Brigandi Case: How a $110,000 AI Hallucination Sanction Rewrites Risk for Legal AI Systems

When two lawyers in Oregon filed briefs packed with fake cases and fabricated quotations, the result was not a quirky “AI fail”—it was a $110,000 sanction, dismissal with prejudice, and a public ethics disaster. [1][5]

For ML and platform engineers, the Brigandi matter is a concrete signal: if your system can move unverified model output into court-facing documents, your organization is in the blast radius. [1][5]

💼 Engineering lens: Treat this case as an incident postmortem on an entire socio-technical stack—model, UX, validation, logging, and governance—not just a story about one careless prompt.

---

1. What Actually Happened in the Brigandi Case (and Why Engineers Should Care)

U.S. Magistrate Judge Mark D. Clarke sanctioned San Diego attorney Stephen Brigandi and Portland attorney Tim Murphy a combined $110,000 for filing AI-assisted briefs that included 15 non-existent cases and eight fabricated quotations. [1][6]

Key facts:

• Judge Clarke called it “a notorious outlier in both degree and volume” of AI misuse and faulted plaintiffs and counsel for not being “adequately forthcoming, candid or apologetic.” [1][6]
• The dispute involved the Valley View winery in Oregon: Joanne Couvrette sued her brothers for control, alleging elder abuse and wrongful enrichment and seeking $12 million. [1][5][6]
• Brigandi, not licensed in Oregon, worked with Murphy, who appeared procedurally; both were sanctioned because they signed filings that put AI-generated citations into the federal record. [1][3]
• The case was dismissed with prejudice; the briefs were “replete with citations from non-existent cases,” and the court noted evidence of a “cover-up” when false references were deleted and refiled without disclosure. [4][5][6]

⚠️ Key shift: This is now a concrete example of how unverified LLM outputs in a regulated workflow can create direct financial liability and reputational damage for anyone deploying such tools. [1][5]

---

2. Where AI Hallucinations Enter Legal Workflows

The technical failure is familiar to anyone working with large language models: when asked for supporting authority, the model confidently produced plausible-looking but fake citations and quotations. [1][9]

How hallucinations got into the briefs:

• The filings were described as “replete with citations from non-existent cases,” suggesting use of AI as an authority generator, not as a retrieval-first assistant. [5][8]
• Judge Clarke noted that an AI tool “once again led human minds astray,” reflecting a misaligned mental model: lawyers treated outputs as authoritative legal text, while the model only sampled likely tokens. [5][7]

💡 Architectural anti-pattern: Letting an LLM fabricate structured legal objects—case names, reporter citations, docket numbers—without deterministic validation is fundamentally unsafe in law and similar domains.

Common risky prompts:

• “Find cases that say X” without retrieval.
• “Fill in” missing citation details from memory.
• Trusting model summaries of cases it just invented.

Without retrieval-augmented generation (RAG) over authoritative case law, strict schema validation, and live lookups to legal databases, even strong models will confidently hallucinate rare or non-existent precedents, especially on niche issues. [9]

📊 Implication: Production legal tools must treat the LLM as a language layer over a verifiable database of law, never as a standalone source of truth for anything that might be filed in court. [5]

---

3. Designing Verification-First Architectures for Legal Citations

The Oregon sanctions flowed directly from non-existent cases being presented as real. Any serious legal AI system must treat “every cited authority exists and is correctly referenced” as a hard invariant. [4][9]

A robust division of labor:

• Retrieval-only for authorities. Cases, statutes, and regulations come only from a vetted corpus or commercial provider.
• LLM-only for narrative. The model summarizes and reasons over retrieved materials but never invents citations or alters reporter identifiers.

Implementation patterns:

• Parse every citation the model emits.
• Normalize it (e.g., Bluebook-style fields) into structured objects.
• Cross-check against a legal database API; unresolved citations are blocked or clearly flagged.

💡 Schema-first output

Use structured outputs (JSON/XML) such as:

{
  "argument_sections": [...],
  "citations": [
    {
      "id": "doc_123456",
      "case_name": "Smith v. Jones",
      "reporter": "F.3d",
      "volume": 999,
      "page": 123
    }
  ]
}

Validate doc_123456 against your authority index before rendering a formatted brief.

For Brigandi-style workloads, a pre-submission gate should hard-block export if even a single citation fails validation, forcing manual review before anything leaves the system. [1][5]

⚡ Containment, not perfection: These guardrails do not stop the model from hallucinating internally, but they ensure fabricated content cannot cross the system boundary into actual court filings.

