Key insights:

• • • The duty of competence now includes technology 鈥� Rules first written for legal knowledge and diligence are being read to cover the tools lawyers use, not just the arguments they make.

    • The rules didn’t anticipate AI, but they still apply 鈥� Formal opinions and updated guidance make clear that generative and agentic AI fall squarely within lawyers鈥� existing ethical obligations.

    • What reaches the court carries the highest stakes 鈥� A lawyer’s technology missteps are most consequential, and most visible, the moment they show up in a filing, an exhibit, or an argument made to a judge.

Every attorney holds a duty of competence, and that duty has always included legal knowledge and sound judgment. Today, however, it now extends to technology as well. Often called 鈥�technical competence鈥�, today鈥檚 duty of competence requires understanding the benefits and risks of the tools used in practice, not just the law itself.

The American Bar Association’s Model Rules of Professional Conduct describe an attorney’s duty of competence as the obligation to provide representation using the (Rule 1.1). Comment 8 of the rule states explicitly, in part: 鈥淭o maintain the requisite knowledge and skill, a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology鈥� .鈥�

Indeed, competence matters well beyond a lawyer’s own file. Every filing, brief, and exhibit is, in effect, a lawyer’s communication to the court, and the court relies on the presumption that what it receives has been prepared with care. When a lawyer misunderstands or misuses technology, whether that means citing an AI hallucination as precedent or mishandling e-discovery output, the failure lands directly on the judge’s desk.

Rules of professional conduct are written broad and durable, so they do not need constant revision. Even so, few could have anticipated the pace of technological change now reaching the legal field.

A lawyer cannot simply plead ignorance when technology goes wrong, and failing to understand the tools being used can lead to ethics complaints, sanctions, or malpractice claims. Perhaps even more damaging, it can just as easily undermine a judge’s confidence in everything else that lawyer submits.

Rules built to last meet a technology no one predicted

Each state and territory maintains their , but most draw heavily on the ABA’s Model Rules as a foundation. Precedent is also shaped by disciplinary proceedings and case law, including a growing number of cases in which judges have sanctioned lawyers for submitting AI-generated content that was never checked for accuracy.

Rules of professional conduct are written broad and durable, so they do not need constant revision. Even so, few could have anticipated the pace of technological change now reaching the legal field. E-discovery was an early example, already raising real questions about what a lawyer could certify to a court in good faith. The more significant shift today is generative AI (GenAI) and agentic AI, capable of producing legal analysis or even taking action with minimal human oversight, often producing material that looks polished and authoritative right up until a judge or opposing counsel checks it.

Recognizing this shift, the ABA issued in July 2024 on the use of GenAI tools in legal practice, confirming that the duty of competence applies squarely to AI. The Formal Opinion requires lawyers to understand these tools well enough to use them responsibly, supervise their output, and avoid overreliance on unverified results 鈥� a standard that matters most the moment a document is filed, or an argument is made in open court.

5 ways to stay ahead of the curve

Given how quickly legal technology is evolving, and how little margin for error exists once something reaches a judge, there are five practical ways that lawyers can maintain their technical competence, including:

1. 1. 1. Pursue CLEs and structured education 鈥� Continuing legal education courses focused on AI, e-discovery, and cybersecurity offer a reliable, credentialed way to stay current. Many state bars now offer, and some require, CLE credits specifically on legal technology, exposing lawyers to real-world case studies that include those involving AI missteps that have drawn judicial sanctions.
      2. Build personal familiarity through hands-on use 鈥� Reading about a tool is no substitute for using it. Treating AI as a “thought partner,” for drafting, brainstorming, or spotting issues, helps professionals develop an intuitive feel for what these tools do well and in which ways they fall short, long before a document ever reaches a court. This low-stakes experimentation on matters that don鈥檛 impact client confidentiality or sensitive data can help build practical AI fluency.
      3. Practice in sandboxes and controlled environments 鈥� Before relying on a new platform in an active matter, test it in a sandbox first. Many vendors and firms now offer walled-off spaces to explore a tool’s features and failure points without exposing real client data.
      4. Review organizational guidelines 鈥� Firms, courts, and bar associations increasingly publish their own AI use policies that cover permitted tools, disclosure requirements, and data security. A growing number of courts also require lawyers to certify or disclose any AI use in filings, making familiarity with local rules as important as firm policy.
      5. Choose the right tools for the task 鈥� Free consumer-grade AI tools may suit general research, but should not be used in matters involving privileged or sensitive information, let alone a court filing. Premium and fiduciary-grade platforms, built with legal-specific safeguards around data handling and auditability, are better suited for substantive casework. Matching the tool to the sensitivity of the task, rather than convenience, is critical.

Used together, these five habits can help lawyers build a durable foundation in AI literacy. While technology will keep evolving, of course, any lawyer who has internalized this approach will be far less likely to be caught off guard once their work is tested in front of a judge.

Why this is about more than just the rules

Maintaining technical competence is not simply about avoiding disciplinary consequence. It is about preserving what the legal system is meant to provide. When lawyers understand the technology they use, including AI, they can apply it to research and review in ways that benefit clients and courts alike.

Just as importantly, technical competence protects the public’s trust in the legal profession, and something more immediate: a judge’s ability to rely on what lawyers put in front of them. In the end, technical competence is not a burden imposed by the rules; rather, it鈥檚 what allows the relationship between lawyers and the court, and the profession itself, to keep working.

For more on AI in the courts, check out the 听鈥� a joint effort by the National Center for State Courts听(NCSC) and the 成人VR视频 Institute (TRI)

Key highlights:

• • The AI strategy-execution gap is an organizational problem, not a tech one 鈥� Among professionals whose firm or department has a stated AI strategy, more than half say that either the strategy is not visible on a daily basis or that the organization has no strategic AI direction at all.

  • The human cost of inaction is building faster than most leaders recognize 鈥� More than 90% of professionals say they are experiencing some degree of this AI-value disconnect, and among them, one quarter is considering leaving their current organization within two years if things don鈥檛 change 鈥� at an estimated replacement cost of $232,000 per employee.

  • Disrupted development compounds talent risk for the next generation 鈥� The flight risk of experienced professionals creates a compounding multiplier effect on the ability of entry-level talent to develop critical skills.

