Hello fellow Prompt Engineers Who Passed the Bar and Senior Partners Who Have Definitely Read the Firm’s AI Policy,
Judges are sanctioning people for fake citations. Junior associates are using ChatGPT at 1 am and billing it as “research.” Legal AI vendors are charging $50,000 a year for software the firm hasn’t finished onboarding. And somewhere right now a partner is explaining AI to his team using information he got from AI, which did not flag this as a problem.
We are all in this together. Let’s get into it. ⚖😂
| Observed: | In the wild, 2024–2025. Six species. One is correct. |
| The Secret User | Denies using AI to anyone who asks. Has four ChatGPT tabs open right now. Billing 0.3 hours as “AI review” and hoping no one asks what that means. |
| The True Believer | Put “AI-forward” in his LinkedIn bio in February 2023. Has deployed four platforms. Uses none with discipline. Recently told a client their matter was “fully AI-optimized,” which is not a thing. |
| The Purist | Still types everything on a desktop he calls “the computer.” Views AI with the suspicion reserved for opposing counsel who is being too friendly. His paralegal is using it. He does not know. |
| The Anxious Junior Associate | Uses AI constantly because it’s 1 am and the partner wants the draft by 7. Also convinced AI is coming for her job. Both things are true. She’s somehow fine. For now. |
| The In-House Evangelist | Bought every seat on the enterprise platform. Onboarded zero team members. Presents the ROI slide at every legal ops conference. The ROI is speculative fiction. The demo was very compelling. |
| The Careful One | Uses AI only for tasks she can verify. Reads every output. Treats it like a first-year associate: useful, unreliable, not to be sent to court unsupervised. Zero incidents. Nobody writes articles about her. |
| Subject: | A brief history of AI-assisted court filings, annotated |
You know Mata v. Avianca. Lawyer. ChatGPT. Six invented cases. Page numbers for opinions that don’t exist. A judge who was not amused and very much on the record. That was not a freak incident. It was a genre-defining entry. The pattern, for those taking notes:
| Step 1 | Lawyer asks AI for supporting cases. AI, constitutionally incapable of saying “I don’t know,” produces them. Real judge names. Plausible dockets. Holdings that are exactly what the lawyer needed. Formatted like Westlaw. |
| Step 2 | Lawyer, relieved, does not check. The deadline was two hours ago and the client is texting. |
| Step 3 | Opposing counsel Shepardizes. Discovers the cases are fiction. Has the best week of his career. Judge schedules a hearing. Lawyer has the worst week of his. |
| Step 4 | Law school sends an ethics newsletter. Nobody reads it. The CLE industry adds a module. Nobody changes anything. Repeat from Step 1. |
The problem isn’t that AI hallucinates. The problem is that it hallucinates with the measured, authoritative confidence of a senior partner who hasn’t been wrong since 2009 and is not flagging uncertainty. The fix is Westlaw. Five minutes. Every time.
| Finding: | The AI is not bad at its job. You are bad at giving it one. |
The No. 1 complaint lawyers have about AI output is that it’s generic. The reason is the prompt. “Write me a motion to dismiss” is not a prompt. It is a wish. A sampler:
| What You Typed | What the AI Understood | Verdict |
| “Write a demand letter.” | A letter. Demanding something. For someone. Tone: vaguely firm. | Overruled. |
| “Write a demand letter. Construction dispute. $340K unpaid. Florida. Done being patient.” | Better. But it still doesn’t know your client, the owner’s defense, or whether you want to leave the door open for settlement. | Under advisement. |
| “You are a senior Florida litigator. Client is the contractor. Owner owes $340K, claims defects that don’t exist. We’ve been patient four months. Draft a letter that signals we are 10 days from filing without saying so. No exclamation points. Flag where I need to insert case-specific facts.” | Exactly that. Output is four times better. Editing time: ten minutes instead of forty-five. | Sustained with prejudice. |
| Warning: | A token budget is not a suggestion. It is load-bearing. |
AI agents can take actions on your behalf: search documents, run tasks, chain instructions together, and operate autonomously while you are in a deposition. They are remarkable. They are also, in token terms, a running faucet connected to a meter you are not watching.
