Sullivan & Cromwell Filed 40+ AI Hallucinations in Court. Self-Verifying AI Would Have Caught Every One.

$2,000 an hour. 40+ fabricated citations.

On Saturday, Andrew Dietderich — co-head of restructuring at Sullivan & Cromwell, one of the most prestigious law firms on the planet — sent an extraordinary letter to a judge.

He was apologizing.

A court filing prepared by the firm contained more than 40 errors generated by AI: fabricated case citations, misquoted authorities, and entirely non-existent legal sources. The letter included a three-page correction list.

The worst part? Sullivan & Cromwell didn’t catch the errors themselves. Opposing counsel from Boies Schiller Flexner did.

CNN reported the story as perhaps the most egregious instance of “workslop” to date — and noted that Sullivan & Cromwell partners reportedly charge around $2,000 an hour for bankruptcy cases.

”Hallucinations” aren’t bugs. They’re how LLMs work.

In his apology, Dietderich described the errors as “hallucinations” — AI tools that “fabricate case citations, misquote authorities, or generate non-existent legal sources.”

He also admitted the firm has safeguards to prevent exactly this. They just weren’t followed.

This is the core problem with policy-based safeguards: they depend on humans remembering to follow them, every time, under deadline pressure. One associate skips the verification step, and 40 fabricated citations go to a federal judge.

As CNN’s analysis pointed out, this isn’t an isolated incident. It happens with surprising frequency across the legal industry. What makes this case notable is that it happened at a firm with the resources, talent, and policies to prevent it — and it still got through.

The “utility gap” CNN identified is real — but solvable

CNN’s article makes a sharp observation: AI works brilliantly for coding (deterministic — code either runs or it doesn’t) but struggles in domains like law, where there are “degrees of functionality informed by value judgments.”

Investor Paul Kedrosky told the Plain English podcast that AI demand estimates are based on the experience of early tech adopters who are “profoundly unrepresentative of the rest of the real world of work.”

This framing is correct but incomplete. The problem isn’t that AI can’t help lawyers. The problem is that standard AI tools generate answers without verifying them against source material.

When a large language model produces a case citation, it isn’t looking it up — it’s predicting what a plausible citation looks like based on patterns in its training data. Sometimes the prediction matches reality. Sometimes it invents “Johnson v. State of California, 2019” out of thin air.

The fix isn’t to stop using AI. It’s to add a verification layer.

What self-verifying AI looks like in practice

Imagine the Sullivan & Cromwell associate had used a document AI tool that works differently:

You ask a question about your case documents — contracts, filings, precedent briefs, whatever you’ve uploaded.
The system builds a Knowledge Graph linking entities, clauses, dates, and legal concepts across every document. “Debtor” on page 3 is connected to “XYZ Corp” on page 47 — automatically.
Every fact in the answer is checked against the source text before you see it. If a claim can’t be traced to a specific sentence in your documents, it’s flagged or removed — not presented as truth.
Every citation is clickable. One click opens the original PDF at the exact sentence, highlighted on the page. You see the source. You verify in seconds.

That’s how Evidoc works.

If those 40+ citations had been verified against source documents before submission, the fabricated ones wouldn’t have survived. Not because a human remembered to check, but because the system checked automatically.

Policy safeguards fail. Automated verification doesn’t skip steps.

Sullivan & Cromwell’s letter reveals the fundamental weakness of procedural safeguards: they rely on every person, every time, following every step.

An associate under deadline pressure skips the manual verification
A partner trusts the output because the tool has been reliable before
Nobody catches it until opposing counsel does

Automated fact-verification eliminates this single point of failure. The system doesn’t have bad days. It doesn’t skip steps under deadline pressure. Every answer is checked, every time.

The pattern is accelerating

This isn’t new. It’s getting worse:

Mata v. Avianca (2023) — Lawyer submitted six fabricated cases generated by ChatGPT. Sanctioned by the court.
Park v. Kim (2024) — Another attorney fined for AI-hallucinated citations.
Multiple state bars (2025–2026) — New ethics opinions requiring lawyers to verify AI output, with penalties stacking up.
Sullivan & Cromwell (2026) — 40+ errors from one of the most elite firms in the world.

If Sullivan & Cromwell — with their resources, policies, and $2,000/hour talent — couldn’t prevent this, what chance does a mid-size firm have with a manual review checklist?

The answer isn’t more policies. It’s systems that verify by design.

How Evidoc prevents this

Evidoc is built for exactly this scenario — professionals who need AI to help them work with documents, but can’t afford a single hallucinated fact:

Self-verifying answers — every fact is checked against source text before display. Unsupported claims are caught and removed automatically.
Sentence-level citations — click any citation to see the exact sentence highlighted on the original PDF page. Not a page reference. The actual sentence.
Multi-hop reasoning — the Knowledge Graph traces connections across documents step by step. Contract → amendment → invoice. Clause → precedent → filing.
Full audit trail — every answer is traceable and defensible. Your supervising partner, judge, or regulator can verify any claim in seconds.

The Sullivan & Cromwell episode isn’t an argument against using AI for legal work. It’s an argument for using verified AI — where the system proves its answers before you stake your reputation on them.

Try Evidoc at evidoc.hulkdesign.com — verified AI document processing with sentence-level citations.