The greatest barrier to AI transformation in professional services is the widening human gap between what organizations promise their people about AI and the spoken and unspoken messages that professionals see and observe in their workplace every day, according to the 成人VR视频 recent , which surveyed more than 1,800 professionals in 62 countries across areas of law, tax, audit, accounting, compliance, risk, and global trade.

Establishing a visible AI strategy

While our research showed that most organizations have a stated AI strategy, the data suggests that it is not translating evenly to employees in their day-to-day work. Among professionals whose firm or department has a formal AI strategy, 35% say that strategy is not visible in their day-to-day experience, and another 17% say their organization has no strategic direction on AI at all. That means that more than half of professionals with fiduciary commitments are working in an environment in which the AI strategy on paper does not match the reality of how their work gets done.

The reasons that professionals say AI strategies are stalling suggest an absence in AI transformation and change agility across the organization. Indeed, professionals say that drivers of their organizations’ struggles to translate AI ambition into measurable results include the fact that the right tools are not yet in place (with 47% of respondents saying this), people are not equipped or trained to work in the intended way (43%), the strategy has not been translated into clear operational priorities (32%), and there is no shared understanding of the AI strategy across the organization (30%).

When strategic clarity around AI exists, it translates into more visible value. In fact, more than two-thirds of professionals in firms and departments with a stated strategy say AI is meeting or exceeding expectations for creating value at work. When there is no understandable AI strategy, less than one-quarter say this.

The quiet accumulation of human costs

The talent consequences of the gap between stated AI strategy and its daily execution are building faster than most leaders recognize. More than 90% of professionals say they are experiencing this gap to some degree. Among them, 1-in-4 is considering leaving their current organization within the next two years if things don鈥檛 change. This can result in an estimated replacement cost of $232,000 per employee, which means the quantifiable financial outlay of this talent flight can add up quickly.

The greatest vulnerability of flight risk sits with mid-career professionals, who often are the most embedded AI users, the most influential in day-to-day operations, and the most impatient with slow adoption. Almost 30% of these professionals would change jobs within two years if AI fails to deliver the value they expect, and 14% say they are considering leaving within the next 12 months.

Of course, the financial risks extend beyond mid-career professionals and could in fact disrupt a generation of early career talent. When experienced professionals leave, they take with them the mentorship and oversight upon which the development of early-career employees depends.

The report shows that 71% of professionals say they believe early-career roles need structured support from experienced peers to develop the skills that are at risk of being displaced by AI. Moreover, nearly half say they are concerned about AI鈥檚 impact on the development of independent judgment and learning through experience. In fact, legal professionals specifically say they expect the timeline for early-career lawyers to develop a level of trusted judgment could be stretched by nearly two years.

Taken together, these factors create a compounding multiplier effect that will almost certainly have a negative impact on organizational performance in the near future, the report suggests.

Recommended actions for employers and professionals

The Future of Professionals Report 2026 details three distinct paths for how organizations can deploy AI:

1. 1. 1. Elevate, which allows AI to handle rote tasks while keeping human expertise at the center;
      2. Scale, which uses AI primarily to increase capacity without increasing headcount; and
      3. Reimagine, which puts AI at the core as it rebuilds operating models and service propositions from the ground up.

The data also makes clear that the organizational and human costs of inaction are multifaceted. There are several concrete steps that both employers and professionals can take now, as outlined in the report, which include:

For employers:

• • • Choose a path and make it visible 鈥� The aforementioned three paths represent genuinely different futures with different commercial models, talent strategies, and definitions of professional value. Organizations must choose one deliberately and make it easily visible at the individual and leadership levels.
    • Close the rift in alignment before it results in talent departure 鈥�听More than one-third of professionals say they are working somewhere where the AI approach does not match their preference. These professionals are almost twice as likely to consider leaving within the next 12 months 鈥� and organizational leaders need to be aware of that.
    • Let strategy drive investments听鈥� Professionals working in organizations with a stated AI strategy are 3-times more likely to say AI is meeting or exceeding expectations for creating value at work compared to those at organizations without a stated AI strategy. This makes the value of establishing a stated AI strategy and ensuring its visible on a daily basis, a clear step for organizational leadership.

For professionals:

• • • Know which future you are working toward听鈥� Almost all professionals say they can see a future in one of the three paths. Understanding which one fits you is the first step to having a productive conversation about where the inconsistency is between your preference and your organization’s direction.
    • Invest in judgment as well as AI tool fluency听鈥� Judgment will always be a human differentiator in regard to AI, and the judgment that professionals apply on top of AI competency is what builds lasting professional value.

Professionals and organizational leaders need to make decisions in regard to their relationship with AI 鈥� and these decisions will determine the future of professional services. Those employers and professionals that choose their path deliberately, work to ensure alignment between strategy and experience, and invest in human judgment as seriously as they invest in technology will be the ones that can turn AI’s promise into a competitive advantage that compounds over time.

You can explore the full听

Key takeaways:

• • • The dulling effect is real, but the root cause is older than AI 鈥� Wherever gatekeeping institutions reward a narrow formula, their output converges long before any chatbot enters the picture. AI then accelerates optimization toward your already selected criteria.

    • Using AI for efficiency alone leaves the creative upside on the table 鈥� While most organizations deploy AI for simple drafting tasks, the bigger payoff comes from using it as a discussion engine 鈥� a sort of sparring partner that pressure-tests ideas and pushes thinking past the first plausible answer.

    • The highest-leverage intervention is reforming what you reward 鈥� The fix for this is upstream of the technology, and it comes from giving people time to make sure unconventional ideas actually survive your organization’s sorting mechanisms.

A tension sits at the center of nearly every serious conversation about AI and organizational strategy, and most leaders can feel it even if they haven’t named it yet.

On one side is the promise that AI can make teams more creative. That it can accelerate brainstorming, provide deeper research, identify hidden connections, and pressure-test ideas before they reach a client or a boardroom. When used well, AI is not a replacement for thinking but an amplifier of it.

On the other side, of course, is the fear that regular AI use is quietly dulling the creativity it’s supposed to enhance. That real fear is that the more people lean on these tools, the more their thinking converges toward the same polished, plausible, and fundamentally safe middle ground 鈥� and that the less people work their creative muscles, the more they atrophy without them realizing it. This trade-off, swapping originality for efficiency, is a losing exchange.

