The Verification Layer Is the Product, Not a Feature
Sullivan & Cromwell apologized to a federal bankruptcy judge last month for filing a brief with citations that didn't exist. In the same month, two lawyers in Oregon were hit with $110,000 in sanctions for the same problem in a winery-ownership dispute. The cases sit at opposite ends of the firm-size spectrum, and the lesson sits in the middle.
If a firm with S&C's review process can't catch a hallucinated cite, your three-paralegal review process may not either.
The two cases
In April, S&C apologized to Judge Martin Glenn of the U.S. Bankruptcy Court for the Southern District of New York. The filing was an emergency motion in the Chapter 15 proceeding, In re Prince Global Holdings. The errors were AI-generated.
The same month, U.S. Magistrate Judge Mark D. Clarke imposed $110,000 in combined fines and attorneys' fees against two lawyers in a Valley View winery-ownership dispute in Jacksonville, Oregon. AI-fabricated citations.
Across all U.S. cases in Q1 2026, AI-hallucination sanctions resulted in $145,000 in sanctions. The Charlotin database, which tracks these cases worldwide, now catalogs 1,494 incidents, at least 915 of them from U.S. courts.
These are not edge cases anymore. They are a quarter-over-quarter pattern.
What the firms that got burned had in common
Different size. Different practice area. Different jurisdiction. Same operational gap.
In every case I've read about — and I've now read a lot of them — the firm treated AI output the way it treats associate work product: assume good faith, spot-check on review. That model doesn't work when the model is the drafter. A first-year associate who's not sure about a cite will flag it. A model will invent one with the same confidence it uses for a real one.
The "we'll catch it in review" assumption was load-bearing for the firm's adoption decision, and that assumption broke. The verification step that was supposed to be a human spot-check turned into the entire control mechanism — but humans are not built to catch every fabricated citation in a 30-page brief at 11:39 pm before a filing deadline.
Why "use AI carefully" isn't the answer
Every law firm policy I've seen on AI use comes down to some version of 'be careful, verify outputs, don't file anything without human review.' Let's be honest, those are wishes, not policies.
The Oregon and S&C cases involved lawyers who presumably already knew to verify. They missed it anyway, because the verification step was a manual checklist sitting on top of a fully automated drafting process. The math of that asymmetry doesn't work. You cannot manually verify every citation in every document at the speed at which the model produces them, especially not when the firm has already factored AI-driven productivity gains into its capacity model.
What must happen is structural. The verification step must be part of the same system that generated the draft — not a downstream review queue, not a paralegal checklist, not a "remember to fact-check" line in the firm's AI policy.
What "verification as architecture" looks like
A few concrete examples of what a verification-first AI feature actually does:
Citation validation. Every case citation the model generates is checked against a real case database before it ever reaches a draft. If the citation can't be matched, it doesn't appear in the draft — the model is sent back to find a real one, or the draft is flagged.
Privilege detection. Documents pulled into a matter are scanned against the firm's ethical-wall configuration before they're exposed to a model. The model never sees the conflicted material, because the platform refused to surface it.
Key-date extraction with provenance. When the model pulls a deadline out of a document, the date is linked back to the exact paragraph it came from. The reviewer doesn't have to verify whether the date exists — they verify whether the paragraph supports it. That's a 30-second check rather than a deep 30-minute review.
Version diff before filing. When a draft is touched by AI, the platform shows a structured diff of exactly what changed and why. Reviewers don't have to reread the whole document to look for ghost edits.
Each of these does what the human reviewer was supposed to do, except it runs automatically within the system that produced the draft. The human review is now confirming that the verification ran — not performing the verification themselves.
What this means for small firms shopping for legal tech
If you are evaluating any platform with AI features right now, the question to ask is not "what AI features does it have?" Every platform will check that box this year. The question is: where does the verification step live?
If the answer is "the user verifies the output," you've bought a generation tool with no safety system attached. You're carrying the entire verification load yourself, and the Oregon and S&C cases are the data on what that actually looks like in practice. (It looks expensive.)
If the answer is "verification is built into the workflow,” now you have something. Look for citation checking against a real source. Look for privilege boundaries the model can't cross. Look for provenance on extracted dates and figures. Look for diff views on AI-touched drafts. None of these are speculative — they are the things that, present or absent, decide whether AI in your practice is an asset or a liability.
The verification layer isn't a feature on the spec sheet. It's the entire reason a firm should be willing to put AI inside its workflow at all. Without it, you've just installed a faster way to write a brief that gets sanctioned.
The next decade of legal software is going to sort itself into two categories: platforms that built verification in from the start, and platforms that bolted it on after a court named them in an opinion. Pick early.