Not Ready For Prime Time: The Current State Of Legal Ethics And AI

from the there-be-dragons dept

I’ve been presenting at UC Law San Francisco Lexlab Law and AI certificate program its past several sessions, as well as some law school classes, on whether lawyers’ use of AI complies with the rules of professional conduct governing how lawyers must comport themselves or risk losing their licenses. The legal industry is keen to reap many of the benefits AI promises, such as streamlining some of the more arduous parts of the job and potentially making legal representation more affordable and/or profitable. And tech vendors are keen to profit from selling their AI systems to this market.

But neither constituency can benefit unless the way AI is used is consistent with those ethical rules. They exist for a reason—to make sure clients’ interests are prioritized and protected—and nothing about AI obviates their need or applicability. As lawyers start to roll these tools into their practice they need to make sure it’s not in a way that violates those rules, and for those vendors eager to sell their tools to the legal profession, they are going to need to make sure they are designed in a way where their use does not.

And we are not yet at a place where compliance can be presumed when AI tools are used. AI use by lawyers remains highly problematic for at least two big reasons: the current unreliability of AI outputs, and the tendency for AI to consume, store, and reuse data it is exposed to, even if that data needs to remain private. With regard to the former issue, the reliability concerns go beyond just the problem of hallucinated citations appearing in legal briefs; it is still the case that clients hire lawyers for their judgment, which so far AI still cannot replace. Maybe someday if AI has developed into something truly autonomous we could simply demand that it take the bar and be accountable to clients like human lawyers currently are, but since that day is not yet here it is critically important that lawyers not abdicate their own judgment in favor of whatever an AI tool might produce. Clients are depending on them, their lawyers, and they remain fully accountable to serve them as the profession requires.

And with regard to the data protection issue, it remains true for any tool lawyers use, AI-based or not: lawyers need to make sure that the tool use does not compromise client information, which they have a duty to protect because really bad things can happen to the client when privacy is not preserved. But AI tools in particular are notoriously greedy about collecting, retaining, and reusing whatever information they can access, unless they are specifically designed not to do so. It is thus critically important for lawyers to make sure that whatever tool they use—including and especially an AI tool—does not have the ability to mishandle or misappropriate client data to which it is exposed.

Each of these major issues then reverberates in a number of the specific ethical rules governing lawyers that they need to abide by. What follows is a closer examination of some of the ways they do.

There are a few things to note at the outset. First, there are more rules governing lawyers than what have been included in the analysis here, such as those relating to the duty to uphold the reputation of the profession, or those relating to advertising, which AI can also implicate. But the ones included are some of the major ones and examples of how AI use can rule afoul of them.

It’s also important to note that the rules can work together, and sometimes are in tension. The duty of zealous advocacy, for instance, can sometimes be at odds with the duty of candor. But AI does not itself resolve those ethical questions. Instead what is important to realize is that an AI use might implicate more than one rule.

As for the rules themselves, although they are rooted in some longstanding principles, what governs lawyers today are a set of model rules the American Bar Association promulgated in 1983 and pretty much every state has since adopted in some form. Because they can vary somewhat in how they were adopted by each state, lawyers need to consult the specific language their state has used to know how the rules apply to them, along with any other guidance and commentary their state’s lawyer regulators have produced. In fact, many states are also busy updating this guidance in order to specifically apply it to lawyer AI use (as is the ABA). But the basic gist of each rule is largely the same for all lawyers and based on the model language articulated by the ABA originally.

These rules also apply to all lawyers and not just litigators. Even transactional, or deal-making, lawyers need to follow them, and so do in-house counsel, who still have clients they owe duties towards, even if it is just one client. So when they declare, on behalf of their client—as, alarmingly, many have—that they will only hire law firms that use AI it is fair to question whether such a priority is indeed consistent with their ethical duties. Perhaps in individual circumstances such an AI-using firm might be preferable, but it won’t be universal; it will depends on the task the lawyers are being engaged to help with and what specific tools they are using. It would after all be contrary to the client’s interests to engage a firm that used an AI tool that compromised the confidentiality of the client’s sensitive information. Plus there remains the question of whether the AI tool being used is really one that can help the job get done in a way that can be trusted, let alone with the savings the client is hoping to see, given the time needed to verify its output.

Rule 1.1: Competence. If you are going to be engaged as a lawyer, you must make sure you know how to do the job, however you find yourself called to do it. Traditionally the duty of competence has, for example, meant that a lawyer could not take on a matter in an unfamiliar area of law, like criminal defense when they were an estate planner (or vice versa), but the rule is not limited to just those situations. Case hallucinations strike at the heart of it, for instance. If you are advocating for a client, you need to know the law relevant to the client’s situation, and if you are submitting hallucinated cases that don’t exist, it strongly suggests that you do not, or else you would have known they were imagined. Which is one reason AI cannot suddenly expand the types of matters a lawyer can take on—it’s simply not reliable enough on its own. The lawyer will still need to know when it is giving good answers and when it is not, but if the lawyer does not already have that competence themselves, then they won’t be able to make that determination.