Both of these intuitions are reasonable and both are, to varying degrees, correct. However, they aren’t equally weighted. The purely cautious camp is taking the bigger gamble, because any competitor that cracks the problem by figuring out how to capture AI’s creative upside while managing the dulling effect gets both the innovation edge and the efficiency gains. The cautious organization doesn’t just miss the upside, it falls behind on both fronts.

The catch is that cracking the problem requires correctly diagnosing what’s actually killing your creativity 鈥� and a prominent recent essay on this exact topic gets it instructively wrong.

A good question, poorly tested

Rebecca Winthrop, a senior fellow at the Brookings Institution and director of its Center for Universal Education, recently published in The New York Times arguing that AI is constricting creative thinking. Her central claim is that while chatbots produce polished language, they鈥檙e masking a narrowing range of underlying ideas 鈥� and this is especially dangerous for students, whose creative development is still taking shape.

The piece is worth reading, and not just as a foil. Winthrop draws on from Georgetown neuroscientist Adam Green, whose team has been tracking the range of ideas in college application essays before and after ChatGPT’s release. Green’s findings related to the before/after tracking study (which have not yet been peer-reviewed) are striking, finding that while post-ChatGPT essays used more diverse and colorful vocabulary, the ideas beneath that language converged. Human judges rated the AI-era essays as more creative, even though the substance had narrowed. In a separate study by Green’s team, cited by Winthrop, human-written essays contributed up to eight-times more novel ideas than AI-generated ones.

The fear is that regular AI use is quietly dulling the creativity it’s supposed to enhance, and the real fear is that the more people lean on these tools, the more their thinking converges toward the same polished, plausible, and fundamentally safe middle ground 鈥� and the less people work their creative muscles.

And Winthrop flags serious concerns that deserve far more attention than they typically get. For example, AI’s homogenizing pressure falls hardest on those students who sit farthest from the mainstream, including neurodivergent students and those from racial and linguistic minorities. That finding alone should be shaping education policy conversations and acting as a warning for innovation-conscious reformers.

Here’s where Winthrop鈥檚 piece stumbles, however, and where it becomes a cautionary tale for organizations that may be thinking about their own AI and innovation strategies. The evidence Winthrop chooses to build her case on 鈥� the college admissions essay 鈥� is possibly the worst genre in American education for measuring whether AI is killing creativity. Because the creativity in college admissions essays was already dead.

I should know. I鈥檓 one of its murderers.

The most templated genre in America

The college admissions personal statement has been reverse-engineered for decades. Well before any large language model existed, applicants had cracked the code: Be damaged, but not too damaged; be resilient but make it look like you did it yourself; and be whole now, because the institution wants guaranteed successes, not risky projects. And all of this must be delivered in a tone that makes the committee feel good about their institution’s role in a meritocratic society. Deviate from this formula and you’re taking a risk, but hit every beat and you’re in the pile that moves forward.

I know this because I lived it recently enough to still remember the specific frustration of trying to fit my own experiences into that template at the age of 17, twisting and contorting experiences I’d actually lived through into the shape I knew admissions readers were looking for while sanding away the human beneath when it didn鈥檛 fit. The authentic version of my story wasn’t what they wanted, the version that hit the beats was.

And there’s a further detail conspicuously absent from Winthrop’s essay: The college admissions consulting industry. It’s enormous, it’s been around for decades, and its entire business model is teaching applicants to write to the template. Some of these consultants charge $5,000 or more, and their product isn’t creativity, it’s optimization. They teach students to identify what the admissions committee rewards and deliver exactly that, with the rough edges smoothed away and the personal experiences torqued into the right emotional shape.

My family took this seriously enough to invest in help, and I was fortunate enough they had the means to do so. I had one of those consultants. Mine cost $2,000, and my parents had to sell my mom’s pinball machine to pay for it. I think sometimes about what it says that the path to higher education ran through a professional who taught me, essentially, to write to a formula rather than to present myself in a way that would have given the committee a more honest, unique portrayal of just who they were letting into their institution. The consultant didn’t make me less creative; the system that made the consultant necessary did.

And this is the blind spot in Winthrop’s argument. She treats pre-ChatGPT essays as the baseline for authentic creative expression, but that baseline was already shaped by an industry dedicated to template optimization. So when Green’s research finds that post-ChatGPT essays use richer vocabulary but converge on familiar ideas, the question worth asking isn’t just whether AI caused a measurable shift (Green’s controlled experiments suggest it did) but whether the underlying ideas were already converged at a more fundamental level that the metrics don’t capture. In essence, all AI may have done is make that convergence more visible while democratizing the surface polish.

There’s an entirely different version of Winthrop’s essay waiting to be written 鈥� one in which the same data tells a democratization story rather than an erosion story. Where a free chatbot gives a first-generation college student the same surface-level advantage that a $5,000 consultant gave wealthier applicants for years. That’s not a comfortable reframe for institutions already invested in the idea that their selection processes brings forth authentic individuality 鈥� but it’s the reframe the data actually supports.

The template always comes first

This isn’t unique to college admissions. Wherever institution rewards a narrow formula, it gets gamed 鈥� and the gaming predates whatever technology that has made it easier.

The video essayist Sarah Z traced exactly this pattern in , which makes the gap in Winthrop鈥檚 argument clearer. When the publishing industry rewarded a specific shape of trauma narrative in the 1980s and 鈥�90s 鈥� suffering resolved through individual resilience 鈥� the template grew so predictable that fabricators outcompeted honest writers. Laurel Rose Willson sold a satanic-ritual-abuse memoir and, years later, a Holocaust-survival story, citing the same self-inflicted wounds as evidence for both. Publishing houses weren’t fooled just because they were careless, they were fooled because they’d built a machine that searched for formula 鈥� and the system that rewards a narrow pattern is the same one that makes it exploitable.

AI doesn’t create that convergence, it just accelerates the optimization toward whatever you’re already selecting for. Blaming AI for homogenized output in an already-homogenized system is like blaming your GPS for traffic on the BQE 鈥� the bottleneck was there long before the tool arrived.

If your organization has ever received a stack of pitch decks, strategy memos, or RFP responses that all hit the same beats in the same order, congratulations! You’ve built your own admissions committee, but don鈥檛 blame AI.

