In November 2026, Sullivan & Cromwell — a firm founded in 1879, with roughly 900 lawyers and a client roster that includes some of the most consequential transactions on Wall Street — admitted that a filing in a potentially billion-dollar Manhattan bankruptcy case contained fabricated case citations and misquoted legal authority. Not a solo practitioner. Not a regional firm stretching beyond its resources. A firm with one of the deepest compliance benches in the legal industry, filing in one of the most scrutinized courts in the country, on a matter large enough to draw institutional attention from the moment it was docketed.
If S&C's internal review process, partner sign-off chain, and malpractice-conscious culture couldn't catch it, the lesson isn't that lawyers need to try harder. It's that the tools they're trying harder with are architecturally incapable of preventing the failure in the first place.
What Actually Happened in the S&C Filing
The specifics reported by Missouri Lawyers Media are stark in their simplicity: a court filing in a large Manhattan bankruptcy matter cited cases that do not exist and misquoted authority that does. S&C acknowledged the error — a notable departure from firms that have tried to minimize or explain away similar incidents — but acknowledgment doesn't undo the exposure. In bankruptcy proceedings of this size, opposing counsel, creditors' committees, and the presiding judge are reading every citation. A fabricated case isn't a typo; it's a credibility event that can taint every other representation the firm makes in the matter.
The uncomfortable part is that S&C is not an outlier case study in negligence. It's a data point in a trend line that's been climbing for three years, now cutting across firm size, practice area, and jurisdiction.
The Scale Problem: 2,262 and Counting
Damien Charlotin, the researcher who has become the de facto record-keeper of this phenomenon, maintains a public tracker of AI-hallucination court decisions. As of early November 2026, it logged approximately 2,262 cases globally, with 1,533 in the United States alone. These aren't anecdotes anymore — they're a measurable, accelerating pattern with its own case law on sanctions.
Recent entries illustrate the range of consequences:
- Two Oregon attorneys were sanctioned more than $110,000 after submitting filings with 15 fabricated citations — a single case, double-digit fake authorities, and a six-figure penalty.
- Courts that once issued modest fines for a first offense are now moving toward license suspensions for attorneys who file unverified AI-generated content, signaling that judicial patience has run out.
- The pattern spans solo practitioners, mid-size firms, and — with S&C's admission — the uppermost tier of the AmLaw 200.
The trajectory matters more than any single incident. Early hallucination cases (2023's Mata v. Avianca being the best known) were treated as novel curiosities. By late 2026, with over two thousand documented instances, they're a recognized category of professional liability risk that malpractice carriers, general counsel, and court clerks are actively pricing in.
Why "Verification by Hope" Fails Even at the Top of the AmLaw 200
Here's the thesis worth sitting with: S&C almost certainly has an AI usage policy. It almost certainly requires citation checking before filing. It almost certainly has partners who reviewed the document before it went out. And the hallucination still made it into a billion-dollar bankruptcy filing.
That's not a training problem. It's not a culture problem. It's what happens when you ask human review to catch errors that are specifically engineered to look correct.
General-purpose large language models — whether accessed through a consumer chat interface or embedded in a per-seat legal SaaS product — generate text by predicting plausible next tokens, not by querying a verified database of case law. A hallucinated citation isn't a garbled sentence a reviewer would catch on a skim; it's a properly formatted case name, a real-sounding reporter citation, and a quote that reads exactly like something a court would say. The entire failure mode is designed to defeat the kind of fast, pattern-matching review that busy associates and partners actually do under deadline pressure.
This is "verification by hope": trust that a human will catch what the model got wrong, after the model has already produced an output that is statistically optimized to look right. It worked, sort of, when hallucination rates were low and usage was light. At the volume AmLaw firms are now running AI-assisted drafting — and at the scale of 2,262 documented failures — hope is not a control. It's an absence of one.
Verification by Design: What Grounded Retrieval Actually Changes
The alternative isn't "use AI less carefully." It's changing what the AI is allowed to say in the first place.
A private, retrieval-grounded AI deployment doesn't ask a model to recall case law from its training data — training data that is frozen at some cutoff date, probabilistically compressed, and entirely capable of inventing a plausible-sounding Smith v. Jones that never existed. Instead, it retrieves actual passages from a verified, indexed corpus — the firm's own case law database, filed briefs, internal research memos, and licensed legal research sources — and constrains the model's answer to what it can point back to in that retrieved material.
The practical difference:
| Capability | Consumer AI assistant | Per-seat legal SaaS | Private / on-prem grounded AI |
|---|---|---|---|
| Source of case citations | Model's training memory | Mix of retrieval + model memory, varies by vendor | Firm's verified corpus + licensed case law index only |
| Citation traceability | None — no source link | Partial, vendor-dependent | Full — every citation maps to a retrievable passage |
| Can it cite a nonexistent case? | Yes, structurally possible | Reduced but not eliminated | Blocked by design — no corpus match, no citation |
| Audit trail for risk/compliance | None | Vendor logs, limited firm visibility | Full firm-controlled logs, permissions, and version history |
| Where full documents reside | Vendor cloud | Vendor cloud (shared-tenant) | Firm infrastructure |
| What leaves the firm per query | Entire prompt + context | Entire prompt + context | Minimal retrieved chunks only, under firm-chosen API terms |
This is the architectural distinction that actually matters — not a slogan about data never leaving the building, but a structural choice about what the model is allowed to invent versus what it's required to retrieve and cite.
