In 2023, two attorneys filed a brief in federal court citing six cases that didn’t exist. The cases had realistic-sounding names, realistic-seeming citations, and zero basis in reality. ChatGPT had made them up. The judge found out. Both lawyers were sanctioned.
That case, Mata v. Avianca, became the AI horror story that every bar association newsletter has referenced at least three times. And it launched a wave of judicial orders, ABA guidance, and state bar ethics opinions that solo lawyers now have to navigate.
Here’s the thing: the problem in Mata wasn’t that lawyers used AI. The problem was that they used a general-purpose chatbot for legal research and then submitted its output without checking it. That’s not an AI problem. That’s a supervision problem. And it’s the exact problem that purpose-built AI agents are designed to eliminate.
This guide covers the real sanctions cases, what courts are actually saying, the ABA rules in play, and how to use AI agents in your solo practice without putting your bar license at risk.
The Cases That Changed Everything
Mata v. Avianca (S.D.N.Y. 2023)
This is the one everyone knows. Attorneys Peter LoDuca and Steven Schwartz used ChatGPT to research a personal injury case. The AI generated citations to cases like Varghese v. China Southern Airlines and Zicherman v. Korean Air Lines. Neither case existed. When opposing counsel couldn’t locate the cases, Judge P. Kevin Castel ordered the attorneys to produce the cases. They couldn’t. After a hearing, Castel imposed sanctions of $5,000 on each attorney and required them to notify the judges named in the fake citations.
The court’s order specifically noted that the attorneys “abandoned their professional obligations” by relying on AI output without independent verification. The lesson here isn’t “don’t use AI.” It’s “don’t use a tool that hallucinates citations for work that requires verified citations.”
Park v. Kim (2d Cir. 2023)
A few months after Mata, the Second Circuit sanctioned an attorney who submitted a brief with citations to nonexistent cases. Again, ChatGPT. Again, no verification. The court ordered the attorney to show cause and ultimately referred the matter for further disciplinary proceedings. The attorney faced potential suspension.
Gauthier v. Goodyear Tire (E.D. Tex. 2023)
Texas federal court, same pattern. AI-generated brief with fabricated citations. The court issued an order to show cause and required the attorney to complete continuing legal education on AI use. The court also warned that future violations could result in dismissal of the case or referral to the state bar.
Kohler v. Inter-Continental Hotels (C.D. Cal. 2024)
By 2024, courts were no longer treating these as isolated incidents. In this ADA accessibility case, an attorney submitted a motion citing cases that either didn’t exist or didn’t say what the brief claimed they said. The court sanctioned the attorney and required written confirmation in future filings that all cited cases had been verified through Westlaw or Lexis. This is now becoming standard in many jurisdictions.
The Pattern Across All These Cases
Every single sanctions case involving AI has the same structure: a lawyer used a general-purpose AI chatbot, didn’t verify the output, and submitted fabricated citations to a court. Not one of these cases involved an AI system that was purpose-built for legal research with verified databases. That distinction matters enormously.
| Case | Court | Tool Used | Sanction |
|---|---|---|---|
| Mata v. Avianca | S.D.N.Y. | ChatGPT | $5,000 per attorney + judicial notifications |
| Park v. Kim | 2d Cir. | ChatGPT | Show cause + disciplinary referral |
| Gauthier v. Goodyear | E.D. Tex. | ChatGPT | Mandatory CLE + future filing restrictions |
| Kohler v. Inter-Continental | C.D. Cal. | AI chatbot | Sanctions + mandatory verification requirement |
What Courts Are Actually Ordering
After the wave of 2023 sanctions, courts started issuing standing orders about AI use. By early 2025, more than 40 federal judges had issued AI-specific orders. They break into a few categories.
Disclosure Orders
Some judges require attorneys to disclose whether AI was used in drafting any filed document. Judge Brantley Starr in the Northern District of Texas requires attorneys to certify that they have not relied on AI-generated research without independent verification. Several other judges require disclosure of which AI tools were used.
Verification Requirements
Other courts require that every cited case be verified through a licensed legal database before filing. This isn’t new as a standard, but courts are now explicitly requiring it as a condition of using AI at all. You can use AI to draft. You cannot use AI citations without checking them in Westlaw or Lexis first.
Training Requirements
A handful of courts have required sanctioned attorneys to complete AI-specific CLE before they can file again. Several state bars, including California and Florida, have issued guidance recommending that attorneys understand how any AI tool they use actually works before relying on it in practice.
The ABA Rules You Need to Know
There’s no ABA rule that says “don’t use AI.” There are several rules that govern how you have to use it.