AI doesn’t create that convergence; it just accelerates the optimization toward whatever you’re already selecting for. Blaming AI for homogenized output in an already-homogenized system is like blaming your GPS for traffic on the BQE. The bottleneck was there long before the tool arrived.

Threading the needle

Of course, none of this means the concern about AI and creativity is unfounded. The dulling effect is real, and anyone who uses these tools regularly has probably felt its subtle gravitational pull toward the center. Or in the way a chatbot’s first suggestion can quietly foreclose any other directions you might have explored on your own, or how it may produce something that sounds polished but carries none of your voice

A different research team 鈥� Anil Doshi and Oliver Hauser, behind the , Winthrop herself points to 鈥� put a name to the mechanism, anchoring. Handed an AI-generated idea, writers locked onto it, narrowing the range of what they produced before they’d really begun.

However, the solution on an organizational level isn’t to restrict the tool; rather it’s to address the institutional and behavioral factors that determine whether the tool narrows thinking or expands it.

In this determination, three things matter most:

First, use AI as a discussion engine, not just an automation tool 鈥� There’s a meaningful difference between asking a chatbot to draft something for you and using it to create something with you. This article is a case in point. I didn’t read Winthrop’s essay and immediately decide to write a response. Instead, I spent almost half an hour talking to Claude about the article, debating the argument, testing my objections, diving into Green鈥檚 research more deeply, and connecting the piece to ideas I’d been thinking about from completely different contexts, such as the Sarah Z essay. This article emerged from that conversation unintentionally, and it would not have existed without it.

Further, the ideas emerged pressure-tested and sharpened through a process that felt more like sparring than delegation 鈥� and that鈥檚 exactly the kind of process organizations should be targeting. Most organizations deploying AI are using it for efficiency 鈥� drafting, summarizing, formatting 鈥� and that’s fine. However, if that’s all you’re doing, you’re leaving the creative upside untouched, and your people are feeling the dulling effect without the compensating benefit. It takes an intentional push from leadership to get teams using AI as a thinking partner rather than a shortcut.

Second, give people time 鈥� This sounds obvious, but it matters specifically because of how AI interacts with time pressure. When people are rushing, they take the first adequate output and move on. With traditional workflows, shortcuts save time at the cost of quality or risk. With AI-assisted workflows, however, shortcuts save time at the cost of originality, because the first output from a chatbot is almost always the most conventional one. It’s the statistically average response, and reaching the edges takes iteration, pushback, and follow-up prompts that challenge the initial direction. That takes time, and if your people don’t have it, they’ll use AI the way a stressed applicant uses a college essay consultant, producing the safest possible version of whatever the system rewards rather than the innovative one which could change the game.

Third, reform what you reward 鈥� This is the intervention that actually addresses the root cause, and it’s the one most organizations will resist because it requires examining their own sorting mechanisms. If your evaluation criteria, your promotion structures, your review processes, and your RFP scoring rubrics all select for the safe and conventional, then AI will only turbocharge that selection.

You’ll get the template faster and more polished than ever, much like the admissions committee that rewards a narrow emotional arc and gets 300,000 identical essays. Or, if your firm rewards the pitch deck that hits every expected beat and takes no risks, AI will produce that pitch deck beautifully 鈥� and you’ll wonder why innovation has stalled.

Again, the intervention is upstream of the tool. What does your organization actually do when someone brings in an unconventional idea? What happens to the proposal that doesn’t fit the template? If the answer is that it gets smoothed out in review or tossed altogether, that’s not an AI problem.

The old traps didn’t disappear

Winthrop is right that creative thinking is something to protect and nurture. She’s also right that AI introduces new pressures that deserve serious attention. And she’s right that the stakes are highest for the people whose perspectives are already farthest from the mainstream.

But the college admissions essay wasn’t homogenized by ChatGPT, it was homogenized by decades of institutional selection pressure that rewarded a single template and penalized everything that didn’t fit. AI didn’t create that problem, it just made the template accessible to everyone, including the families that couldn’t previously afford $2,000 and a pinball machine to get their kid across the threshold.

Similarly, your organization’s creative output won’t be determined simply by which AI tools you adopt. It will be determined by what your leadership rewards, what your processes select for, and whether your people have the time and incentive to push past the first plausible AI-supplied answer.

The technology is new, but the traps are very old. And if you want to use AI to make your organization more innovative, the place to start isn’t the tool 鈥� it’s the template.

You can find more听听from the 成人VR视频 Institute here

Key insights:

• • • Law firms have moved from restricting GenAI use (Don鈥檛 use tools that leak client data) to mandating it (Incorporate AI into your practice and market our firm鈥檚 GenAI capabilities)鈥斕齆either phase has given rank and file lawyers what they really need: Guidance on in which instances GenAI actually helps deliver better, cheaper, and faster legal services, where it introduces serious professional risk, and how to tell the difference.

    • GenAI鈥檚 capacity to transform legal work for the better is real, but so is its capacity to degrade it听鈥擥enAI can significantly boost speed and quality on tasks involving breadth, synthesis, or straightforward analysis, but it can weaken performance on complex judgment and revision tasks 鈥� especially for stronger professionals 鈥� by encouraging overconfidence, missed issues, and superficial reasoning.

    • A use-mode framework can close the听gap鈥� A proposed governance framework can give law firm leadership a practical tool for identifying in which situations GenAI enhances legal work, where it introduces serious risk, and where professional judgment is non-negotiable.

This article synthesizes findings from the author鈥檚 paper,

Your law firm undoubtedly has a policy around generative AI (GenAI), which probably tells lawyers to avoid tools that leak client data, admonishes them to look out for hallucinations, and encourages them to incorporate AI into their practice to satisfy client demands.

However, it likely does not tell them which cognitive functions they should delegate to GenAI, which they should not, and where the line between the two is absolute. In the space between restriction and mandate, lawyers are making consequential decisions about GenAI delegation every day. Meanwhile, most law firms have not addressed that space with meaningful governance.

GenAI can make legal work worse

GenAI鈥檚 capacity to transform legal work for the better is real, but so is its capacity to degrade it. Most law firm leaders know that AI can hallucinate; yet far fewer know that it can make expert legal judgment and work product actively worse.

