Legal Ethics in AI

Defining AI Layers: Traditional, Generative, and Agentic

• Taxonomy of AI Categories: It is essential to distinguish between different forms of Artificial Intelligence to avoid conflating the technologies and their respective ethical risks.
  • Traditional AI: This form of AI has been established for a significant period and is integrated into various foundational technologies.
  • Generative AI (Gen AI): Emerged notably in $2022$ with the launch of ChatGPT. This is the primary driver of current legal ethical controversies and the technology most law firms are currently attempting to integrate or regulate.
  • Agentic AI: The next evolution of the technology, which larger corporate firms and in-house counsel are already transitioning toward.
• Geographic Adoption Trends: In regions like Charleston, firms often exhibit a slower, more conservative, and risk-averse approach to adopting Generative AI, whereas larger national or corporate firms are moving into the Agentic space faster.

Retrieval Augmented Generation (RAG) and the Persistence of Hallucinations

• Misconceptions Regarding Legal Databases: A common inaccuracy among legal professionals is the belief that using Retrieval Augmented Generation (RAG) within closed legal systems (e.g., Lexis, Westlaw, or Bloomberg) eliminates the risk of hallucinations. This is false; hallucinations still occur in legal-specific environments.
• Pattern Recognition Hallucinations in Case Law: Generative AI operates on pattern recognition rather than logical legal reasoning. In complex documents like Supreme Court cases, the AI may observe the facts/analysis followed by multiple dissents. It often mistakenly identifies the final dissent as the controlling ruling of the case because it is the concluding part of the pattern.
• Candor to the Tribunal: Under the rules of professional conduct, an attorney certifies the accuracy of every document filed with the court. An attorney is ethically responsible for any hallucinations or errors in a filing they sign, even if an associate or an AI tool produced the content.

Statistical Landscape of AI in Legal Practice

• Prevalence of Hallucination Cases:
  • There are currently over $1,000$ (specifically $1,092$ reported) hallucination cases worldwide.
  • Over $700$ (specifically $769$) of these cases are within the United States.
  • Every state in the U.S. has recorded filings involving hallucinations.
• Professional Learning and Education:
  • A $2025$ Thomson Reuters study indicates that most professionals are learning about AI voluntarily rather than through mandatory requirements.
  • Currently, no jurisdiction has a specific Generative AI training requirement, though Florida requires a general technology CLE (of which AI is a small component).
• Adoption Statistics (CLIO Study):
  • $59\%$ of legal professionals report widely adopting AI and believe it has a positive impact on revenue.
  • $40\%$ of legal professionals use legal-specific AI solutions (e.g., Protégé, Co-Counsel, Harvey).
  • $79\%$ of legal professionals use AI generally, but a significant portion uses generic, non-legal tools like ChatGPT, Gemini, Claude (referred to as "Plot"), or Perplexity.
• Small Firm and Solo Practitioner Risk: $64\%$ of firms with $2$ to $9$ attorneys and $62\%$ of solo practitioners use ChatGPT. These practitioners are at higher risk because they often serve as their own IT departments and may lack the staff to vet AI-generated content for hallucinations.
• Market Investment: Tech leaders have pledged over $5+ \text{ billion}$ dollars for AI education in $2025$.

Ethical Implications and the Model Rules of Professional Conduct

• Rule $1.1$ and Comment $8$: While Rule $1.1$ focuses on competent representation through legal knowledge and skill, Comment $8$ (adopted by many but not all jurisdictions) explicitly extends this to technology. There is an ongoing debate as to whether failure to utilize AI constitutes incompetence, or if the ethical failure is purely "bad lawyering" when an attorney fails to review an AI-assisted filing.
• ABA Formal Opinion $512$: The American Bar Association (ABA) recently released this opinion, asserting that existing model rules apply to Generative AI. Attorneys must consider AI's impact on:
  • Competent legal representation.
  • Confidentiality of client information.
  • Communication with clients.
  • Supervision of employees and third-party vendors.
  • Charging of legal fees.
• Engagement Letters: South Carolina attorneys are increasingly redrafting engagement letters to explicitly disclose the use of Generative AI. Sophisticated clients are responding by negotiating these clauses or striking them to limit AI use to specific administrative tasks rather than legal work.