---

4. Governance, Logging, and Accountability in High-Risk Domains

Judge Clarke criticized the plaintiffs and their counsel for lacking candor and highlighted an attempted cover-up once the bogus citations were exposed. [1][4]

He also noted circumstantial evidence that Couvrette herself may have generated some AI drafts, but held the attorneys responsible because they signed the filings. [5][6]

For engineering teams, this demands a trustworthy audit trail showing who did what, with which tool, and when.

Minimum logging for a legal AI platform:

• User identity and role.
• Model version and tool configuration.
• Prompt templates and raw prompts.
• Full prompt–completion pairs for any court-facing draft.

Role-based controls and workflow constraints:

• Require human review and sign-off for any filing-ready document.
• Persistent UI disclaimers that outputs are drafts requiring independent verification.
• Restrict high-risk features (e.g., authority generation) to trained users.

📊 Risk monitoring: Build alerts for:

• Unusually high numbers of new authorities in a single matter.
• Repeated citation-validation failures.
• Users bypassing suggested review paths.

These governance and observability practices allow organizations, when AI errors occur—as in the Oregon vineyard lawsuit—to show process discipline rather than negligence. [5][10]

---

5. Implementation Blueprint: Safer Legal AI Systems After Brigandi

In Brigandi, hallucinations produced case-ending sanctions and a six-figure penalty that dwarfed prior Oregon appellate sanctions, where the largest had been $10,000. [1][5][6]

Legaltech engineers should assume similar exposure wherever unverified AI text can reach a court, regulator, or opposing counsel, and ensure filing-ready documents emerge only after checks and human review.

A pragmatic stack:

• Vector database over vetted opinions (e.g., Elasticsearch, Qdrant, pgvector) powering RAG for case discovery.
• Authority index keyed by citation and document ID for deterministic lookup.
• LLM layer limited to summarization, comparison, and reasoning over retrieved documents.
• Validation service that inspects drafts, resolves every citation, and blocks or annotates unresolved references.

To help stakeholders visualize this, it is useful to model the end-to-end workflow from first draft to filing, showing exactly where retrieval, validation, and human review prevent hallucinated citations from escaping into the record.

flowchart LR
    title Verification-First Legal AI Workflow to Prevent Hallucinated Citations

    A[Lawyer drafts] --> B[Query AI assistant]
    B --> C[Retrieve corpus]
    C --> D[LLM drafts narrative]
    D --> E[Validate citations]
    E --> F{Unresolved cites?}
    F -- Yes --> G[Manual review]
    F -- No --> H[Court filing]

    style C fill:#3b82f6,color:#ffffff
    style E fill:#22c55e,color:#ffffff
    style F fill:#f59e0b,color:#000000
    style G fill:#ef4444,color:#ffffff
    style H fill:#22c55e,color:#ffffff

💡 Evaluation under pressure

Before deployment, run offline tests where you:

• Prompt the model for obscure or adversarial citations.
• Force edge cases like “find a Ninth Circuit case that says X” when none exists.
• Push outputs through your verification pipeline and log residual hallucination rates.

Use results to set conservative thresholds—for example, no unverified citations in auto-export mode; drafts with unresolved items must be watermarked and limited to internal use.

To avoid Brigandi-style failures, roll out capabilities gradually:

1. Start with internal research memos and email summaries.
2. Move to low-stakes filings (routine discovery motions, status reports).
3. Only then enable AI-assisted drafting for dispositive motions or appellate briefs. [5][4]

⚠️ Documentation is part of the product

Maintain clear, versioned documentation of:

• Model choices and training constraints.
• Guardrails and validation logic.
• Operational limits and recommended use cases.

If a judge or regulator later scrutinizes your tooling, you want to show the system was intentionally engineered to minimize hallucination-driven harm, not casually bolted onto billable workflows.

---

Conclusion: Designing for Hallucinations, Not Around Them

The Brigandi sanctions turn AI hallucinations from a modeling quirk into a quantified operational risk in legal practice: one incident, $110,000 in penalties, and a case dismissed with prejudice. [1][5]

The root failure was architectural: the model was treated as an authority instead of as a language layer on top of verifiable legal data.

A safer, verification-first design includes:

• Grounded retrieval from authoritative corpora.
• Strict citation validation and schema-constrained outputs.
• Mandatory human review before filing.
• Governance, logging, and monitoring that establish accountability.

⚡ Action step: If you design or operate legal AI tools, use this case as a checklist. Audit every path by which unverified authorities might escape your system, add retrieval and validation layers, and stress-test workflows with adversarial prompts long before they touch live matters or real clients.