The best evidence of this dynamic comes from a with consultants from the Boston Consulting Group, who were given similar tasks and allowed to use various levels of AI assistance, including no AI. For professional tasks requiring breadth and option generation, GenAI delivered, showing that output quality improved by 40% and consultants worked faster. For tasks requiring judgment and synthesis, however, something unexpected happened. Consultants using GenAI were 19% less likely to produce correct solutions than those working without it.

Governing GenAI鈥檚 uneven performance requires asking a question that most law firms are not asking: What cognitive function is being delegated to GenAI at each step in the workflow?

The same pattern appears in research evaluating GenAI use in legal analysis. An empirical in the Journal of Legal Education confirmed that AI dramatically improves performance on straightforward analysis while producing no measurable benefit for complex reasoning. And in the case of complex reasoning, GenAI use also introduced recurring failures, such as jumping to conclusions, missing less obvious issues, and generating confident prose that masks superficial analysis.

from the University of Minnesota focused on legal tasks showed that GenAI assistance on a synthesis task improved performance by nearly 60% and produced a surprising downstream benefit. Those participants who used AI for synthesis outperformed the control group on the subsequent independent reasoning task even after GenAI was removed. However, when GenAI was introduced at the revision stage, the picture changed. GenAI helped weaker performers, but it actively degraded the work of stronger ones. Indeed, the best lawyers in the study produced worse revised work product when they used GenAI than when they worked without it.

A use-mode governance framework

Given all these findings, governing GenAI鈥檚 uneven performance requires asking a question that most law firms are not asking. Instead of determining whether GenAI is appropriate for a particular deliverable 鈥� such as a brief, a contract, or a board presentation 鈥� the governance question instead should be: What cognitive function is being delegated to GenAI at each step in the workflow?

My proposed framework, outlined below, organizes common GenAI uses into seven recurring modes following the sequence in which lawyers actually use GenAI to produce legal work product. Then, governance controls are calibrated to the risk profile of each mode.

Modes 1 and 2: Retrieval and organization

At the mechanical end of the cognitive spectrum are two distinct functions. In retrieval mode (Mode 1), a lawyer reviewing a merger agreement asks GenAI to identify every representation and warranty in the document. In organization mode (Mode 2), a litigator reviewing 50 depositions asks GenAI to construct a timeline from the testimony. The first locates material that already exists. The second arranges it into a usable structure. No new content is created in either case, and both uses are low-risk and should be actively encouraged, subject to modest verification controls. Firms that unduly restrict these use modes are leaving value on the table.

Mode 3: Summarization

Summarization (Mode 3) introduces selection risk. In this mode, GenAI chooses what to emphasize, include, and omit. Consider a lawyer preparing a board presentation on the results of an internal investigation. GenAI can condense dozens of witness interviews into key points and themes in minutes; however, a summary may focus on procedural detail while missing credibility issues that a lawyer would immediately recognize as material. The appropriate control is to mandate meaningful review by a lawyer with first-hand knowledge of the source material. A lawyer encountering the summary cold has no reliable way to evaluate what GenAI missed.

Mode 4: Candidate generation

Mode 4 is exploratory. A lawyer drafting a brief might ask GenAI to generate a list of potential arguments, propose alternative framings, or identify supporting authority. This candidate material expands options and accelerates iteration. The work product is not filing-ready and must be treated as provisional. GenAI can suggest, but a lawyer must decide.

The authority verification obligation at this stage deserves special emphasis. GenAI will identify cases, summarize holdings, and weave them into an argument structure. Thus, the output will read fluently and cite real-looking cases. However, a lawyer cannot assume the model has accurately characterized the holdings or context, and any authority cited in an external filing must be independently read and verified. GenAI can help find the cases, but a lawyer must read and apply them.

Mode 5: Editing and rewriting

In Mode 5, a lawyer asks GenAI to tighten a dense contract provision or restructure a wordy paragraph, risking, of course, unintended meaning change. An edit may read cleanly while subtly narrowing a representation, softening a covenant, or eliminating a carve-out. The revision risk is not hypothetical. The University of Minnesota study referenced above found that stronger performers produced worse work product when GenAI revised their independently produced memos. In this mode, a lawyer must confirm that the edit produced no shift in meaning and introduced no new factual assertions.

Mode 6: Critique and stress-testing

Mode 6 may be the most underutilized GenAI capability. Before filing a brief or presenting to regulators, a lawyer can ask GenAI to identify weaknesses in their argument. In this way, GenAI finds vulnerabilities before adversaries do; and unlike every other mode, the risk here runs in one direction. Lawyers who skip this step are missing one of GenAI鈥檚 core value propositions. Law firms鈥� governance frameworks should not merely permit it but actually require it in appropriate cases.

Mode 7: Evaluation and decision

The boundary against AI delegation becomes absolute when GenAI is asked to evaluate or decide. A lawyer advising a board on whether an event requires disclosure cannot delegate that determination to GenAI. A litigator assessing settlement value cannot outsource probability judgments because these are core expressions of professional responsibility. In this mode, GenAI may inform background analysis, but it may not substitute for lawyer judgment in making the call. This is a categorical prohibition 鈥� professional judgment cannot be delegated.

Going forward with GenAI

Law firm leaders who have moved their GenAI policy from restriction to mandate without governing the space between have not finished the job. Their lawyers are making consequential decisions about GenAI use every day without the guidance they need and deserve.

The use-mode framework presented above gives firm leadership a practical tool for filling that gap. It identifies the instances in which GenAI enhances legal work, where it introduces serious risk, and where professional judgment is non-negotiable. Firms that govern at that level will capture GenAI鈥檚 value; and those firms that do not will have policies that look serious but govern nothing important.

The views expressed in this article are solely those of the author in his individual capacity and do not represent the views, positions, or opinions of Foley & Lardner LLP, its partners or clients, or the University of Wisconsin Law School.

Key takeaways:

• • • Build due diligence into the process 鈥� Make human-rights due diligence routine from the decision to build or buy through deployment by mapping uses to standards, assess severity and likelihood, and close control gaps to prevent costly pullbacks and reputational damage.