There is also a related concept that is becoming more and more important to the legal profession: technical competence. Do you know how to effectively use the tools you are using? The answer needs to be yes, particularly to ensure that your tool use is not creating problems you were not aware of, especially with respect to protecting client data. As discussed further below, lawyers have a duty to protect client information, which means they need to know how to use their digital tools properly to ensure it remains protected. Concerns about technical competency predate AI, given the risk of potential hacker exfiltration, but they apply just as readily to an AI system exposed to client data that may then train on it.

Rule 1.3: Diligence. This duty goes hand in hand with the duty of competency but requires a bit more. It’s a lawyer’s job to do the job. And the whole job, not just the bits of the job they like best. One supposed promise of AI is that it can increase the volume of matters a lawyer can take on. But can it actually? Because on each matter it will still be the lawyer’s judgment on the line. So while AI might offer some time savings on certain tasks associated with a representation, and that might seem to create capacity to take on more matters, the lawyer will still need to have the bandwidth to competently provide any representation they’ve been engaged to provide. There are physical limits to what one lawyer can do, even with the help of AI, and especially given that they still need to be able to review whatever results an AI tool might give them to help along the way and that supervisory function will also require time.

This diligence rule also incorporates the general lawyer duty of zealous advocacy. Before relying too heavily on an AI tool, one needs to ask whether an AI can actually itself deliver zealous advocacy. Can it deliver all the strategic thinking needed to look out for the client and their objectives in evolving circumstances? And would whatever it recommended also comport with the rest of the ethical rules? If the AI can’t meet all those requirements—and it’s doubtful whether any truly can right now—the lawyer will still need to.

There is also a new wrinkle developing with respect to this rule and case hallucinations: increasingly courts are imposing a duty on litigators to police their opponents’ briefs for false citations. If they do not, they may not be able to recover fees for litigating against the fiction once they are finally brought to the court’s attention. Diligence appears to now require that sort of review (although in some instances it is framed as part of the duty of competence).

Rule 1.4: Communications. Clients are entitled to be kept apprised of how things are going with their representation and be able to contribute to it. On the one hand, if AI does create some time savings for lawyers, it may help ensure they have more time to communicate with clients. But if AI is used in a way that replaces lawyer judgment, and, worse, is applied to tasks where it performs them in a way that lacks adequate transparency, then clients won’t be given enough information such that they can assist with their own representation. The lack of transparency is especially an issue for agentic AI, where instead of manually working with an AI a step at the time the lawyer is trusting the AI to do a more complex task, because the AI may not even realize that it needs to report out what it has already done and be able to get more input before proceeding.

Furthermore, applying AI to client communications themselves also raises the risk of exposing confidential client information. Such a task would inevitably require the AI to learn about the client in order to know what to tell them, but if what it learns does not stay local to the law firm then its confidentiality is no longer protected.

Rule 1.5: Fees. This rule requires lawyer fees to be reasonable, which some have suggested may require using AI. The thinking is that if technology makes some lawyer tasks more efficient, it would be unreasonable to bill for the time taken to perform the tasks without the technology. And as a general proposition this view may be correct: we may, for instance, no longer consider it reasonable to bill for the time needed to research in a law library when so much information is now digitized. But it does not follow that using AI is similarly necessary to use, especially not when there are so many issues still associated with its use. Indeed, it’s even questionable whether it can provide true savings given that lawyers must still take the time to review anything resulting from an AI tool before it is relied upon. Perhaps in some in some situations there can be a savings, but, given the current state of the technology, the savings are not nearly certain enough at this point to support the inference that AI use is something that must be implicitly required.

There also is a side issue of when the costs associated with AI use can be passed through to a client. The answer may depend on guidance from the specific state regulating the lawyer, but there is some precedent, from the period when lawyers started regularly adopting Lexis and Westlaw research tools, and on fixed subscription rates, that technology services that are paid at a flat rate cannot be pro-rated among clients, and only costs directly associated with a client’s representation can be passed through to the client on their bill, at least not without the client’s prior written consent.

And it should not need to be said, but apparently, given some known cases it must be: if the AI use does in fact result in a time savings, then the client gets the benefit of those savings. Lawyers cannot bill for the time it would have taken to perform the task without the AI, only the time it did take, even if it went faster thanks to the AI.

Rule 1.6: Confidentiality of Information. This rule, in many ways, is the ballgame. If a technical tool, including AI, cannot protect the confidentiality of client communications, then it cannot be used by lawyers.