Rule 1.1: Competence
The ABA amended Comment 8 to Rule 1.1 in 2012 to include technology competence. The comment says lawyers should “keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology.” In practice, this means you’re expected to understand the basic capabilities and limitations of AI tools you use in your practice. You don’t need to know how transformer models work. You do need to know that general-purpose chatbots can hallucinate citations and that this is a known risk.
Rule 1.3: Diligence
Diligence means you complete work thoroughly and on time. Submitting AI-generated research without verification isn’t diligent. Courts have cited this rule in sanctions orders because submitting unverified citations is the opposite of pursuing a client’s matter “with reasonable diligence and promptness.”
Rule 1.6: Confidentiality
This is the one solo lawyers sometimes miss. When you paste client information into a general AI tool, you may be disclosing confidential client data to a third party. OpenAI, Google, and Anthropic all have data retention and training policies. Some offer enterprise agreements that provide stronger confidentiality protections. Without one, you’re potentially violating Rule 1.6 every time you paste a client’s facts into ChatGPT.
AI agents deployed specifically for your firm, connected to your own matter management system, don’t have this problem. The data stays in your environment.
Rule 5.3: Supervision of Nonlawyer Assistance
The ABA’s Formal Opinion 512 (2024) directly addressed AI use. The opinion concluded that AI systems used by lawyers constitute “nonlawyer assistance” under Rule 5.3, which means you have a duty to supervise the AI’s work output. You can’t just accept what an AI generates. You have to review it with the same care you’d apply to work product from a paralegal or contract attorney.
ABA Formal Opinion 512 (2024): The Full Picture
This opinion is worth reading in full if you use AI in your practice. The key points:
- Lawyers must supervise AI output under Rule 5.3
- Using AI doesn’t reduce your professional responsibility for the work product
- You must understand enough about the AI tool to supervise it meaningfully
- Confidentiality obligations apply to client data entered into AI systems
- Fee arrangements must still meet reasonableness standards even when AI reduces the time spent
Why General AI Tools Create These Problems
General-purpose AI chatbots are trained on massive datasets that include legal documents, but they’re not connected to live legal databases. When you ask ChatGPT about a legal question, it generates a response based on patterns in its training data. It doesn’t look up actual cases. It predicts what a plausible-sounding answer would look like, and sometimes that prediction includes case names and citations that seem real but aren’t.
This is called hallucination, and it’s not a bug that’s going to get fixed. It’s a fundamental characteristic of how large language models work. The model doesn’t know what it doesn’t know. It generates confident-sounding output regardless of whether the underlying facts are accurate.
For legal research, this is catastrophic. You can’t submit a brief with citations you haven’t verified. And verifying every citation generated by a hallucinating AI defeats the purpose of using AI for research in the first place.
How AI Agents With Verified Databases Are Different
AI agents built for legal work don’t rely on the model’s training data for citations. They’re connected to live, verified legal databases. When a research agent looks up a case, it retrieves the actual case from a verified source, not a prediction of what the case might say.
The architecture is different in a few important ways.
Retrieval-Augmented Generation (RAG)
Purpose-built legal AI uses a technique called RAG, where the agent first retrieves actual documents from a verified database and then generates its response based on those documents. The citations come from the database, not from the model’s memory. If the case doesn’t exist in the database, the agent doesn’t cite it.
Verified Source Integration
Agents connected to Westlaw, Lexis, or Fastcase are pulling from the same verified sources you’d use manually. The agent doesn’t hallucinate a citation because it’s not guessing. It’s retrieving. The difference between a retrieval-based system and a generation-based system is the difference between looking something up and making something up.
Audit Trails
Good legal AI agents log every source they used to generate a response. You can see exactly which cases were retrieved, from which database, and when. That audit trail is how you supervise the AI’s work under Rule 5.3. You review the sources, not just the output.
| Feature | General AI Chatbot | Purpose-Built Legal Agent |
|---|---|---|
| Citation source | Model training data (hallucinated) | Live verified legal databases |
| Verification | None built in | Every citation traced to source |
| Client data privacy | May be used for training | Stays within firm environment |
| Audit trail | None | Full source log for supervision |
| Rule 5.3 supervision | Requires manual review of everything | Source log enables efficient supervision |
| Rule 1.6 compliance | Requires enterprise agreement | Built for confidentiality compliance |
State Bar Ethics Opinions: What Your State Is Saying
Beyond the ABA, most state bars have issued guidance or are in the process of doing so. A few key examples:
California
The California State Bar’s 2023 guidance on generative AI emphasizes competence, supervision, and confidentiality. California has particularly strict confidentiality rules, and the guidance makes clear that pasting client information into a public AI tool is risky. The guidance also notes that using AI doesn’t shift professional responsibility, the lawyer remains responsible for all work product regardless of how it was generated.