    • Identify risks early on 鈥� Use practical methods to identify risks early by engaging end users and running responsible foresight and bad headlines

    • Use due diligence to build trust 鈥� Treat due diligence as an asset and not a compliance box to tick by using it to de鈥憆isk launches, uncover user needs, and build durable trust that accelerates growth and differentiates the product with safety-by-design features that matter to buyers, regulators, and end users.

AI is reshaping how we work, govern, and care for one another. Indeed, individuals are turning to cutting-edge large language models (LLMs) to ask for emotional help and support in grieving and coping during difficult times. 鈥淯sers are turning to chatbots for therapy, crisis support, and reassurance, and this exposes design choices that now touch the right to information, privacy, and life itself,鈥� says , co-Founder & Principal at , a management consulting firm that specializes in human rights and responsible technology use.

These unexpected uses of AI are reframing risk because in these instances, safeguards cannot be an afterthought. Analyzing who might misuse AI alongside determining who will benefit from its use must be built into the design process.

To put this requirement into practice, a human rights lens must be applied across the entire AI lifecycle from the decision to build or buy to deployment and use, to help companies anticipate harms, prioritize safeguards, and earn durable trust without hampering innovation.

Understanding human rights risks in the AI lifecycle

Human rights risks can surface at every phase of the AI lifecycle. In fact, they have emerged in efforts to train these frontier LLMs in content moderation functions and now, are showing up elsewhere. For example, data enrichment workers who refine training data, and data center staff, who power these systems, are most likely to face labor risks. Often located in lower鈥慽ncome markets with weaker protections, they face low wages, unsafe conditions, and limits on other freedoms.

During the development phase, biased training sets and the probabilistic nature of models can generate misinformation or hallucinations, and these can further undermine rights to health and political participation. Likewise, design choices often can translate into discriminatory outcomes.

Unfortunately, the use of AI-enabled tools also can compound these harms. Powerful models can be misused for fraud or human trafficking. In addition, deeper integration with sensitive data can heighten privacy and security risks.

A surprising field pattern exacerbates the risk when people increasingly use AI for therapy鈥憀ike support and disclose issues related to emotional crises and self鈥慼arm. In particular, this intimacy widens product and policy obligations, which include age鈥慳ware safeguards and clear limits on overriding protections.

Why human rights due diligence is urgent

That鈥檚 why human rights due diligence must start with people, not the enterprise. By embedding human rights due diligence into the lifecycle of AI, development teams can begin to understand the technology and its intended uses, then map those uses to international standards. Next, a cross functional team gathers to weigh benefits alongside harms and to consider unintended uses. Primarily, they need to answer the question, 鈥淲hat happens if this technology gets in the hands of a bad actor?”

From there, the process demands an analysis of severity 鈥� which assesses scale, scope, and remediation, and the likelihood of each use. The final step involves evaluating current controls across supply chains, model design, deployment, and use-phases to identify gaps.

The biggest barrier in layering in a human rights lens into to AI is the need for speed to market. The races to put out minimally viable products accompanied by competitive pressure can eclipse robust governance, yet early due diligence may prevent costly pullbacks and bad headlines. Article One鈥檚 Poynton notes that no one wants to see their product on the front page for enabling stalking or spreading disinformation. Building safeguards early “ensures that when it does launch, it has the trust of its users,” she adds.

How to embed safeguards without slowing teams

The most efficient path in translating human rights into the AI product lifecycle is to turn policy principles, goals, and ambitions into actionable steps for the engineers and the product teams. This requires the 鈥渆ngineers to analyze how they do their work differently to ensure these principles live and breathe in AI-enabled products,鈥� Poynton explains. More specifically, this includes:

Identifying unexpected harms 鈥� One of the most critical, yet difficult components of the human rights impact assessment is brainstorming potential harms. Poynton recommends two ways to make this happen: First, engage with end users to help identify potential harms by asking, 鈥淲hat are some issues that we may not be considering from the perspectives of accessibility, trust, safety and privacy?鈥� Second, run responsible foresight workshops at which individuals play the parts of bad actors to better identify harms and uncover mitigation strategies quickly. Pair that with a bad headlines exercise that can be used to anticipate front鈥憄age failures. Then, ship with these protections in place, pre鈥憀aunch.

Implementing concrete controls 鈥� Embedding safety-by-design should cover both content and contact, a lesson from gaming in which grooming risks require more than just filters. Build age鈥慳ware and self鈥慼arm protocols, including parental controls and principled policies on overrides. Govern sales and access with customer vetting, usage restrictions, and clear abuse鈥憆esponse pathways. In the supply chain, set supplier standards for enrichment and data center work that include fair wages, safe conditions, freedom of association, and grievance channels.

Treating due diligence as value-creating, not box-checking 鈥� Crucially, frame due diligence as an asset rather than a liability. 鈥淢ake your product better and ensure that when it does launch, it has the trust of its users,” Poynton adds.

Additional considerations

Addressing equity must be front and center. Responsible strategies include diversifying training sets without exploiting communities and giving buyers clear provenance statements on data scope and limits.

Bridging the digital divide is equally urgent. Bandwidth and device gaps risk amplifying inequality if design and deployment assume privileged contexts. In the workplace, Poynton stresses that these impacts will be compounded, from entry-level to expert roles.

Finally, remember that AI鈥檚 environmental footprint is a human rights issue. “There is a human right to a clean and healthy environment,” Poynton notes, adding that energy and water demands must be measured, reduced, and sited with respect for local communities, even as AI helps accelerate the clean energy transition. This is a proactive mandate.

You can find out more about the ethical issues facing AI use and adoption here

Key takeaways:

• • • Shifting from engineering to architecture 鈥� Focusing solely on building better AI models and engineering solutions leads to isolated, non-reusable outputs. Instead, organizations should build AI into the broader enterprise, emphasizing reusable, machine-readable intelligence that integrates with business operations and data structures.

    • Regulation as opportunity for reusability and efficiency 鈥� Regulatory frameworks are not just compliance burdens; they also are catalysts for sustainable AI. By mandating standardized, machine-readable data, these regulations force organizations to design systems for reuse, enabling operational efficiency and scalable innovation.

    • Reusable enterprise is the path to sustainable reinvention 鈥� The future of AI leadership lies in building adaptable, reusable data and AI infrastructures. When standardized data, AI models, and regulatory compliance reinforce each other, organizations can continuously reinvent themselves, support multiple business outcomes from the same information assets, and achieve compound returns on their investments.