Every step costs tokens. Every tool call, document read, and sub-task costs tokens. An agent you deploy with “review all documents in this folder” will review all documents in that folder. Thoroughly. Then it will present you with a bill that makes your paralegal’s overtime look reasonable.
| Ruins your day: | “Review the contracts.” |
| Does not: | “Review these twelve contracts for indemnification clauses only. Flag deviations from our standard template. Stop after those twelve.” |
Scope your agents like you scope a matter. The agent does not have a concept of “seems like enough, wrapping up.” It has instructions. The quality of those instructions is the difference between a useful tool and a very interesting cloud invoice.
That 12% utilization problem? It doesn’t happen when the tool is actually built for legal work.
CoCounsel by Thomson Reuters is the GenAI built specifically for legal professionals — trained on authoritative legal sources, not the open internet. Research, drafting, document review, deposition prep. The work that takes hours, done in the time it used to take to find the right binder.
The Coffee Infrastructure Working Group has been reviewing AI tools since 2024. CoCounsel didn’t wait for the committee.
| Disclosure: | A totally unbiased comparison that will make our sponsors mildly uncomfortable |
| Factor | General AI (ChatGPT, Claude, Gemini) | Legal-Specific AI (CoCounsel, Harvey, etc.) |
| What it costs | $20/mo (personal) to $30/user/mo (enterprise) | Requires a demo call, a proposal, and a procurement conversation that will haunt you |
| Citation accuracy | Will invent cases. Confidently. With page numbers. See Exhibit B. | Connected to actual legal databases. Meaningfully better. Still verify. Always verify. |
| Integration timeline | Open a tab. Start typing. The IT department finds out later. | “Six weeks” per the pitch deck. Eight months per the IT coordinator who has stopped returning calls about it. |
| Actual firm usage | What associates are using right now, regardless of firm policy | What was announced at the all-hands with champagne. Currently at 12% utilization. |
| Who takes the blame | You. Alone. At a bar hearing. | Still you. The vendor has indemnification language. It does not cover bar hearings. The sales rep was very confident about this. |
| Best for | First drafts, client summaries, brainstorming, translating legalese into English your client might actually read | Contract review at volume, legal research, deposition prep, document review that would otherwise cost $80K in associate hours |
| Authority: | RPC 1.6 would like a word. The word is: no. |
| What happened: | Pasted client’s acquisition term sheet into a free-tier consumer AI product. |
| What this is: | A third-party data processing event your client did not authorize, not covered by your engagement letter, and very interesting to your malpractice insurer. |
| The fix: | Enterprise agreements. Data processing terms. Your firm’s AI policy, which was covered at the February all-hands titled “AI and Data Governance: What You Need to Know.” You had a conflict. RPC 1.6 did not. |
The legal industry’s relationship with AI has moved through all five stages of professional grief in eighteen months. Denial. Anger (the brief had made-up citations). Bargaining (“we’ll use it only for non-privileged first drafts under appropriate supervision as defined by a committee”). Depression (another standing order). Acceptance (“fine, show me the prompt thing”).
Most of you are at Acceptance. Some are still at Bargaining, refining the committee scope. The Purist is at Denial. His paralegal is at Acceptance and has everything handled.
Here is the part nobody says at legal tech conferences because it doesn’t sell a platform: the real advantage right now isn’t which tool you bought. It’s whether you know how to use it. Prompting well. Scoping agents correctly. Knowing when the general-purpose version is fine and when you need the legal-specific one. Verifying before you file. That’s craft now. It’s learnable. Back when I was doing document review at 2 am on a case that settled before trial… anyway. The tools would have helped. Use them correctly.
Walter, Editor-in-Law
Licensed to practice sarcasm in all 50 states. AI use disclosed. Citations verified. Mostly.
Which species are you? Best prompt disaster you’ve witnessed? Best AI output that was actually good? Anonymous dispatches welcome. Research purposes only. No DMS logs will be subpoenaed.
Forward this to the partner who just explained AI to his team using information he got from AI. He needs this more than you do.