Florida
Florida Bar Ethics Opinion 24-1 concluded that lawyers may use generative AI but must ensure they understand the tool, supervise its output, and protect client confidentiality. The opinion specifically addressed the risk of using AI for research and noted that citation verification is non-negotiable.
New York
The New York State Bar Association issued a comprehensive report in 2024 covering AI use in legal practice. The report recommends that lawyers “understand the limitations of generative AI, including its tendency to hallucinate,” and requires that all AI-generated legal research be independently verified before submission.
Texas
Texas has been more aggressive than most states, with multiple courts (not just bars) issuing AI orders. The Texas Disciplinary Rules require competent representation, and the State Bar has issued guidance that using AI without understanding its risks violates this standard.
Disclosure Requirements: What You Have to Tell Clients and Courts
Court Disclosure
As of 2026, about 45 federal courts have standing orders requiring some form of AI disclosure. These vary in scope. Some require disclosure if any AI was used in drafting. Others only require disclosure if AI was used for research. A few require disclosure of the specific AI tools used.
Before you file in any federal court, check that court’s standing orders. The Northern District of Texas, the Eastern District of Texas, and several others have extensive AI requirements. Failure to comply with a standing order is itself sanctionable, even if your underlying work is solid.
Client Disclosure
The ABA has not yet required blanket client disclosure of AI use, but several state bars have moved in this direction. California’s guidance suggests disclosure is a best practice. A handful of states are moving toward requiring disclosure as part of the fee agreement.
Practically speaking, adding a simple AI use disclosure to your engagement letter is smart risk management even where it’s not required. Something straightforward: “Our firm uses AI tools to improve efficiency. All AI-generated work product is reviewed and verified by a licensed attorney before use in your matter. Client information is not shared with public AI services.”
Fee Disclosure
ABA Opinion 512 raised a question that solo lawyers need to think about: if AI reduces the time you spend on a task, do you have to pass those savings to the client? The opinion doesn’t require it, but it notes that fees must remain reasonable. If you’re billing by the hour and AI reduces a 4-hour research task to 30 minutes, billing 4 hours is ethically questionable.
The practical answer for most solos is to move toward flat fees where AI efficiency benefits the firm’s margin rather than creating ethical billing questions. Agents accelerate your throughput. Flat fees let you capture that value cleanly.
How to Use AI Agents Safely in Your Solo Practice
None of this means you should avoid AI agents. It means you should deploy the right kind of agents in the right way. Here’s the framework that keeps you compliant and productive.
1. Use Agents Connected to Verified Sources
For any work that involves legal citations, use an agent that pulls from Westlaw, Lexis, Fastcase, or a comparable verified database. Don’t use general chatbots for legal research. The risk isn’t worth it. Purpose-built legal research agents retrieve from verified sources. General chatbots predict. Those are fundamentally different activities.
2. Keep Client Data in Your Own Environment
Use agents deployed within your practice management system or through a provider that guarantees data confidentiality. Don’t paste client names, case facts, or identifying information into public AI tools. Your Rule 1.6 obligations apply every time you interact with an AI system.
3. Review the Sources, Not Just the Output
When an agent generates a research memo or draft brief, your review should include checking the cited sources directly. If the agent provides an audit trail of sources used, review that list. Click through to the cases. Confirm they say what the agent says they say. This is what Rule 5.3 supervision actually looks like in practice.
4. Check Court-Specific AI Orders Before Filing
Before filing in any federal court, search that court’s standing orders for AI requirements. The Federal Judicial Center maintains a list of courts with AI-specific requirements. Add this to your pre-filing checklist. It takes 2 minutes and can prevent a sanctions motion.
5. Document Your Review Process
Create a simple internal protocol: when you use AI to assist with a filing, note in your file what AI was used, what output was generated, and what verification steps you took. If you ever face a sanctions inquiry, this documentation shows that you exercised reasonable supervision. It’s the difference between a lawyer who got sloppy and a lawyer who had a process.
6. Update Your Engagement Letter
Add a simple AI disclosure paragraph to your engagement letter. Cover: (1) you use AI tools to improve efficiency, (2) all work product is attorney-reviewed, and (3) client data is not shared with public AI services. This isn’t a legal requirement in most states yet, but it’s good practice, and it prevents client complaints if they find out you used AI after the fact.
The Agents That Carry the Lowest Risk
Not all AI agents carry the same risk profile. Here’s how to think about it.