Across industries, executives are confronting an uncomfortable truth: AI projects are delivering outputs, not outcomes.

For years, organizations have poured time and capital into the mechanics of AI 鈥� the algorithms, the computation power, the data pipelines, and the engineering teams to support them. Yet results remain uneven. Models keep getting larger, but lasting, reusable business value hasn鈥檛 followed.

The problem isn鈥檛 the math, it鈥檚 the mindset.

Too many enterprises have tried to engineer AI into existence instead of architecting it into the enterprise. The focus has been on perfecting models, not integrating them into the broader data and operational fabric of the business. The assumption has been that a technically superior model naturally creates a competitive edge. It doesn鈥檛.

Without consistent governance, shared definitions, and reusable data structures, every AI initiative becomes its own isolated experiment. One line of business builds a credit-risk model. Another develops an environmental, social, and governance (ESG) classifier. A third deploys a generative assistant for customer support. Each team moves fast, but none build on each other鈥檚 work. The result is a proliferation of proofs of concept 鈥� impressive on paper but disconnected in practice.

For years, organizations have poured time and capital into the mechanics of AI 鈥� the algorithms, the computation power, the data pipelines, and the engineering teams to support them. Yet results remain uneven.

And this fragmentation carries a financial cost. Every new model adds complexity 鈥� new pipelines, new monitoring requirements, and additional governance checkpoints. These systems rarely scale together, and as integration demands grow, executives find themselves in a paradox: Make massive investments in AI infrastructure yet see declining agility and uncertain ROI.

The AI engineering mindset has optimized the structural parts, not the whole when it comes to a production solution set. In general, it has produced models that predict, but not organizations that learn.

In short, the AI engineering mindset has reached its limit 鈥� a sign that AI is entering sustainable growth cycles. Many leaders are beginning to realize that they don鈥檛 need more AI engineers, rather they need system designers who can embed intelligence into reusable business frameworks 鈥� all while navigating a regulatory environment increasingly defined by machine-readable data standards such as the Financial Data Transparency Act (FDTA) and Standard Business Reporting (SBR).

Regulation as catalyst, not constraint

At first glance, FDTA and SBR may appear to be just another layer of regulatory complexity. They are not. In fact, they represent one of the most powerful architectural opportunities available to organizations today.

By mandating machine-readable data standards, these frameworks force companies to design for reuse. They turn what once felt like a compliance exercise into an infrastructure strategy 鈥� one that connects regulatory requirements directly to operational efficiency. Build once. Reuse often.

For decades, compliance has been treated as a cost of doing business. Under FDTA and SBR, it can become the scaffolding of reinvention. Machine-readable, standardized data provides the foundation for models that are verifiable, shareable, and reusable across domains. Reporting ceases to be an afterthought and becomes a living data layer that fuels forecasting, stress testing, and product innovation.

When viewed through this lens, regulation isn鈥檛 an obstacle; it鈥檚 the blueprint for sustainable AI. It forces clarity, consistency, and interoperability 鈥� qualities every enterprise says it wants, but few achieve voluntarily. Regulation may finally deliver what AI engineering alone could not: The discipline of reusability.

From proofs of concept to proofs of architecture

For most organizations, AI success has been measured by the number of proofs of concept completed, or how fast a model moves into production. However, the real test of maturity isn鈥檛 how many experiments you run, it鈥檚 how easily those experiments can be scaled, reused, or extended.

That鈥檚 where the next evolution lies. We are now shifting from proofs of concept to proofs of architecture. And that means the question leaders should be asking isn鈥檛, 鈥�Did it work once?鈥� but 鈥�Can it work again, and with half the effort?鈥� Only when a single domain鈥檚 data can serve multiple regulatory, compliance, and analytical purposes, can the enterprise start to gain compound returns on its information assets.

When viewed through this lens, regulation isn鈥檛 an obstacle; it鈥檚 the blueprint for sustainable AI. It forces clarity, consistency, and interoperability 鈥� qualities every enterprise says it wants, but few achieve voluntarily.

This approach turns data from a static resource into a dynamic capability. AI is no longer something you deploy; rather, it鈥檚 something you design for reuse.

Engineering adaptability

Organizations that embrace this shift are learning to engineer adaptability rather than one-off innovation. Their data and AI systems act like interchangeable components, each capable of supporting new regulations, mergers, or market disruptions without starting from scratch.

Some industry examples of this development include:

• • • Financial services 鈥� Stress-testing data used for regulatory compliance can also inform pricing analytics and liquidity simulations, reducing cycle time between audit and strategy.
    • Healthcare 鈥� Patient outcome models built for quality reporting can be reused to predict staffing needs or optimize clinical supply chains, extending beyond compliance and into operations.
    • Legal and compliance sectors 鈥� AI used for document classification under discovery protocols can be repurposed for internal policy audits or ESG disclosure mapping, turning regulatory data into a strategic asset.
    • Manufacturing and supply chain 鈥� Sensor and maintenance data initially used for safety reporting can drive predictive production planning and carbon-emission forecasting under emerging sustainability standards.
    • Public sector and critical infrastructure 鈥� Data collected for transparency and open-data mandates can be reused to model risk exposure across utilities, cybersecurity, and climate resilience programs.

In each of these cases, the same information infrastructure supports different outcomes. That鈥檚 the hallmark of a reusable enterprise.

The above chart鈥檚 interconnected components illustrate how standardized data, reusable AI, and regulatory compliance can reinforce one another to create a continuous cycle of enterprise reinvention 鈥� standardized data supports reusable AI, which in turn enhances reporting and regulatory alignment. The result is a virtuous loop that replaces isolated projects with scalable, data-driven reinvention.

A call to reusable leadership

The next phase of digital leadership won鈥檛 be defined by how sophisticated a company鈥檚 models are, but instead by how seamlessly those models integrate into decision-making.

The leaders who succeed will be those who align AI investments with evolving regulatory and data standards. Their organizations will speak a common data language in which AI, compliance, and analytics operate within a shared architectural framework.

