Shadow AI in Legal: The Hidden Risk Threatening Attorney-Client Privilege, Bar Licenses, and Client Trust

The Brief That Never Existed

In the spring of 2026, attorneys Van R. Irion and Russ Egli submitted briefs to the Sixth Circuit Court of Appeals in the case of Whiting v. City of Athens, Tennessee. The briefs contained more than two dozen citations. The court found that the cited cases “did not exist, did not include the quoted language claimed, or did not discuss or support the proposition for which they were cited.”

Both attorneys were sanctioned $15,000 each—$30,000 combined—the stiffest financial sanction the court could impose. The matter was forwarded to the chief judge for potential disciplinary proceedings.

The citations had been generated by AI. Neither attorney had verified them before filing.

This is not an isolated incident. It is one entry in a database that now tracks over 1,369 legal decisions globally involving AI-generated hallucinations—fabricated cases, invented quotations, citations to real cases that stand for the exact opposite proposition claimed. The count was 200 a year ago. It grows by five to six new documented cases every day. And that database, maintained by researcher Damien Charlotin at HEC Paris, only captures what courts caught. Fabricated authorities pass undetected constantly, especially in matters that settle or where opposing counsel lacks the resources to verify every citation.

What Shadow AI Looks Like in a Law Firm

Shadow AI in legal practice refers to any use of AI tools by lawyers, paralegals, legal assistants, or other firm personnel without the knowledge, approval, or oversight of the firm’s IT or compliance leadership. In a law firm context, it is pervasive and takes many forms:

• A partner drafting a client memo by pasting a confidential litigation strategy into ChatGPT for faster summarization
• An associate using a personal AI subscription to generate a first draft of contract language, without disclosing the tool to the firm
• A paralegal feeding deposition transcripts into an unapproved AI platform to generate summaries—uploading protected client communications to a third-party server in the process
• A litigation team using a consumer AI tool for case research, trusting its citations without verification
• A junior attorney inputting client financial records into an AI document analysis tool that has never been vetted by the firm’s data security team
• A legal secretary using a free AI writing tool to draft correspondence containing privileged information

According to the 2025 Clio Legal Trends Report, 79% of legal professionals use AI tools in their work—but 44% of law firms have no formal AI governance policy in place. Thomson Reuters research found that only 30% of law firms currently have a specific AI policy. The gap between adoption and governance is where every one of the profession’s most serious AI risks lives.

The Hallucination Crisis: By the Numbers

The most visible Shadow AI problem in the legal profession is hallucination—AI generating confident, plausible-sounding legal citations that simply do not exist.

• Over 1,369 legal decisions globally have been documented involving AI-hallucinated content submitted to courts (Charlotin AI Hallucination Cases Database, 2026)
• The figure includes over 1,227 fabricated citations as of early 2026—a number that was 200 just one year prior
• 518+ US cases since January 2025, growing at 30–50 new documented cases per month
• General-purpose LLMs hallucinate on approximately 58% of federal case research queries (GPT-4); Llama 2 hallucinated on approximately 88%—meaning the free tools most commonly used in Shadow AI scenarios are wrong more often than they are right (Stanford/Yale Research, Dahl et al., 2024)
• Even legal-specific AI tools with RAG architecture—specifically designed to ground outputs in verified sources—hallucinate 17–34% of the time
• Every practice area has been affected: personal injury, commercial litigation, family law, bankruptcy, employment, immigration, IP, and consumer protection
• Who has been caught: solo practitioners, mid-size firms, Am Law 100 firms including Gordon Rees and Boies Schiller, and even a federal DOJ attorney

Real Incidents: What Shadow AI Has Cost Lawyers and Their Clients

⚖️ Whiting v. City of Athens — Sixth Circuit Sanctions Two Attorneys $30,000 (March 2026)

The Sixth Circuit’s most significant AI sanctions order to date. Attorneys Van R. Irion and Russ Egli submitted briefs containing more than two dozen fake or misrepresented citations across three consolidated appeals. The court imposed $15,000 per attorney in punitive fines payable to the court registry, plus full reimbursement of the opposing party’s attorney fees on appeal, and referred the matter to the chief judge for potential disciplinary proceedings.

