Brooks v. Lowe’s: Attorney Fined $1,000 After Trusting Claude AI to Fix Its Own Hallucinations — Without Human Review | Advocate Prakhar

Case at a Glance

Full Citation	Brooks v. Lowe’s Home Centers, LLC (Wilkins Sanctions)
Court	United States District Court (district not specified in excerpt), May 18, 2026
Date	May 18, 2026
Category	AI Hallucination & Sanctions
Jurisdiction	USA (Wisconsin)
AI Tool Used	Claude AI (Anthropic)
Outcome/Sanction	$1,000 personal sanctions against attorney Paul Wilkins; must be paid personally, cannot be reimbursed by law firm

Background

Paul Wilkins represented a plaintiff in an employment case against Lowe’s Home Centers. He used the Claude AI platform to draft a brief. A human law clerk reviewed the output and discovered that Claude had hallucinated quotations. Wilkins then confronted Claude with the identified errors — and trusted Claude alone to correct them, without human verification of the corrections. Claude generated new fabrications in response. Wilkins filed that second output into the record without review. The parties reached a settlement, but the court still addressed Wilkins’s AI misconduct.

The AI Issue

The central question was not merely whether AI was used — the court acknowledged AI use is not inherently improper — but whether the attorney’s process for using AI met a reasonable standard of professional care. Wilkins knew Claude had hallucinated (his law clerk found the errors). He had the knowledge and the opportunity to verify the corrections. Instead, he trusted the AI to self-correct and filed the result without any human review of the corrections. The court found this was fundamentally incompatible with professional responsibility.

What the Court Decided

• Using AI to draft legal briefs is permissible — but the attorney remains personally responsible for the accuracy of every citation and quotation [AI use policy].
• Knowing that an AI has hallucinated in a draft and then trusting that same AI to self-correct, without human verification, is an independent and serious failure of professional judgment [compounded negligence].
• Attorney Wilkins is personally sanctioned $1,000 — and the sanction must be paid from personal funds, not reimbursed by his law firm [personal accountability].
• The AI self-correction problem illustrates a specific danger: AI systems when prompted to fix errors may generate plausible-looking alternative fabrications rather than accurate corrections [AI self-correction failure mode].

“Mr. Wilkins then confronted Claude with the identified errors and entrusted Claude to correct them. Instead, Claude just made up more stuff. Mr. Wilkins filed that second output into the record without review.”

— United States District Court, Brooks v. Lowe’s, May 18, 2026

The India Angle

Indian Law Equivalent

Under the Advocates Act, 1961, an advocate who supervises the preparation of court documents — whether by a junior, a paralegal, or an AI tool — remains personally responsible for the accuracy of those documents. Rule 18 of the Advocates Act empowers Bar Councils to address professional misconduct including negligent preparation of court filings. The specific failure mode here — trusting AI to self-correct without human verification — is analogous to signing a brief without reading it, which is directly prohibited by BCI conduct rules.

Bar Council Rules

Bar Council of India Rule 15 (fair conduct), Rule 22 (no false statements to court), and the general duty of competence implied by Rule 14 (upholding court’s dignity) all apply. The Brooks case adds a specific lesson about AI self-correction: an advocate who receives a report of AI hallucinations in a draft and then delegates correction to the same AI tool — without personal verification — has not fulfilled the duty to review before filing.

Practical Advice for Indian Advocates

• Never trust AI to correct its own hallucinations — when your AI tool identifies or is told about an error, verify the corrected version yourself using an independent legal database before filing.
• Establish a two-person verification protocol for AI-assisted briefs: the drafter using AI and a separate reviewer who checks every citation independently, without reference to the AI’s output.
• If AI corrections are necessary after initial review, treat the corrected version as a new first draft requiring the same level of verification as the original — not as a finished product.

Quick Takeaways

• Trusting AI to self-correct its own hallucinations without human review is a sanctionable failure of professional judgment.
• Personal liability for AI misconduct cannot be shifted to your law firm — courts can require personal payment.
• The AI self-correction failure mode — generating new fabrications when prompted to fix old ones — is a documented and dangerous pattern.

---

Deep Dive: The Self-Correction Trap: Why Asking AI to Fix AI Hallucinations Is Dangerous

Brooks v. Lowe’s introduces a failure mode that lawyers using AI tools need to understand specifically: the self-correction trap. When Wilkins discovered (through his law clerk) that Claude had hallucinated quotations, he faced a choice: verify the corrections himself, or ask Claude to fix them. He chose the latter. Claude generated what appeared to be corrections. But without independent verification, there was no way to know whether the ‘corrections’ were accurate or simply new hallucinations dressed up as fixes. He filed the result without checking. The court’s summary of what happened — ‘Claude just made up more stuff’ — is a precise description of a well-documented AI behavior.

Large language models, when told they made an error, do not actually go back and verify the correct information. They generate a new response that is more likely to satisfy the feedback they received. If told ‘this case citation is wrong,’ an LLM will produce a different citation that looks more like a real case — but it may still be fabricated. The LLM has no memory of what is actually in legal databases; it only knows what pattern of text is statistically likely to follow from the prompt. This means that asking an AI to self-correct hallucinated citations is not a verification step — it is a second roll of the same dice, with similar odds of another hallucination.

The court’s decision to require personal payment — not law firm payment — of the $1,000 sanction reflects a principle appearing in several 2026 AI sanctions cases: when the misconduct results from an individual attorney’s failure of professional judgment, rather than a systemic firm-wide failure, the financial consequence should fall on that individual. This creates a genuine personal incentive to take the verification step seriously. If law firms routinely absorbed AI-sanction costs, individual attorneys would face no personal financial consequence for shortcuts in AI use.

For Indian advocates, the Brooks case is a practical guide to what not to do when AI errors are found mid-process. The correct protocol when a law clerk or colleague identifies an AI hallucination in a draft is: (1) remove the hallucinated content entirely, (2) search for the correct authority independently using Manupatra, SCC Online, or IndianKanoon, (3) manually verify that the correct authority says what you need it to say, and (4) insert the verified citation. At no step in this process should the AI be re-queried with instructions to ‘fix’ the error — the AI is not a reference database, and its self-corrections cannot be trusted without independent verification.