Third-Party Vendors and Data Privacy

• Supervisory Responsibility (Rule $5.3b$): Since $2012$, this rule has included the supervision of third-party vendors and external resources. Attorneys are ethically responsible for how a vendor (e.g., an e-discovery platform) manages, saves, or trains on client data.
• Privacy Concerns with Generic Models:
  • ChatGPT Enterprise Policy: OpenAI's enterprise-level policy admits they may share personal information with third parties without notice for business objectives. They hold data for $30$ days and cannot guarantee that humans do not review the input.
  • Harvey: Currently regarded as the "gold standard" for legal AI because it is a self-contained, closed universe that does not train on outside information, making it safer for maintaining client confidentiality.
• Industry Consolidation and Mergers:
  • In July $2025$, Clio purchased Vlex (which had previously acquired Fastcase) for $1 \text{ billion}$ dollars.
  • Thomson Reuters acquired Co-Counsel to integrate it into Westlaw.
  • These mergers raise ethical questions regarding data access and cost-effectiveness when a firm's legal research tool is absorbed into a larger case management ecosystem.

Prompting Ethics and Risk Evaluation

• Prompt Engineering: The power of AI lies in dynamic or multi-piece prompting. However, the ethics of prompting starts with what data is included in the prompt.
• Risk-Level Categorization for Added Information:
  • No Risk: Publicly available case opinions.
  • High Risk: Client-confidential information.
• Redaction Strategies: Attorneys should use templates or redact sensitive information before inputting data into AI to mitigate privacy risks.
• Case Illustration (MUSC): Librarians at the Medical University of South Carolina (MUSC) discovered medical residents were inputting patient files into ChatGPT for efficiency. This necessitated emergency HIPAA training and the implementation of policies against the use of public AI for sensitive medical data.

The Global Tracking of Hallucination Cases and Sanctions

• Tracking Platform: A website managed by an attorney named Damian tracks hallucination cases world-wide.
• Sanction Trends:
  • "Falling on the Sword": Attorneys who admit to errors early (as seen in the Mata v. Avianca context, though spelled "Mata Ivaca" in discussion) generally face lower monetary damages and more opportunities to refile.
  • Passing the Buck: Law firms that blame associates or secretaries without having an established AI policy are more likely to receive large monetary sanctions, attorney's fee assessments, and referrals to state bars.
• Unapologetic Conduct: In cases where attorneys deny using AI despite obvious hallucinations, sanctions are significantly more severe, potentially leading to disbarment.

Recent Legal Developments and Unauthorized Practice of Law (UPL)

• South Carolina Interim Policy: South Carolina's policy distinguishes between AI and Generative AI. It contains sections specifically for attorneys that function as a "catch-all," echoing Formal Opinion $512$ by holding attorneys responsible for how Gen AI correlates with professional responsibility rules.
• Work Product vs. Attorney-Client Privilege: A New York court rejected a claim where a litigant argued their interaction with ChatGPT was privileged. The court ruled that ChatGPT is not an attorney and provides explicit warnings stating such, meaning no attorney-client relationship exists.
• New York Senate Bill: A proposed bill aims to address the Unauthorized Practice of Law (UPL) by chatbots, targeting tools that provide legal advice to pro se litigants.
• Pro Se Litigants and OpenAI Lawsuit: A significant case involves a pro se litigant who reopened a settlement (originally for $77$ units/dollars) based on ChatGPT's assessment that her previous attorney had "gaslighted" her. This led to a flurry of frivolous filings, prompting a life insurance company to sue OpenAI for the Unauthorized Practice of Law since ChatGPT functioned as the woman's "lawyer."

Questions & Discussion

• Q: Is company use of generative AI tools subpoenaable?
  • A: Yes. While companies like OpenAI are fighting these subpoenas (e.g., the Barat's case), they have been forced to provide some dialogue transcripts. Use of AI is potentially discoverable in litigation.
• Q: Was there litigation regarding the MUSC patient file situation?
  • A: This information was shared among information professionals at a conference to highlight the need for education; the specific recourse or litigation following that incident was not detailed, but it served as a catalyst for policy change.
• Q: How does this play out with pro se litigants and hallucinations?
  • A: Courts are seeing an influx of frivolous cases. While pro se litigants are sometimes ordered to pay attorney's fees for egregious behavior, the larger trend is firms suing the AI providers for UPL when the AI actively encourages additional litigation (as in the gaslighting case).