💰 Valley View Winery — $110,000 Sanction, Case Dismissed With Prejudice (2026)

The single most expensive AI hallucination sanction in US legal history. US Magistrate Judge Mark D. Clarke imposed $110,000 in combined fines and attorneys’ fees against two lawyers in a winery ownership dispute in Jacksonville, Oregon. Three court filings contained 23 fabricated legal citations and 8 false quotations generated by AI. Judge Clarke ordered the lawsuit dismissed with prejudice. The client—who had been suing her brothers for control of a family winery and seeking $12 million—lost her case entirely because her attorneys submitted AI-generated citations they never verified.

🔄 Gordon Rees Scully Mansukhani — Three Incidents in Six Months (Oct 2025–Feb 2026)

Gordon Rees, an Am Law top-100 firm with $759 million in gross revenue in 2024, became the defining repeat-offender case of the current enforcement wave. The firm submitted a bankruptcy filing containing “inaccurate and non-existent citations” in October 2025. After receiving sanctions and promising updated AI policies, two months later Magistrate Judge Carolyn Delaney issued a formal reprimand in a California federal court. In February 2026, despite prior sanctions, policy updates, and public embarrassment, the firm allegedly submitted yet another brief containing fabricated authority. Three incidents at a major firm in six months illustrates precisely why policy updates without workflow controls do not prevent Shadow AI incidents from recurring.

🏥 Lacey v. State Farm — $31,000 in Fees Against Ellis George and K&L Gates (May 2025)

Special Master Michael R. Wilner imposed approximately $31,000 in fees against two major law firms, jointly and severally, after approximately 9 of 27 citations in a supplemental brief were wrong and at least 2 cases did not exist. Both Ellis George and K&L Gates—prominent, well-resourced firms—were caught submitting AI-generated citations that had never been verified against primary sources.

⚠️ Wadsworth v. Walmart — Morgan & Morgan Attorney Sanctioned, Pro Hac Vice Revoked (Feb 2025)

Attorneys at Morgan & Morgan—the largest personal injury firm in the United States—were sanctioned after attorney Rudwin Ayala used the firm’s own internal AI tool, “MX2.law,” to generate case law citations for motions in limine, producing eight fabricated cases. Judge Kelly H. Rankin imposed sanctions of $3,000 on Ayala, $1,000 each on two senior partners, and revoked Ayala’s pro hac vice admission. The judge stated: “A finding of subjective bad faith is not required to impose sanctions.” This case is significant because the tool involved was the firm’s own approved internal AI—even governed, purpose-built legal AI tools can produce hallucinated citations without proper human verification.

🚫 DOJ Attorney Fired — Federal Government Not Immune (Late 2025)

An Assistant US Attorney in the Eastern District of North Carolina was found to have submitted a government brief containing fabricated quotations attributed to appellate decisions and the Code of Federal Regulations. Magistrate Judge Robert T. Numbers II issued a sharply worded order finding fabricated quotations, misstatements of case holdings, and false or misleading statements about how the errors occurred. The attorney was subsequently fired. The federal government itself was not immune to the consequences of unverified AI use.

⛔ Colorado Suspension — Attorney Suspended for Lying About AI Use

Attorney Zachariah Crabill received a 90-day bar suspension—not merely a fine—after filing fabricated ChatGPT citations in a custody case and then lying to the judge about their origin. The lie transformed an AI hallucination mistake into a candor violation, which courts treat as a categorically more serious offense. The case establishes that the cover-up compounds the original Shadow AI misconduct.

Beyond Hallucinations: The Deeper Shadow AI Risks for Legal

1. Attorney-Client Privilege Destruction

Attorney-client privilege is one of the most powerful protections in the law, shielding confidential communications between lawyers and clients from disclosure in litigation. That protection depends on the communication remaining confidential.

When an attorney pastes client communications, litigation strategy, settlement positions, or legal advice into a consumer AI tool through a personal account, that information is transmitted to a third-party server the attorney does not control. The AI provider is not a party covered by attorney-client privilege. The result: privileged information has been voluntarily disclosed to a third party outside the attorney-client relationship, which can waive the privilege—potentially permanently.

