three main competitors and the sector's growth rate over the last two years. In seconds a ready-made slide comes back: round percentages, a source cited alongside, even a trend curve drawn in. You almost took it straight to the board. Except AI isn't plugged into any real industry report, it filled in the pattern of a plausible market, and the cited source doesn't exist. And next year's investment decision was about to rest on those numbers. AI speeds up the deck's structure and the narrative, that's real. But the truth of the market data comes from your real source, a paid report, your own research, auditable public data, and the decision to invest is yours. Anchor AI in the real reports, check every number against the original source, and only then take the slide to the board.
Thiago here. Let me be direct with you: AI doesn't "know" case law. It predicts text that looks like case law. When you ask "give me the precedents," it doesn't consult a court, it completes a pattern. And completing a pattern with total confidence is exactly what it does best, including when the pattern is false. The danger isn't AI getting it wrong in an obvious way. It's getting it wrong beautifully.
AI speeds up the drafting and structure of the opinion. The truth of the sources and the strength of the argument are yours, and remain yours. Whoever signs, answers for it.
01The deadly trap: case-law hallucination
There's an error in legal work that isn't like the others. A typo, you fix. A weak thesis, you reinforce. But a fabricated citation that makes it into an opinion and reaches the client, the judge, the opposing party, that error destroys the case and your credibility in the same stroke.
Case-law hallucination happens because the model was trained to sound convincing, not to be true. When you ask for a ruling and it doesn't have a real one available, it doesn't answer "I don't know." It builds one. Case number in the right format, coherent summary, a judge with the name of a real person. All false, all confident.
It has already happened for real. Lawyers were sanctioned by a court in the United States for citing six rulings that AI made up. They didn't check. They figured the machine wouldn't lie. The machine didn't lie, it just completed a pattern. The responsibility was on whoever signed.
The economic frame is cold: drafting an opinion with AI got cheap and fast. But a single false citation costs the case, costs the client, costs your reputation. The cost of checking is minutes. The cost of not checking is your career. There's no trade here that's worth it. Fair?
02The four-beat choreography
The right way isn't to ban AI. It's to choreograph it. Four beats, always in the same order, without skipping any.
Beat 1, you anchor. You don't ask AI to remember case law. You give it the real sources: the passages of the law, the rulings you already have in hand, your trusted base. This is what's called RAG, retrieve before generating. AI works with what you provided, not with its memory.
Beat 2, AI drafts. With the real passages in hand, it puts together the draft of the opinion: structure, argumentation, connecting the points. Here it's fast and good. This is its job.
Beat 3, you verify. Every citation, one by one. Does the law exist? Does it say what the draft claims it says? Is the ruling real? Does the summary match? You open the source and look. You never delegate this.
Beat 4, you stand behind it. The final argument is yours. You adjust the thesis, take the position, sign. Because the signature is yours, and the signature is what stands behind it.
03How to really anchor
Anchoring isn't saying "only use reliable sources" in the prompt. That's a request, not anchoring. AI has no way of knowing what's reliable if you don't put the source in front of it.
Anchoring is pasting the material. You open the real ruling, copy the summary and the relevant passages, and hand them to AI together with the question. You paste the statute with the exact text. You give it the legal scholarship with the page number. AI starts working inside a box: only what you provided exists for it.
In a more mature setup, this becomes a RAG base: your library of case law and legislation connected to AI, from which it retrieves the passages before drafting. This is what we connected back in N.jur.2, when we talked about connecting your sources. The logic is the same: AI doesn't invent when it doesn't need to invent, because the true answer is already on the table.
The instruction that closes the box is explicit: "Use exclusively the passages below. If something isn't in the passages provided, write 'not found in the sources' instead of filling it in." This sentence changes the model's behavior. It takes off the pressure to invent and gives it permission to say it doesn't know. And "I don't know" is an honest answer, one you can work with. A fabricated citation isn't.
04The division of labor: what is AI's, what is yours
Every good choreography has a clear division of roles. Here it's simple and non-negotiable.
Notice the boundary: everything AI does is form and speed. Everything that's yours is truth and responsibility. The drafting gets fast, that's real and worth it. But the speed doesn't carry over into the verification. You don't check faster because AI wrote faster. The time you saved drafting, you reinvest checking. That's the deal.
And that's why the verification isn't an optional step for whenever there's time. It's the step that justifies everything. An anchored and verified opinion is an asset. A generated and unverified opinion is a bomb with your name on the fuse.
05The verification ritual
Checking a citation isn't "taking a quick look." It's a ritual, with fixed steps, that connects directly to the idea of the audit we go deeper into in N.jur.7.
For every citation AI put in the draft, three questions:
Does the source exist? You open the court's website, the legislation portal, the official database, and locate it. If you can't find it, the citation goes. No exceptions, no "it must just be a typo in the number."
Does the source say what the opinion claims? Existing isn't enough. The real ruling might be about something else, or say the opposite of what the draft suggests. You read the relevant passage and confirm it supports the argument. AI is great at taking a real ruling and describing wrong what it decided.
Is the citation complete and faithful? Correct number, correct year, correct court, context preserved. A ruling taken out of context becomes a half-truth, and a half-truth in an opinion is a weakness the opposing party will exploit.
Only after the three right answers does the citation stay. Mark each one as verified. If you didn't check it, it doesn't go in. The rule is binary: verified or out. There's no "probably right." Fair?
Do it now
Take your real task: a real legal opinion or memo you need to produce (or one you produced recently) and run the full choreography:
- ANCHOR: gather the real sources before opening AI. Paste the statute
passages, the rulings you have, the relevant legal scholarship. Include the closed-box instruction: "Use exclusively the passages below. If something isn't there, write 'not found in the sources'."
- DRAFT: ask AI for the draft of the opinion based only on these passages.
- CHECK: list every citation that appeared in the draft. For each one, answer
in writing: does it exist? Does it say what the opinion claims? Is it complete and faithful? Mark each citation as [checked] or [out].
- STAND BEHIND IT: rewrite the final thesis in your own words and your position.
Sign mentally: would you defend this in front of a judge?
Write down: how much time AI saved on the drafting, and how much time the verification took. That's your real trade-off between speed and responsibility.
Practice
1. You ask AI "give me three STJ rulings on intercorrent prescription, with case numbers" without providing any source. What most likely happens?
2. What is the correct division of roles in the anchored-opinion choreography?
3. While checking a citation, you find the real ruling on the court's website, but it deals with a different topic and doesn't support the draft's argument. What do you do?
For the board
On the mechanicsthe AI does not know case law. It predicts text that looks like case law, and completes the pattern even when the pattern is false.
On the frightening errorthe danger is not getting it badly wrong, it is getting it beautifully wrong.
On verificationa real ruling described wrongly is still a failure. Check that it exists, then check what it actually says.
Thanks for the feedback. It helps sharpen the next lesson.