A Wisconsin DA Used AI to Draft Court Filings. 74 Criminal Counts Were Dismissed.
By Don Ho, Co-Founder & CEO, Kaizen AI Lab
Published: February 1, 2026
TL;DR: A Wisconsin district attorney used AI to draft filings in a 74-count criminal case. The AI hallucinated case law. The court caught it. The case was dismissed. Not reduced. Not continued. Dismissed. I spent 19 years in legal practice, and I've never seen a tool make it this easy to commit professional malpractice at scale. The verification framework at the bottom exists because apparently we need one.
---
74 Counts. Zero Convictions.
A district attorney's office in Wisconsin had a 74-count criminal case. That's a major prosecution. Multiple charges, likely involving significant harm, representing months of investigation and preparation.
Someone in the office used an AI tool to draft legal filings for the case. The AI generated citations to legal authorities: case names, reporters, page numbers, legal principles. The kind of authoritative-sounding citations that belong in a professional legal brief.
The citations were fabricated. The cases didn't exist.
The court discovered the hallucinated citations. The consequence wasn't a slap on the wrist or an order to refile. The court dismissed the case. All 74 counts.
A defendant facing 74 criminal charges walked free. Not because the charges lacked merit. Not because the evidence was insufficient. Because the prosecutor's office submitted AI-generated fiction as legal authority, and the court determined that the integrity of the proceedings was compromised.
How AI Hallucinations Work
To understand how this happens, you need to understand what AI hallucination actually is.
Large language models don't retrieve information from a database. They predict the next word in a sequence based on patterns in their training data. When you ask for legal citations, the model doesn't look up real cases. It generates text that looks like legal citations based on patterns learned from millions of documents containing legal citations.
Sometimes the generated citation happens to match a real case. Sometimes it doesn't. The model doesn't know the difference. It has a concept of "plausible" based on statistical patterns, not a concept of "real."
The result: AI can generate a perfectly formatted citation to a case that never existed, argued before a court that doesn't hear that type of case, published in a reporter that doesn't cover that jurisdiction. Every element looks correct individually. The combination is fabricated.
This is a fundamental characteristic of how these models work, not a bug that gets patched in the next update. The hallucination rate can be reduced through better training, retrieval augmentation, and fine-tuning. It cannot be eliminated. Any workflow that relies on AI-generated factual claims without independent verification is vulnerable.
The Pattern Is Accelerating
The Wisconsin DA case follows a trajectory that's becoming distressingly familiar.
In June 2023, attorneys in Mata v. Avianca submitted a brief to the Southern District of New York containing six fabricated case citations generated by ChatGPT. Judge Castel sanctioned both attorneys. The brief became the single most-cited example of AI misuse in legal proceedings.
In 2024, multiple attorneys across different jurisdictions were sanctioned for submitting AI-generated citations without verification. Each time, the attorney's defense was some variation of "I didn't know the AI could make things up." Each time, the court responded that ignorance of the tool's limitations doesn't excuse professional responsibility.
What makes the Wisconsin case different is the scale of consequence. Mata v. Avianca resulted in sanctions and fines. The Wisconsin case resulted in the dismissal of 74 criminal charges. Real charges, presumably backed by real evidence, against a real defendant.
The trajectory is clear. The consequences are escalating.
The Tool Made Malpractice Effortless
Here's where my legal background makes this especially infuriating.
I practiced law for 19 years. I've reviewed thousands of briefs. I've caught cite-checked errors in associate work, in opposing counsel's filings, in my own drafts at 2 AM. Verification was never the hard part. The hard part was the analysis, the strategy, the argument construction. Checking whether Smith v. Jones actually said what you claimed it said took five minutes on Westlaw.
What AI did was invert the effort curve. It made the easy part (drafting) instantaneous and the hard part (verification) feel optional. When you get a polished, professional-looking brief in 30 seconds, the psychological temptation to skip the "boring" verification step is enormous. The document looks done. It reads like competent work product. Every visual cue signals "this is finished." The only thing missing is accuracy.
