Unethical AI use in legal filings on the rise in Oregon and the US, along with penalties

Two Oregon lawyers were recently fined $110,000 for including fictitious cases in filings that were AI “hallucinations.” This is not the first such case in Oregon, and General Counsel of the Oregon State Bar Ankur Doshi says it’s unlikely to be the last.

There are hundreds of cases across the U.S. where the use of AI has resulted in incorrectly cited cases, fabricated quotes, cases that don’t exist, or all three. Doshi says using AI this way is antithetical to the sworn duties and responsibilities of the profession. Courts are increasingly imposing higher fines for these violations and state bar associations are also meting out discipline, which can range from admonitions to disbarment.

Doshi joins us to share more about the regulations and guidance in place for using AI in legal proceedings.

Note: The following transcript was transcribed digitally and validated for accuracy, readability and formatting by an OPB volunteer.

Dave Miller: From the Gert Boyle Studio at OPB, this is Think Out Loud. I’m Dave Miller. Not long ago, an Oregon lawyer used AI to help him write some legal filings. He ended up citing numerous completely made up cases. But the fine handed down by the Oregon Court of Appeals in March was very real, $10,000. Then earlier this month, news broke of a separate Oregon case with a much bigger penalty. A federal judge ordered two lawyers to pay $110,000 because of AI fabrications in their legal filings. It made national news as possibly the largest such fine anywhere in the country.

These weren’t the first cases of hallucinatory AI filings and they likely won’t be the last. Ankur Doshi is General Counsel for the Oregon State Bar. He joins us now. It’s great to have you on the show.

Ankur Doshi: Thank you for having me, Dave.

Miller: I mentioned these two recent stories that have made the news. How many other instances of AI hallucinations in Oregon filings are you aware of?

Doshi: So far, we’ve seen about five filings that have occurred in terms of cases. This is surprisingly a national trend that is taking a lot of us by complete surprise and it’s been percolating through the system. When the public gained access to generative AI tools back in 2022, we didn’t really expect this to become such an issue. But we received notice that there was the first type of this case occurring within the nation pretty early on, within about six to eight months after OpenAI – ChatGPT, at this particular instance – became available. Even after we saw that case came out, we didn’t really expect it to become an enormous issue either in terms of court sanctions or in terms of disciplinary proceedings.

However, what we’ve seen since then is this has become an issue that has become very percolating throughout the system itself. As of today, there’s been about 900 cases throughout the United States that have specifically referenced issues, where cases have been hallucinated. And of those, about 350 have been committed by lawyers themselves. So it’s become something that we’re seeing, not just within our own state but, across the nation as well.

Miller: When you say “lawyers themselves,” is it because sometimes it’s people representing themselves who are doing this?

Doshi: Yes. Because of the fact that generative AI tools are available to members of the public, we are seeing individuals who are representing themselves, who are going into court per se, using this to draft their own pleadings. That has actually become a very interesting issue in and of itself. We do see about 600 or so cases throughout the nation that deal with that type of issue. The part that we’re seeing a lot more of and that we’re concerned about, in terms of discipline aspects of it, is what is happening when lawyers are utilizing these tools and utilizing them improperly, which is causing a lot of these hallucinations and fabrications to occur.

Miller: One of the cases I mentioned was about the revocation of a marijuana production license. Another, the $110,000 one was an inheritance battle. Is there a pattern in these cases and are they also happening in criminal cases?

Doshi: Generally, we haven’t seen a pattern, aside from just attorneys wanting to be able to use this tool effectively and efficiently. Since generative AI tools came out earlier, in 2022, we’ve seen them being integrated in a lot of our legal tools. So a lot of larger firms are using them, smaller firms are using them. As the Bar, we’ve been working especially hard to make sure that attorneys are educated about how to utilize them. But what we’re seeing is that there’s not necessarily any type of pattern that’s specifically involved, about which attorneys are utilizing these and which ones are utilizing them improperly.

Instead, it’s a lot more just [a] misunderstanding of how generative AI works; whether or not the product that comes out of the generative AI is something that is reflective of actual research and thought versus something that is essentially a fabrication of cases that is utilized to respond to whatever prompt that the attorney put in. One of the first cases we had in discipline was an attorney who had been practicing since 1990 and has had a substantial amount of experience. So we’re not necessarily seeing any type of pattern of newer attorneys utilizing this or attorneys within a certain age set.

Miller: What do you think is behind this? I mean, in the cases where lawyers are just passing off made up cases in filings as their own work, is this just plain old laziness?

Doshi: I think there are a number of factors that are creating a confluence that is causing this to become such a substantial topic and such a substantial issue. One is that, as lawyers, we’re all under substantial pressure to get work done as fast as possible. We generally bill by the hour for most attorneys. And for attorneys who work on a contingency basis, they are billed by their efficiency. So there’s always an intentional pressure for us to work effectively. AI does represent a substantial amount of time savings and efficiencies for attorneys who integrated [it] into their work process properly. It does have a substantial amount of capability for being able to review our work and draft our work. But there is this need to understand how it works, and part of that is understanding the capabilities of it.

Within our own professional rules, we have what’s known as a duty of competence, in that we are supposed to be competently able to represent our clients. Part of that duty of competence is to be able to understand any technological items that we use in our representation. So one of the things that we are seeing a lot of is that attorneys don’t necessarily understand how AI functions and that when a prompt is put into a generative AI program, it doesn’t necessarily create something that is reflective of research that would be done by an attorney who would be working on the product itself. We’ve seen this because of just the number of different types of errors we see. We sometimes have attorneys who utilize a generative AI tool to check their work. When they use it to check their work, what happens is the product that they get back actually changes the citations in their actual work itself.