California recently passed a bill restricting the use of generative AI in regulated professions including law, with a central provision prohibiting attorneys from entering any confidential, personal identifying, or nonpublic information into a public generative AI system, specifically to prevent breaches of attorney-client privilege.

2. Model Rules Violations (ABA Formal Opinion 512)

The ABA’s Formal Opinion 512, issued July 29, 2024, clarifies how existing Model Rules apply to AI:

• Rule 1.1 — Competence: Lawyers must understand AI capabilities and limitations, including hallucination rates. A lawyer using a consumer tool that hallucinates 58% of the time without understanding that risk is failing their duty of competence. AI literacy is now a professional obligation.
• Rule 1.6 — Confidentiality: Entering client data into a consumer AI platform without a data processing agreement or enterprise privacy guarantee is a direct violation. Reasonable efforts to prevent inadvertent disclosure are required.
• Rule 3.3 — Candor to the Tribunal: Lawyers must review AI outputs and correct errors before filing. Submitting a hallucinated citation is a false statement of law under Rule 3.3(a)(1), regardless of whether the lawyer knew it was false. Ignorance is not a defense.
• Rule 5.3 — Supervision: Supervising attorneys are responsible for AI-generated work product signed under their names. The Lacey v. State Farm and Wadsworth v. Walmart cases both imposed sanctions on supervising partners who signed briefs without adequately reviewing AI-generated content.

As of early 2026, over 35 state bar associations have issued formal guidance on AI use, with requirements that vary materially by jurisdiction.

3. Work Product Doctrine Exposure

Work product protection shields attorney mental impressions and legal theories from opposing counsel. When those “mental impressions” are generated by an AI system operating outside the firm’s secure environment, the doctrine’s protection becomes uncertain. Courts have not yet fully addressed whether AI-generated work product retains the same protection as human-generated work product—but the argument that an AI system’s output is not the attorney’s mental impressions is one opposing counsel will make.

4. Client Data Exposure

Legal clients share among the most sensitive information in their lives: pending divorces, criminal matters, business disputes, health records, immigration status, financial difficulties, and trade secrets. When law firm employees use unauthorized AI tools to process this information, it flows to third-party servers with no client consent, no disclosure, and no governance. The firm’s engagement letter almost certainly does not mention AI data processing. This creates potential liability under GDPR for EU clients, state data breach notification laws, and emerging AI-specific legislation—independent of whether a breach ever occurs.

5. Malpractice and Client Harm

When a client’s case is damaged by an AI hallucination—lost due to sanctions, dismissed with prejudice, or weakened by opposing counsel’s discovery of fabricated citations—the client suffers real consequences. The Valley View winery case is the starkest example: a woman’s $12 million case dismissed entirely because her attorneys submitted AI-generated citations they did not verify. Professional liability insurers are beginning to update underwriting requirements to address AI governance documentation.

6. The Shadow AI Ban Paradox

Many firms have responded to AI risk by banning AI tools outright. This reliably backfires. When firms ban AI without providing approved alternatives, lawyers under pressure turn to free, consumer-grade tools on personal devices—losing all firm visibility into where client data is going. New lawyers entering firms today arrived AI-native; AI is embedded in their research workflows from law school. Banning these tools is not a compliance strategy—it is a Shadow AI creation strategy.

The Regulatory and Disclosure Landscape: Navigating 35+ Sets of Rules

• Federal courts: No uniform rule. Individual judges increasingly issue standing orders requiring AI disclosure before filing. The SDNY and EDTX (Local Rule AT-3) are among those with specific requirements.
• State courts: Highly variable. Illinois, California, New York, Delaware, and 30+ other jurisdictions have issued AI-specific policies or are proposing them. The patchwork grows monthly.
• ABA Formal Opinion 512 (2024): Establishes that competence requires understanding AI capabilities and limitations, that confidentiality rules apply to AI data sharing, and that candor requires human verification of AI outputs before submission.
• Emerging affirmative duty: Courts are beginning to suggest attorneys may have an affirmative duty to flag AI hallucinations in opposing counsel’s briefs when discovered—transforming AI verification from a self-protective practice into a professional obligation toward the court.