The Honest Nuance: Full Corpus Control vs. Minimized Exposure
It's worth being precise here, because the lazy version of this argument ("we never send data anywhere, they send everything") isn't accurate and doesn't hold up to scrutiny from a sophisticated CIO.
In a RAGbase-style private deployment, the firm can still use the same frontier LLM providers that power consumer and SaaS tools. The difference is architectural, not categorical:
- What stays on the firm's infrastructure: the full document corpus, the retrieval and indexing layer, vector stores, agentic workflows, permissions, and complete audit logs. The firm controls who can query what, which documents are indexed, and how long everything is retained.
- What leaves, per query, under the firm's own API terms: only the minimal retrieved chunks necessary to answer the specific question — not the entire client file, not the full matter history, not the firm's whole knowledge base.
That's the difference between a shared-cloud legal AI tool, where the firm's prompts and context typically transit and sometimes persist on vendor infrastructure under the vendor's terms, and a private architecture where the firm decides what minimal footprint leaves at all, and to whom. For a billion-dollar bankruptcy matter with multiple creditor constituencies and press attention, that distinction is not academic — it's the difference between a defensible data governance story and a vendor contract nobody in risk management has fully read.
What This Means for Risk Committees and Innovation Leads
Tools like Claude Cowork, Harvey, CoCounsel, Lexis+ Protege, and Legora have each pushed the market forward on usability and have driven real adoption inside AmLaw firms — that's not in dispute. But usability and grounding are different design axes, and the S&C filing is a reminder that a slick interface doesn't guarantee a hallucination-proof output. The question an innovation committee should be asking isn't "which tool has the best demo" — it's "which architecture makes a fabricated citation structurally impossible to produce, versus merely less likely."
For sovereignty-critical workloads — bet-the-company litigation, billion-dollar restructurings, matters with regulatory exposure, anything where a fabricated citation becomes a press headline — that question has a specific answer. Three things to evaluate before the next major filing goes out:
- Does the tool retrieve from a verified corpus, or generate from model memory? Ask the vendor directly whether citations are retrieval-grounded with source traceability, or produced from the model's general training and then spot-checked.
- Can every citation in a draft be traced to a specific retrieved passage? If the answer is "not reliably," the tool is still operating on verification by hope, regardless of how good its interface is.
- What is the actual data flow per query — not the marketing claim, but the architecture? Full corpus and agent layer on firm infrastructure, with only minimal chunks sent to the model, is a materially different risk profile than full prompt-and-context transiting to a shared-tenant cloud product.
Firms serious about this shift are also building internal verification habits that mirror what tools like case search are designed to support — cross-referencing every AI-suggested authority against a live, authoritative index before it reaches a filing, not after a judge flags it.
From Fines to License Suspensions: The Regulatory Trajectory
The courts' posture is hardening fast. Early hallucination sanctions, including the widely cited Mata v. Avianca matter, drew fines in the thousands of dollars and public embarrassment. The Oregon case — $110,000-plus for 15 fabricated citations — shows the penalty scaling with the severity and volume of the fabrication, not staying flat. And the move toward license suspensions, now reportedly under active consideration in multiple jurisdictions, changes the calculus entirely: this is no longer a cost-of-doing-business fine a firm can absorb. It's a career-ending sanction for the attorney who signed the filing, and a reputational event for the firm whose name is on the letterhead.
For a firm the size of S&C, with 900 lawyers generating thousands of filings a year, the math on unmanaged AI risk doesn't stay flat either — it compounds with volume. The firms moving fastest to grounded, auditable AI infrastructure aren't doing it because it's trendy. They're doing it because the sanctions regime just changed, and the old mitigation (a policy memo and a reminder email) was never a control — it was a hope.
The S&C filing won't be the last billion-dollar matter where an AI-generated citation makes it past review — the tracker's 2,262 entries and climbing say that much. What it clarifies is where the fix actually lives: not in a stricter attorney certification, not in a longer training deck, but in the architecture that decides whether a hallucinated case can be produced at all. Firms evaluating their next AI deployment should treat citation traceability and corpus grounding as non-negotiable procurement criteria, not nice-to-haves — and should read the fine print on where their documents actually live before the next filing goes out under the firm's name. For a deeper breakdown of what grounded, private AI deployment looks like in practice, and how it compares across the current market, the AI for law firms guide is a reasonable next stop.
Frequently Asked Questions
Did Sullivan & Cromwell actually admit to filing fabricated AI citations?
How many AI hallucination cases have courts documented so far?
How does private or on-premise AI actually prevent hallucinated citations?
Related Articles
AI for Law Firms in 2026: The Complete Guide to Choosing, Deploying, and Owning Legal AI
Comprehensive guide to AI adoption for law firms in 2026 — agentic AI, proprietary vs SaaS, privilege implications, pricing, and the ownership model.
Your AI Vendor's Moat Is Your Data. Here's How to Take It Back.
How SaaS AI vendors build competitive moats from your firm's usage data — the shared learning paradox, the dilution problem, and why proprietary AI keeps the compounding advantage with you.
98% of AmLaw 200 Firms Use AI — But Most Still Can't Search Their Own Files
98% AI adoption, but most law firms still can't search their own institutional knowledge. The gap between external AI tools and internal document access — and how to close it.
Agentic AI for Law Firms: What It Actually Means in 2026
What agentic AI actually means for law firms — plain-English definition, what the big players are doing, real deployment examples, and how custom agents differ from SaaS workflows.
RAGbase builds private AI systems for law firms: deployed on the firm's own infrastructure, zero data retention, full ownership.
See How RAGbase Works on Your Data
30-minute call. We scope your use case and show the system live.