The people using these tools aren't all incompetent. Some of them are. But some are competent professionals who got seduced by a tool that made their output look better than their process deserved. That's a design problem as much as a user problem.
Lessons for Every Industry
If you think this is a legal industry problem, expand your frame.
Healthcare: An AI generates a treatment recommendation citing a clinical study that doesn't exist. A physician follows the recommendation without verifying the study. Patient harm results.
Financial services: An AI generates a compliance report referencing regulatory guidance that was never issued. The report is submitted to a regulator. The company faces enforcement action for misleading submissions.
Engineering: An AI generates a structural analysis citing safety standards with fabricated specifications. The analysis is approved without independent verification. A structure is built to incorrect specifications.
Insurance: An AI generates a claims assessment citing policy language that doesn't match the actual policy. The assessment is used to deny a claim. The insured sues.
Every industry that uses AI to generate documents containing factual claims faces the same risk. The specific consequences vary by domain. The underlying vulnerability is identical.
The VVAC Framework: Verify, Validate, Attribute, Certify
Every organization using AI for document generation needs a named, documented verification process. I'm calling this VVAC because four steps shouldn't need a consultant to explain.
Step 1: Verify (Every Factual Claim)
Every factual assertion in an AI-generated document gets checked against a primary source. Every citation gets looked up. Every statistic gets confirmed. Every name, date, and figure gets validated.
For legal citations: Westlaw, LexisNexis, or the court's own records. For academic citations: the journal database. For regulatory citations: the agency's published guidance. No exceptions. No spot-checking. Full verification.
Step 2: Validate (Against Domain Standards)
A domain expert reviews the entire document for logical coherence, appropriate application of cited authorities, and accuracy within the professional standards of the field. Verification confirms the facts exist. Validation confirms they're used correctly.
For legal documents, the validator must be a licensed attorney familiar with the subject matter. For medical documents, a qualified medical professional. The AI might cite a real case that says exactly the opposite of what the document claims. Verification catches fake citations. Validation catches misapplied real ones.
Step 3: Attribute (Document the AI's Role)
Record which portions of the document were AI-generated, which model produced them, what prompts were used, and what human modifications were made. This creates accountability and enables pattern analysis. If one model consistently hallucinates citations in employment law but performs well in contract review, you need that data.
Step 4: Certify (Personal Accountability)
A named individual signs off on the final document, certifying that all four steps were completed. Their name goes on it. Their reputation backs it. Their license is on the line.
This is the step that creates the behavioral incentive. When your name is attached to a certification that you personally verified the AI's output, the temptation to skip verification disappears. It's one thing to submit an unverified AI draft. It's another to sign a certification that you verified it when you didn't.
Implementation Requirements
Your organization needs a written policy that specifies:
- Which document types require VVAC
- Who is qualified to perform each step
- What primary sources are acceptable for verification
- Maximum turnaround time for each step
- Consequences for skipping or falsifying any step
- Documentation and retention requirements
Post the VVAC framework where your team works. Train on it quarterly. Audit compliance monthly.
The Stakes Keep Rising
The Wisconsin DA case is the highest-consequence AI hallucination incident in the legal system so far. The emphasis is on "so far."
As AI tools become more widely adopted in high-stakes environments, the consequences of unverified output will scale with the stakes. Criminal prosecutions dismissed today. Medical misdiagnoses tomorrow. Structural failures next.
The time to build verification processes is before the incident. Not after. The VVAC framework takes 30 minutes to implement and costs nothing. The alternative costs careers, cases, and in some industries, lives.
---
Kaizen AI Lab helps organizations implement AI systems with proper output verification, human review protocols, and compliance frameworks that prevent AI-generated errors from becoming organizational catastrophes.
Take the AI Compliance Readiness Assessment: acra.kaizenailab.com
Learn more: kaizenailab.com
Book a call: cal.com/dhoesq/kaizen