This rule is broader than attorney-client privilege, although AI that doesn’t protect client confidentiality may fail to protect attorney-client privilege as well. The principle behind this rule is that clients can’t get effective representation from their lawyers unless they can be candid with them. Which means that lawyers end up as vessels for all sorts of sensitive information that clients need to feel safe divulging. This rule helps make it safe for them to divulge it. It also applies to any information a client divulges, including their identity. If any such information is going to be shared with an AI, it can only be with an AI tool that can be trusted to maintain its secrecy, which means likely not uploading it anywhere and certainly not using it for any later purpose that won’t be local, such as training.

Which right away means that general purpose, freely available AI tools more than likely are unsuitable for legal work. There has already been at least one judicial ruling finding that the terms of service of one such tool dispelled any reasonable belief that confidentiality would be protected. It is incumbent on lawyers—including as part of their technical competency—to review the terms of service and privacy policy of whatever AI tool they want to use to understand if there’s any danger of information associated with the client’s representation leaving the control of the lawyer. It will also likely necessitate lawyers using only paid versions of AI products in order avail themselves of ones with terms of service that are adequately protective—and it will need to be only the right paid product, as there can be critical differences in privacy practices even among one vendor’s family of paid products. Which does unfortunately raise the issue that it may be only large law firms who will be able to afford using AI products that provide adequate assurance on an enterprise level, and smaller firms only able to afford “business” packages will be left out of being able to use AI.

There is also the unfortunate situation that clients may already be inadvertently waiving their own potential attorney client privilege by uploading information to AI systems—especially free ones—to ask them questions or even just have them collate their information into a more usable form. In addition to the issues associated with relying on the AI’s output, it means their private business may have just been absorbed into the larger system and now be discoverable by others. This ship may be sailing before clients engage their lawyers, but once engaged lawyers may have an obligation to educate their clients not to use AI this way because it risks losing attorney-client and work product protection, and as a practical matter may make information findable by an adversary. In fact, consider whether as part of a lawyer’s zealous advocacy they might have the obligation to ask AI about their opponent’s business, to see what information has already been revealed. On the other hand, lawyers should be aware that courts are considering whether their own AI prompts may be discoverable. Perhaps not, as work product or privileged, but it is an evolving area.

Rules 1.7-1:11: Conflict of Interest. Conflicts of interest are of serious concern to the lawyering profession. We need to make sure that lawyers are committed to zealously advocating for their clients and not pull their punches out of a sense of loyalty to someone else or some other interest. How AI may implicate these rules is something still developing, but one way to be aware of again relates to protecting client confidentiality, because it is conceivably possible that if an AI trains on client information, it may learn sensitive information about that client that later AI use may reveal, possibly in a context adverse to them.

Rule 3.3: Candor toward the Tribunal. This rule is about the integrity of the legal process. We’ve built a system where the idea is that with all the facts and law on the table, we’ll be able to reach a just conclusion. Reality may or may not be so simple, but this rule is about making sure the process has the best quality of facts and law available to it.

On the facts front, concerns are coming up more often, such as in the context of AI-generated police reports, and we’ve even had occasions where expert reports ended up with hallucinations, and another occasion where a lawyer was just trying to get AI to format a citation, and instead the AI went ahead and hallucinated a whole cited source. There is also a report from Brazil where a lawyer wrote a brief with hidden white-on-white text intended to do an injection prompt on any AI system the brief was run through. The court deemed it an act “offensive to the dignity of justice” and the lawyers were sanctioned.

And then there is the question of hallucinated law, which seems to be an epidemic. Note that the issue here is not just invented citations for fake cases but bad summaries of holdings from actual cases, or otherwise invented quotes. Even taking AI off the table this rule about candor already obligates a lawyer to disclose to the court adverse authority. This obligation persists even in the face of the duty for zealous advocacy, although it can provide an opportunity for effective advocacy to be so candid because if that adverse authority is out there, the opponent might dig it up, so the duty to be candid gives the lawyer the ethical space to get out in front of it and try to minimize its effect. But the point of this rule is to make sure the court has the benefit of all the law it needs to consider in reaching a just result, which is why it would violate it to give it hallucinated law that at best obscures applicable law, if not outright deceives the court.

The strange thing though is that it apparently needs to be said that not only should lawyers not deceive courts by submitting briefs with hallucinated citations, but, once called out for it, the solution is not to double-down and submit even more hallucinations in briefs arguing why the lawyer should not be sanctioned for the first hallucinations. Oddly, such behavior seems to be happening a lot, and these cases are where the resulting sanctions have so far been most severe.

Rule 3.4: Fairness to Opposing Party and Counsel. This rule joins the duty of candor to also help make sure our adversarial system can function. But it applies to more than just legal filings; it is intended to discourage obstructive strategy altogether, including by not depriving opponents of evidence they are entitled to.

