Key Takeaways
- The Florida Supreme Court amended Rule 2.515(d)(2) on May 28, 2026, requiring attorneys to certify the existence and accuracy of all cited legal authorities, with explicit sanctions for AI-generated false citations.
- The 11th U.S. Circuit Court’s July 10, 2026, reprimand of Anthony Sabatini for submitting briefs “replete with fake and hallucinated citations” illustrates the immediate disciplinary consequences attorneys face under Florida’s new rules.
- Florida Bar Ethics Opinion 24-1, approved January 19, 2024establishes that AI output falls under the attorney’s non-delegable supervision duty under Rule 4-5.3, meaning no AI tool can shift liability for a false filing. Florida attorneys who sign a court filing now certify, under sanction, that every cited authority actually exists. The Florida Supreme Court’s amendment to Rule 2.515(d)(2), which took effect June 15, 2026, was a direct response to a string of disciplinary cases involving fictitious AI-generated citations. Non-compliance carries consequences up to dismissal of proceedings and fee-shifting awards.
The Hallucination Problem
Generative AI models are language prediction engines, not fact retrieval systems. They produce text that sounds authoritative, and in legal contexts that means plausible-sounding case names, docket numbers, holdings and statutory language that simply do not exist. The 11th U.S. Circuit Court’s July 10, 2026, reprimand of Florida attorney Anthony Sabatini, whose briefs were found to contain multiple fabricated citations in an employment case, is the most prominent recent example. The Second District Court of Appeal separately referred an unnamed attorney to The Florida Bar for citing imaginary authorities, triggering formal disciplinary proceedings.
Florida Bar Ethics Opinion 24-1, approved January 19, 2024acknowledged AI’s utility in legal practice while making clear that the attorney’s duties of competence under Rule 4-1.1 and supervision of non-lawyer assistants under Rule 4-5.3 are non-delegable. When AI functions as a non-lawyer assistant, its output requires at least the same scrutiny as work product from a junior associate. The opinion explicitly warns against fabricated citations and misquoted law. Under the amended Rule 2.515(d)(2), sanctions now include striking documents, dismissing proceedings and awarding attorneys’ fees, giving courts a uniform statewide enforcement mechanism. The broader compliance exposure for firms that deploy AI without adequate governance extends well beyond Florida.
Vetting AI Tools
Before integrating any AI tool into legal workflows, firms need to conduct genuine due diligence, not just a review of vendor marketing materials.
Start with data retention and confidentiality policies. Rule 4-1.6 of the Florida Bar Rules requires protection of client confidentiality, and many cloud-based AI tools use input data for model training. That creates real exposure. Firms should confirm that vendor terms include enforceable confidentiality agreements and data isolation protocols. On-premise deployment or secure, isolated cloud instances are worth the additional cost for any client-sensitive work.
Go beyond vendor claims when assessing a tool’s actual accuracy. Request independent audits or benchmark results that specifically test performance on legal research tasks, not general-purpose language benchmarks. Direct engagement with the vendor’s technical team about training data, model architecture and documented error rates will surface information that a product demo will not. Some legal AI platforms, including those integrated with LexisNexis and Westlaw pull citations directly from verified legal databases rather than generating them, which meaningfully reduces hallucination risk compared to general-purpose large language models, though it does not eliminate it.
Run a controlled pilot before firm-wide adoption. Test the tool on hypothetical cases or non-confidential matters where correct outcomes are already known. Assign attorneys and paralegals to document error patterns, particularly instances of fabricated or misrepresented citations. That internal data shapes the verification protocols the firm will need to build.
Verification: The Non-Negotiable Steps
This is where compliance with Rule 2.515(d)(2) is won or lost. Every legal authority generated or cited by an AI tool must be cross-referenced against its original primary source, physically located in a reputable legal database or official reporter. Do not rely on hyperlinks the AI provides; they can point to incorrect or non-existent pages. Verify that the cited text, holding and factual context actually support the proposition for which the authority is offered. The rule’s language is specific: authorities must “exist and are accurately cited.”
Scrutinise quotations with particular care. AI models are prone to paraphrasing that subtly shifts meaning, or to fabricating quotes outright. Any direct quotation must be matched word-for-word against the original source. Case summaries require the same treatment: confirm that the AI’s distillation reflects the court’s actual reasoning and holding, not a plausible-sounding reconstruction of it.
Run conventional citation-checking software as a final automated pass: Shepard’s on LexisNexis, KeyCite on Westlaw. These tools check precedential value and citation format, and can flag non-existent case numbers or incorrect reporter references. They are not a substitute for primary-source verification, but they catch a category of mechanical errors that human review can miss under time pressure. The output review bottleneck is a known pressure point in AI-assisted workflows; building in automated checks reduces the load on attorneys doing final sign-off.
Establish a mandatory peer review requirement for any filing that incorporates AI-assisted content. A second attorney or experienced paralegal should independently review citation accuracy and the fidelity of legal arguments to the authorities cited. The Florida Bar’s competence and supervision standards effectively require this; building it into workflow policy makes the obligation explicit and auditable.
Firm-Wide Policy and Governance
Verification workflows are only as reliable as the policies that mandate them. Firms need written AI usage policies that specify permissible tools, mandatory verification steps and the confidentiality requirements under Rule 4-1.6. Those policies should reference Rule 2.515(d)(2) and Ethics Opinion 24-1 directly and define consequences for non-adherence. Restricting client-specific legal work to approved platforms, and prohibiting the use of general-purpose AI tools for that work, is a practical baseline.
Training cannot be a one-time event. All legal professionals, from partners to paralegals, need grounding in how hallucination actually works, what the firm’s verification protocols require and what the ethical obligations are. Practical exercises, identifying AI-generated errors in mock filings, are more effective than passive instruction. As AI tools evolve and new ethics guidance emerges from The Florida Bar, training content will need to follow.
Designate a partner or senior attorney to own AI governance. That person, or a small committee, should track amendments to Florida Bar rules, evaluate new tools against the firm’s compliance requirements, update internal policies and confirm that training stays current. Centralised oversight prevents the fragmented adoption patterns that tend to produce the verification gaps behind most disciplinary referrals.
The Sabatini reprimand and the Second District referral are not cautionary edge cases. They are the foreseeable consequence of deploying AI tools without verification discipline. Florida’s amended Rule 2.515(d)(2) removes any ambiguity about where responsibility sits: with the attorney who signs the filing, not the tool that helped draft it.
Originally published at https://autonainews.com/how-to-verify-ai-output-to-avoid-florida-rule-2-515d2-sanctions/
