A federal court in Tennessee's Western District recently sanctioned the Reaves Law Firm under Rule 11 after the firm failed to produce evidence that it had verified AI-generated legal citations before filing them in court. The sanction wasn't primarily about AI hallucinations; it was about the firm's inability to demonstrate that any verification process existed at all.
This case reveals a critical gap in how most organizations approach AI governance: the distance between having a policy and producing evidence that the policy was followed.
What the Court Expected
The court's show-cause order made three specific demands of Reaves Law Firm:
- Confirm whether cited cases existed and supported the propositions claimed.
- Verify whether quoted language actually appeared in the cited cases.
- Describe the verification process used before filing.
The firm produced an internal email titled "Mandatory Ethical AI Training & Reporting Protocols for All Staff." It couldn't demonstrate the email was distributed firm-wide or that any training occurred. When the firm pointed to a departed general counsel and "restructured" internal processes, the court rejected these explanations. Without contemporaneous records of verification, the firm stood exposed, not just for citing nonexistent cases, but for having no documented process to prevent it.
The court noted that the firm's namesake had signed the pleadings personally. Blaming departed personnel didn't eliminate responsibility. The sanctions followed.
Four Findings That Matter for Your Program
1. Policy Statements Are Assertions, Not Evidence
Many organizations now include language in AI policies requiring "human oversight" or "responsible use." These statements create expectations but don't prove compliance. The Reaves firm appears to have had such policies. What it couldn't produce was proof that anyone followed them for the specific decisions at issue.
2. Verification Processes Require Contemporaneous Records
The court specifically asked Reaves to describe its verification steps. An affidavit based on memory wouldn't have been persuasive, especially after personnel changes. Organizations that document verification as it happens stand in a fundamentally different position than those that attempt to reconstruct events after a challenge arrives.
3. Accountability Requires Named Owners
Reaves attempted to assign blame to a former general counsel. The court attributed responsibility to both the firm and the attorney whose name appeared on the filings. Your AI governance program needs identified individuals assigned to specific decisions, not general statements about oversight.
4. Warning Signals Demand Process Review
After the initial allegations of AI hallucinations, Reaves filed two additional pleadings containing similar defects. The firm received a signal and didn't stop to review its processes. Effective governance includes mechanisms to escalate concerns and suspend processes when problems surface.
What This Means for Your Compliance Program
Rule 11 has existed for nearly 90 years. The Reaves case doesn't create new obligations; it exposes how easily existing duties can be outsourced to AI systems and how difficult it becomes to demonstrate appropriate oversight after the fact.
Your organization likely faces similar pressures in different contexts. Regulators reviewing your data privacy practices, auditors examining your SOC 2 Type II controls, boards asking about AI risk management, or insurance carriers investigating a claim. Each represents a moment when someone will ask you to prove your governance program functions as described.
The question isn't whether you have an AI policy. It's whether you can produce evidence of compliance when challenged.
Action Items by Priority
Immediate: Identify Consequential AI-Assisted Decisions
Map where AI systems influence decisions that carry legal, financial, or reputational consequences. This includes:
- Customer communications that create contractual obligations.
- Risk assessments that determine pricing or eligibility.
- Compliance determinations that affect regulatory status.
- Financial analyses that inform material disclosures.
For each category, document what human review is required and by what standard.
Within 30 Days: Design Record-Keeping for Verification
Create templates or system fields that capture:
- What the AI system produced (input and output).
- Who reviewed the output and when.
- What judgment was exercised (what was checked, what standard was applied).
- What conclusion was reached (approved as-is, modified, rejected).
These records must be created contemporaneously, not reconstructed later.
Within 60 Days: Assign Named Accountability
Identify by name and role who owns verification for each consequential decision category. Update job descriptions and performance expectations accordingly. Ensure these individuals understand they may need to defend their decisions under scrutiny.
Ongoing: Implement Escalation Triggers
Define the signals that should prompt process review:
- Quality control failures or near-misses.
- Customer complaints about AI-generated content.
- Internal audit findings.
- Regulatory inquiries.
- Personnel changes affecting oversight roles.
When these signals appear, your program should include a documented pause, review, and correction process before resuming operations.
Quarterly: Test Evidence Production
Conduct internal exercises where you simulate a show-cause order or audit request. Ask: If we had to produce evidence of our verification process for the last 90 days, what would exist? Can we demonstrate compliance for specific decisions, or only assert general practices?
The Standard You'll Be Held To
The Reaves case demonstrates that courts expect organizations to produce contemporaneous evidence of verification processes, not after-the-fact explanations. This expectation will extend beyond legal filings to any domain where your organization claims human oversight of AI systems.
A governance program that can demonstrate Define, Record, Own, and Guard functions under pressure, with actual records, occupies fundamentally different ground than one that cannot. The time to build that capability is before the show-cause order arrives.