Why Legal Professionals Turn to Shadow AI

• Billable hour pressure is constant. AI tools that generate a contract summary in 30 seconds or produce a first-draft research memo in minutes are nearly impossible to resist when the alternative is hours of billable work.
• Approved tools are slow to arrive. Legal IT procurement is lengthy. Security review, data processing agreements, partner approval, and training requirements can span months. By then, associates have been using the consumer version for a year.
• Law school graduates arrive AI-native. The Class of 2026 was exposed to Lexis and Westlaw’s generative AI tools during their first year. They arrived at firms already using AI for research and document drafting.
• In-house counsel is driving adoption. Corporate clients increasingly ask whether outside counsel uses AI tools for efficiency—and question whether firms not using AI are providing cost-effective representation. The pressure to adopt comes from the client relationship as well as from within the profession.

What Law Firms Must Do in 2026

Step 1: Create a Realistic AI Policy — Not a Ban

A defensible AI policy for law firms must: specify which AI tools are approved for which categories of work; define what client and matter information may and may not be entered into AI systems; address data residency and privacy requirements; require Business Associate Agreements or Data Processing Agreements with any AI vendor that may process client data; establish mandatory human verification requirements for AI-generated citations before filing; and include disclosure obligations aligned with each jurisdiction where the firm practices.

Step 2: Treat Citation Verification as Non-Negotiable

Every AI-generated legal citation must be verified against primary sources—Westlaw, Lexis, the court’s official records—before any document is filed. Even the best legal-specific AI tools hallucinate 17–34% of the time. Consumer tools hallucinate on more than half of legal research queries. Practically, this means:

• Requiring attorneys to document that every citation in an AI-assisted filing has been independently verified
• Building citation verification into matter workflow, not leaving it as an individual attorney’s judgment call
• Logging AI tool use in matter files so supervisors and risk management can assess exposure
• Building verification checklists into filing workflows that cannot be bypassed under deadline pressure

Step 3: Protect Client Confidentiality at the Tool Level

No client matter information should enter any AI system that has not been evaluated and approved by the firm’s data security team. This requires: formal vendor evaluation specifically assessing data retention and training data use; enterprise agreements prohibiting use of firm data for model training; DLP tools that detect when privileged information is being transmitted to unauthorized AI platforms; and clear staff guidance on exactly what categories of information may enter which systems.

Step 4: Build AI Literacy into Professional Development

ABA Formal Opinion 512 establishes that competence includes understanding AI capabilities and limitations. Effective AI training for legal professionals must cover: how AI hallucination works and why legal research is particularly susceptible; specific verification steps for each type of AI-assisted work; confidentiality rules that apply to AI data sharing; current AI disclosure requirements in each jurisdiction where the firm practices; and practical demonstrations of AI hallucination to calibrate trust in AI outputs.

Step 5: Address the Supervision Gap

Rule 5.3 makes clear that supervising attorneys are responsible for AI-generated work product signed under their names. Partners and senior associates need explicit guidance on their supervision obligations in AI-assisted work. A supervising attorney who signs a brief without verifying AI-generated citations has, in the eyes of the court, personally vouched for those citations.

Step 6: Monitor Disclosure Requirements Continuously

The patchwork of AI disclosure rules across federal and state courts is expanding rapidly. Firms need a process for tracking new standing orders, local rules, and bar guidance as they are issued. Failure to disclose AI use when required creates a candor violation independent of whether the AI-generated content was accurate.

Quick Reference: Shadow AI in Legal at a Glance

Sources: Damien Charlotin AI Hallucination Cases Database, PlatinumIDS Legal AI Analysis, Clio Legal Trends Report 2025, Thomson Reuters Generative AI in Professional Services 2025, NC Bar Association AI Policy Guidance 2026, ABA Formal Opinion 512 (2024), NexLaw AI Hallucination Sanctions Guide 2026, Stanford/Yale Dahl et al. 2024.