Miller: When you say “changes citations,” meaning a lawyer has done hopefully a good job or at least is not making cases up for their arguments. Then they put this in a generative AI model to check it and then the model actually fabricates cases in trying to check the human work?

Doshi: What we’re seeing a little bit of, in that aspect, is that the attorney requests that the AI checks the structure, the grammar. And what actually comes back is that it’ll change the citations themselves. Then as we go further and further, it gets more severe. When the attorney’s asking the AI to draft certain proportions, do the research, it’ll start generating false citations, false quotations from the case. If it gets to the point where attorneys are actually asking it to respond to an opposing counsel’s brief, it will respond to that prompt in a way to draft arguments against that brief, even though there may not necessarily be cases on point about that brief.

Miller: What you’ve just described is an incredibly unhelpful tool, just disastrously unhelpful. But my understanding is that it’s very common now for big law firms, and I guess even smaller or medium sized ones, to pay for access to a trainable AI model, as opposed to just off-the-rack, free ChatGPT, or even a paid version of one of these models that you can’t train. But these firms – and correct me if I’m wrong – can train these models, feed it relevant case law, relevant opinions or briefs, or writing style – this is how, at our firm, we do this stuff – so they can get more tailored output for their mergers and acquisitions or whatever they happen to do at their firm. Which seems very different than just going to ChatGPT and saying, “hey, write me a filing for this case with these parameters.”

How common is what I just described, in terms of what wealthier law firms, bigger law firms are doing?

Doshi: One thing I must emphasize on this is a lot of what I speak about, when we speak about fabrications, it sounds very much anti-AI, that attorneys should eschew any type of utilization of AI – and that is not necessarily the case. I think it does have a place. We do see it integrated a lot, especially into larger law firms that have the resources to create their own models.

One interesting aspect about this is that some of the technology is at a cost level that a lot of smaller firms can utilize to be on the same level as a lot of larger law firms because they’re able to process discovery with it faster. But the key element within utilizing it well, versus potentially creating an issue where you have fabrications, is that there has to be that human element who checks the work and is able to respond to it.

Miller: Just like you would for an early career associate or a clerk, right? I’d want a lawyer, a more established lawyer, to say, “I got this work product and now I’m going to check it to make sure it’s right.”

Doshi: Exactly. And one of the reasons that this has been such a substantial concern to the courts is because whenever we file something with the court, we sign our name to it. And the attorney who signs their name to a pleading, that’s submitted to the court, is ultimately responsible for everything in that pleading to be truthful and accurate. As one of our professional responsibilities, we have a duty of candor to the court. In keeping with that candor, when we sign our names to a pleading, we are responsible for that regardless of any associate who’s drafted it, any of our employees who drafted it, because of the fact that we have the supervisory duties over them.

Miller: That’s where the buck stops. So what are the repercussions, at the state bar level in Oregon, if someone is found to have knowingly used AI inappropriately?

Doshi: It depends a lot on the circumstances and the severity. Right now, we have about five investigations ongoing with this type of situation occurring. When this happens in a court, in addition to sanctions that are imposed, a common issue is that the court will specifically refer them to the Oregon State Bar for disciplinary proceedings. As of right now, we’ve only had one that has concluded with a sanction, where the attorney was admonished last year. However, we do still have five going on. The sanctions can range from admonition, all the way up to suspension and potential disbarment, depending on the case.

Miller: Is an admonition like, “hey, we saw you, don’t do this again, that was bad,” a kind of tisk-tisk, or is it more serious than that?

Doshi: For admonitions it’s generally kind of along the lines of a letter indicating that we have established that there is a rules violation, please keep this in mind for your future occurrences. But it’s kind of the first stage in a number of potential sanctions that can occur. A lot of that depends on the circumstances around it and whether or not there’s aggravating factors or mitigating factors. Especially, in terms of mitigating factors, where the attorney comes out and lets the court know that that particular instance might have been utilized by AI, they make specific actions to resolve that claim. Versus aggravating actions, where the attorney may try to hide the ball or paper over the fact that they utilize AI.

So there’s a lot of variance and sanctions, but part of that variance is based on the circumstances of the issue that occurred in the AI.

Miller: If opposing counsel, if judges, now need to be more aware than they have ever been that citations could just completely be made up, or maybe the name of a case is a real one, but the details or the quotes are completely made up, does all this have to be checked now, in a way that it wouldn’t have been 10 years ago? And is that going to sort of slow down and gum up the system?

Doshi: The effect of fabrications, specifically on the court system and the ability of the court system to operate, is a substantial concern to the court. That’s one of the primary factors that’s extremely aggravating about this. Because these citations are fabricated, it does create additional work for the court to have to go back and check it. But in addition to that, it creates additional work for opposing counsel on it, which can result in higher attorneys’ fees for the opposing side on it. One of the parts that we’re seeing a lot of, and one of the reasons that resulted in this very high sanction that was $110,000, was that the sanction was the cost to opposing counsels to review and indicate that these briefings had fabricated cases.

Miller: Which apparently they had done numerous times but which didn’t prevent more fabrications from being provided to the court.

Doshi: Of course. And the aspect of that is that, when you see a fabricated case, your first thing to think is not to say it’s a fabricated case. Your first thing to say is that this is a typo? Do I need to look further? But it’s not only just fabricated cases, it’s fabricated statements of law. So you have entire arguments that have no basis in law, which strikes directly at our precedent-based system.

Miller: Ankur Doshi, thanks very much.

Doshi: Thank you.

Miller: Ankur Doshi is General Counsel for the Oregon State Bar.

“Think Out Loud®” broadcasts live at noon every day and rebroadcasts at 8 p.m.

If you’d like to comment on any of the topics in this show or suggest a topic of your own, please get in touch with us on Facebook, send an email to thinkoutloud@opb.org, or you can leave a voicemail for us at 503-293-1983.