Which matters, because one use of AI is in support of document review and production. Quicker and more cost-effective document review is a white whale many hope AI can finally slay, but there is a danger in presuming that it has already replaced the need for human review. It may be able to make it more efficient by clumping up documents based on the likelihood of them being relevant or privileged, but there are consequences to not getting attorney judgment to make the final call, including potentially waiving privilege, overproducing, or under-producing and ending up subject to sanctions. And the risk remains: as the AI is learning, what is happening to the client information it is learning?

Rule 5.1: Responsibilities of Partners, Managers, and Supervisory Lawyers. It is not enough for a lawyer to just be ethical on their own; they are also responsible for the ethical conduct of those who work with and for them. Courts are increasingly sanctioning not just lawyers who have submitted briefs with hallucinated citations but also their bosses and law firms, as well as co-counsel. Law firms need to have policies about what AI use is acceptable and make sure everyone is trained on it, which also means that if someone wants to use a new tool, it has to be vetted first to make sure it can be trusted by any lawyer at the firm. Furthermore, lawyers need to be judicious about what other lawyers they co-counsel with or sponsor for pro hac admission to their local court, because any name on a brief will be responsible for its content.

This supervision obligation is also why lawyers cannot simply rely on output an AI tool generates; they must verify it. And they must use non-AI systems to do it, so that if bad information has corrupted the original AI tool, the system that’s used to verify the results won’t be vulnerable to the same corruption.

Rule 5.3: Responsibilities Regarding Nonlawyer Assistance. Supervisory duties do not just involve other lawyers; lawyer are also responsible for non-lawyers who work with and for them. For instance, if a firm secretary steals client funds, the lawyers are on the hook, because it is there duty to make sure such things don’t happen. It also means they need to manage vendors that they would outsource any work to, such as printers and document review vendors, as well as any technology vendors, including AI vendors, all of whom need to be vetted to see if they can be trusted with the sensitive material they will inevitably work with.

Rule 5.4: Professional Independence of a Lawyer. There’s also a rule that law firms can’t be owned by non-lawyers. The reason for this rule is the same as the reason for the conflicts of interest rules: it is critically important that lawyers have no incentive to do anything but zealously advocate for their clients. The fear is that, if firms could be owned by non-lawyers, then the profit pressures felt by owners without these other ethical concerns would overtake that client-first orientation. But especially in periods, like now, where there’s a gold rush to invest in certain technology and offer equity and other financial incentives to underwrite it, there ends up being a lot of pressure put on this rule, with lawyers being tempted to sell equity stakes in their firms, and non-lawyers being keen to buy them. But the rule exists for a reason. Legal representation is so important to liberty and due process that a right to it is enshrined in the Constitution. Whereas no one is owed the profession as a profit center, nor could society afford for it to be recast as a normal profit-seeking business, which would come at the expense of the critical constitutional purpose we need it to serve.

Rule 5.5: Unauthorized Practice of Law. The unauthorized practice of law is ultimately defined mostly by local law, but the model rules themselves say that only those with licenses and subject to bar regulation are entitled to practice law. There have already been instances with some AI vendors trying to be hired to advise and litigate cases, in pretty clear contravention of this rule and associated local law. Plus there are states trying to pass bills preventing chatbots from dispensing legal advice.

The tricky thing is that it’s not always clear what is meant by dispensing legal advice. And there is some tension between these rules and access to justice concerns—in some situations it may perhaps be better to have a non-lawyer with some relevant expertise help people, rather than them potentially going without any support at all. But non-lawyers could get clients into a lot of trouble with ill-informed advice, and the same danger exists with AI. Some of these new tools, especially when well-designed for the legal realm, may be useful. But some, especially those of general purpose, are frequently terrible, and yet people still try to use them for legal advice and depend on unreliable if not completely wrong advice—in addition to risking the disclosure of private information in ways that may also tempt trouble.

In any case, lawyers certainly shouldn’t be letting AI tools dispense legal advice through their practice. When they get output from an AI they need to make sure it is output that is appropriate to rely on. Which is why the idea of AI providing significant savings seems such a dubious proposition, because either the lawyer is spending the time researching themselves, for instance, or, consistent with their supervisory duties, they are spending the time reviewing the research the AI produced, and in a system that itself is not AI and thus subject to the same information corruption issues.

AI is not yet in any sort of fit state where anyone can afford to rely on what it produces as the final, correct word on anything. Maybe someday it will be, but until we are sure that day has arrived, the ethical rules governing lawyers prevent clients from being unwilling beta testers. Even when lawyers get knowing, written consent from clients to use AI—which would be a best practice anyway—there are still pitfalls everywhere. And even if AI use is something clients may think they want, it may still be something lawyers are obligated to advise against.