In April 2026, a federal judge in Oregon ordered two lawyers to pay roughly $110,000 — the largest sanction of its kind the state had seen — after they filed briefs built on cases that didn’t exist. The citations were invented by a generative AI tool, copied into the filing, and signed without anyone checking whether the cases were real.
That number gets the headlines. But the thing worth sitting with is why it happened, because it wasn’t really about the AI. It was about a missing 15-minute step. The lawyers used ChatGPT (or something like it) as a research engine, trusted the output, and never opened a single one of those cases to confirm it was real before putting their name on the document.
If you’re a solo or run a small firm, here’s the honest version: you can absolutely use ChatGPT to save hours a week. Lots of attorneys already do. The ones getting sanctioned aren’t the ones using AI — they’re the ones who skipped the part where a human checks the work. This post is about that part.
What’s actually happening in 2026
The sanctions are real, they’re recent, and they’re piling up.
- Oregon, ~$110,000. A federal magistrate judge sanctioned two attorneys for filing briefs with non-existent cases and fabricated quotations generated by AI and submitted without verification. The court called it an outlier in scale — and a warning to everyone else.
- Oregon Court of Appeals, $10,000. In Henry Doiban v. Oregon Liquor and Cannabis Commission (March 2026), an attorney’s brief contained citations and quotes that, in the court’s words, “do not exist anywhere in Oregon case law.”
- Kansas, $12,000 in Rule 11 sanctions. Five attorneys in Lexos Media IP v. Overstock.com were fined between $1,000 and $5,000 each for fabricated citations — with the lawyer who actually used the AI and skipped the check taking the biggest hit.
Researchers tracking this have now logged hundreds of AI hallucinations in U.S. court filings, and at least 25 federal district courts have standing orders requiring attorneys to certify whether AI was used and confirm a human reviewed everything. The 2026 message from the bench is blunt: “the standards were unclear” stopped being a defense a while ago.
Here’s the part the headlines skip. Adoption has completely outrun governance. Clio’s 2026 Legal Trends data shows 79% of legal professionals now use AI in some form — and among solos and small firms specifically, the majority do. But roughly 55–57% of solo and small firms have no written AI policy at all, and 44% of firms overall have no formal AI governance. So most of the profession is using these tools with no rules, no checklist, and no safety net. That gap is exactly what the courts are punishing.
The one rule that keeps you off the sanctions list
Strip away the legal-tech marketing and the rule is almost embarrassingly simple:
AI writes drafts. You verify facts. You sign the filing — so you own every word in it.
Every sanctioned lawyer broke the second half of that sentence. The AI gave them a confident, professional-looking citation, and they filed it without confirming it existed. The model wasn’t lying to them so much as doing exactly what these tools do: predicting text that looks like a real case. Sometimes the case is real. Sometimes it’s a beautifully formatted ghost.
This also tracks with the ethics rules. ABA Formal Opinion 512 (issued July 2024) didn’t invent new duties — it applied the ones you already have to AI. The relevant ones:
- Competence (Rule 1.1): you have to understand the tool well enough to know it can hallucinate — and verify its output.
- Confidentiality (Rule 1.6): before you put client information into a tool, you have to understand how that tool stores and uses the data — and for “self-learning” tools that may train on your inputs, get the client’s informed consent.
- Candor to the tribunal (Rule 3.3): anything you file has to be accurate. “The AI wrote it” is not a shield.
- Fees (Rule 1.5): you can’t bill a client for the hours you spent learning the software.
Translation: treat AI like a sharp but unsupervised junior. Useful for a first pass. Never the final word.
The 3-step citation check (do this before anything gets filed)
This is the whole ballgame. It takes about 15 minutes and it’s the difference between “saved three hours” and “$110,000 and a bar complaint.”
Step 1 — Pull the case yourself, in a real database. For every authority the AI gives you, open it in Westlaw, Lexis, or a court’s own docket. Not “search the name and see a few results.” Open the actual opinion. If you can’t find it, it doesn’t exist — delete it and move on. (This is also why attorneys on the front lines say things like “never trust an AI-generated citation unless it came from CoCounsel inside Westlaw” — i.e., a tool wired directly to the real reporters.)
Step 2 — Confirm the quote and the holding. A real case can still be misused. Read the part the AI quoted and confirm (a) the quote actually appears, word for word, and (b) the case actually stands for what the AI says it does. Hallucinated quotations inside real cases got attorneys sanctioned in 2026 just as often as fake cases did.
Step 3 — Sign it as if no AI touched it. Before you file, ask yourself: “Would I put my bar number on this if I’d written every word myself?” Because legally, you did. The signature is the certification. As one small-firm lawyer put it after watching colleagues get sanctioned: policies live in a document, but verification has to live in the workflow. Make Step 1 and Step 2 a non-negotiable habit, not a poster on the wall.
Notice what the check is not: it’s not “don’t use AI.” You can still hand ChatGPT a messy fact pattern and get a clean intake summary in two minutes. You can still turn six bullet points into a first-draft demand letter. You just don’t file the legal authorities until a human has pulled each one.
What this means for you
If you’re a brand-new solo: This is your edge, not your enemy. You don’t have a paralegal, so AI handling the intake summary and the first draft of routine letters genuinely buys back hours. Start there — the no-court-filing tasks — and build the 3-step check into anything that touches a brief from day one.
If you run a 2–10 lawyer firm: Your biggest risk isn’t you — it’s the associate or staffer who quietly pastes a client’s name into the free ChatGPT to “just speed something up.” Write a one-page AI policy this week (we have a free template here) and make the verification step part of how filings get reviewed, not an honor system.
If you mostly do transactional work: Lower courtroom risk, but the confidentiality rule still bites. Be deliberate about what client data goes into a consumer tool versus an enterprise one with a data agreement.
If you’re a litigator: Treat AI as research prep — a way to orient fast on an unfamiliar area — and then do your real research in the databases. The fastest way to lose a judge’s trust in 2026 is a fabricated cite, and they are actively looking.
What ChatGPT can’t do for a lawyer (be honest with yourself)
- It can’t do your legal research. It can summarize, brainstorm, and draft. It cannot reliably tell you what the law is, and it will invent authorities that sound perfect.
- It can’t keep a secret it was never told to keep. Consumer ChatGPT is not a privileged channel. Don’t paste client-identifying facts, full case files, or anything you’d be horrified to see in a data breach into a general tool — de-identify, or use a tool with a real confidentiality agreement.
- It can’t certify accuracy — only you can. The signature, and the liability, are yours.
- It can’t replace judgment. It has no idea which argument will actually land with this judge in this jurisdiction. That’s the part clients pay you for.
- It can’t make a no-policy firm safe. The tool is only as careful as the workflow around it.
The bottom line
The lawyers who got sanctioned in 2026 weren’t reckless because they used AI. They were reckless because they let AI write and sign. Keep those two jobs separate — AI drafts, you verify, you own it — and ChatGPT becomes one of the best junior assistants a solo ever had: fast, tireless, and cheap, as long as you never let it near the “file” button unchecked.
If you want the full, safe workflow — from intake summary to first draft to a verification habit you’ll actually keep — our AI for Lawyers (Practical) course walks through it step by step, ethics rules included. And if the confidentiality side is what worries you most, ChatGPT Without the Liability is built around exactly that.
Use the tool. Just never sign what you didn’t check.