Low risk: Intake agents that handle initial client contact, collect intake information, schedule consultations, and send confirmation emails. These agents aren’t doing legal work. They’re doing administrative work that any paralegal could handle. No citation risk, no confidentiality risk if properly configured, no competence issues.
Medium risk: Document drafting agents that generate contracts, motions, or letters based on templates. The risk here is competence: you need to review what the agent drafts before sending or filing. The agent can get the structure right and miss a jurisdiction-specific requirement. Your review catches this. The risk is manageable with a solid review process.
Higher risk (but manageable): Research agents. These carry the highest risk because errors in research show up in court filings. But a research agent connected to verified databases and providing source citations is genuinely safer than an attorney rushing through manual research under time pressure. The key is using the right tool and having a verification step.
| Agent Type | Risk Level | Primary Rule in Play | Key Mitigation |
|---|---|---|---|
| Intake agent | Low | Rule 1.6 (confidentiality) | Keep data in firm environment |
| Scheduling agent | Low | Rule 1.6 | No client matter data needed |
| Billing agent | Low-medium | Rule 1.5 (fees) | Attorney review of invoices |
| Document drafting agent | Medium | Rules 1.1, 5.3 | Attorney review before sending/filing |
| Research agent | Medium-high | Rules 1.1, 5.3, court orders | Verified database + source review + court order check |
What Hello Paralegal Does Differently
Hello Paralegal builds AI agents specifically for solo law firms. The agents are designed around the compliance requirements in this guide. Intake agents keep client data within your firm environment. Research agents connect to verified legal databases, not general web crawls. Every agent provides an audit trail of its work so you can supervise meaningfully under Rule 5.3.
The sanctioned lawyers in Mata and the other cases didn’t have a purpose-built legal agent. They had a general chatbot that was never designed for legal work. That’s the gap. You need agents that were built for the environment you operate in, with the compliance requirements you face.
Solo lawyers have less margin for error than large firms. One sanctions order can damage your reputation in ways that a BigLaw associate can survive and you can’t. The right agent stack doesn’t just make you more efficient. It makes you more careful than you could be manually, because it surfaces sources, maintains audit trails, and flags when it’s operating outside its verified knowledge base.
The Takeaway
The lawyers who got sanctioned weren’t bad lawyers. They were good lawyers who grabbed the wrong tool. They treated a general-purpose chatbot like a research database, and it behaved like a general-purpose chatbot. The court didn’t care that they didn’t know better. Rule 1.1 requires you to know better.
The answer isn’t to avoid AI. The answer is to use the right AI. Agents built for legal work, connected to verified databases, operating within your confidentiality requirements, and providing the audit trails you need to supervise their work. That’s how you get the efficiency without the sanctions exposure.
Check the court-specific orders before filing. Update your engagement letter. Use agents connected to verified sources. Review the sources, not just the output. Document your process.
That’s it. That’s the whole framework. The lawyers who get sanctioned skip one of those steps. Don’t skip any of them.
Frequently Asked Questions
Do I have to disclose to courts that I used AI?
It depends on the court. As of 2026, roughly 45 federal courts have standing orders requiring some form of AI disclosure. Check the standing orders for every court where you file. State courts vary widely, with some requiring disclosure and others not yet addressing it.
Is it an ethics violation to use ChatGPT for legal research?
Not automatically, but submitting AI-generated citations without independent verification has resulted in sanctions in multiple courts and could constitute a violation of Rules 1.1 and 5.3. Using ChatGPT and then verifying every citation in Westlaw or Lexis is defensible. Using ChatGPT citations without verification is not.
Do I have to tell clients I use AI?
Most state bars recommend disclosure as a best practice, though it’s not uniformly required yet. Adding an AI disclosure paragraph to your engagement letter is smart risk management regardless of whether your state requires it.
Can I bill clients for time that AI handles?
ABA Opinion 512 requires that fees remain reasonable. Billing full hourly rates for tasks that AI completes in a fraction of the time is ethically problematic. Most solo lawyers using AI agents are shifting toward flat fees to avoid this issue entirely.
What’s the safest way to use AI agents in my practice right now?
Start with intake and administrative agents. These carry the lowest risk and provide the highest immediate ROI. When you add research agents, use only tools connected to verified legal databases and review the source citations before filing anything. Keep client data out of public AI tools. Document your review process for every AI-assisted filing.
How is an AI agent different from ChatGPT?
A purpose-built AI agent for legal work retrieves information from verified sources rather than generating it from training data. It’s connected to live databases, operates within your firm’s data environment, and provides audit trails of the sources it used. ChatGPT and similar general tools generate plausible-sounding text based on training patterns, without access to live verified legal databases.
Leave a Reply