As FDTA and SBR converge globally, the line between compliance and competitiveness will blur. What once felt like regulatory overhead will become the foundation of reusable intelligence. Reinvention, in this sense, isn鈥檛 a campaign or initiative 鈥� it鈥檚 a discipline. This is not AI as a project; it鈥檚 AI as infrastructure and the architecture of continuous reinvention.

For executives navigating 2026鈥檚 convergence of regulation, consolidation, and automation, the difference between thriving and merely surviving will depend on whether they can build organizations that learn, adapt, and continuously reinvent themselves through data.

You can find more blog posts听by this author here

Key takeaways:

• • • Embrace value, risk, and execution 鈥� for good and bad 鈥斕齈rofessional services firms must weigh the value of AI applications against potential risks, embracing both successes and failures as learning opportunities to improve responsible adoption.

    • Ethical oversight is everyone鈥檚 responsibility 鈥斕鼸nsuring responsible AI use in professional services requires active participation from all members of an organization, not just legal or IT teams.

    • Human creativity and feedback remain essential 鈥斕齏hile AI can generate ideas and accelerate processes, human judgment, creativity, and continuous feedback provide the proper pathways for ethical decision-making and successful integration.

AUSTIN, Texas 鈥� With the professional services world now squarely into the AI era, it鈥檚 clear that the speed of business is quicker than ever. Clients expect results in hours or even minutes rather than days, while generating documents can happen at the click of a button. Ask a research question, and a machine can intuit what you鈥檙e looking for with striking accuracy.

Alongside these business changes, however, it鈥檚 clear that the ethics of technology usage within professional services is shifting just as quickly. 鈥淓very time you come and do a talk with a group of people, within four weeks if not sooner, it鈥檚 changed,鈥� says Betsy Greytok, Associate General Counsel in Responsible Technology at IBM. 鈥淪o, it really does require you to keep on your toes.鈥�

Ensuring that AI is used responsibly is paramount within professional services than in other professions, given the ethical and regulatory constraints placed on legal, tax, audit & accounting, financial services and risk, and more. During a recent session, A Unified Field: Ethical Considerations amid AI Development and Deployment, at the 成人VR视频 Institute鈥檚 2025 Emerging Technology and Generative AI Forum, panelists describe an ethical world that should be tackled as a challenge, rather than shied away from as an unsolvable risk.

Or, as Paige L. Fults, Head of School at the AI-centric Alpha School & 2-Hour Learning program, put it: 鈥淣ot being afraid of replacement, but leaning into repurpose.鈥�

Embracing success 鈥� and failure

John Dubois,听the Americas AI Strategy Leader at Big 4 consultancy Ernst & Young, says he regularly gets questions from customers about AI and how they should use it, given that there are new AI applications arising seemingly every day. 鈥淭he way we describe it is a balance,鈥� Dubois explains. 鈥淟et鈥檚 start with value. If we know there鈥檚 value in something, then we can figure out the risk behind it, then we can figure out how we can execute.鈥�

Just as importantly, however, this focus on value, risk, and execution can also aid professional services firms when an AI plan fails. For example, Dubois cites an MIT report from August 2025 that showed , often because of flawed integration. Embracing the value, risk, and execution strategy from the beginning not only allows for better chances of success, but even in the event of failure, 鈥渨e actually have a better shot at mitigating, when it does fall down.鈥�

This sort of planning is not limited to just one group, Dubois says, noting that ethical oversight is seen as a key responsibility of everyone in the organization. He explains that E&Y has an internal implementation of OpenAI that has 150,000 distinct users each month. Because of an internal process called SCORE that removes customer data at the source, E&Y鈥檚 instance of OpenAI is largely clear of customer data 鈥� but it鈥檚 still not perfect.

E&Y has set a culture so that if someone sees proprietary data when using GenAI to develop a proposal or create a PowerPoint, they not only delete the data before use, but work to scrub it from the system entirely. 鈥淚t is all of our job to ensure that whatever you鈥檙e putting into that system or extracting out of that system, you鈥檙e cleansing,鈥� Dubois says. 鈥淚t鈥檚 not the job of the general counsel, or the risk team, or the IT team, it鈥檚 all of our job.鈥�

When it comes to keeping up with AI ethics in a rapidly advancing space, professionals can rely on the same methods they have been employing for years to solve ethical quandaries: human creativity.

IBM鈥檚 Greytok agreed, noting that she鈥檚 part of an internal review board that examines major AI-related projects for ethical issues. There is a board review at the beginning of the development process to determine how risky a use case is, and then the system will give a response, considerations, and steps. If there is an issue, the board is empowered to stop development, even on a major project.

She drew an analogy to writing a paper in high school, in which there is a marked difference between simply turning in the paper, proofreading your own work, and asking a friend for peer review feedback. 鈥淭hat鈥檚 what you want, is that disagreement, because that鈥檚 critical thinking.鈥�

She adds: 鈥淭he researchers sometimes get so excited about what they鈥檝e discovered that they forget to look at the other side of what can happen. You should want that. You shouldn鈥檛 be punished for saying, Is this the right thing or not?鈥�

The importance of feedback

Fults says that at the Alpha School, AI is not only baked into the curriculum, it . Students spend just two hours a day on academics, led by AI tools that are supplemented by off-line learning on a variety of subjects by in-person instructors that fill in the gaps that AI is not able to provide.

It鈥檚 a revolutionary concept but not a static one. Fults notes that 鈥渢he two-hour learning model has already changed so much since I鈥檝e been part of the school,鈥� and the instructors have a Slack channel on ways to find improvement that receives hundreds of messages a day.

It鈥檚 through this marrying of human intuition and the possibilities of the technology that Fults says she believes the school has found success and used AI ethically within education. 鈥淓ven though we have this tool, the human levers, the motivational levers that are happening day to day, actually make it work,鈥� she says, insisting that she 鈥渃an鈥檛 just hand [the technology] to any school鈥� without the corresponding processes in place.

Dubois and Greytok also call feedback a crucial part of the process in order to overcome AI barriers. Dubois tells the story of a large retailer that bought satellite images to determine footfall within a store. Shoppers, however, felt that was a privacy risk, and the idea was almost scrapped. Then, however, the legal and IT teams worked together to come up with an idea: Can you track clothing, but not faces, to get the same information of where within the store shoppers were going?

鈥淚t鈥檚 a creative workaround to get us to the same thing,鈥� Dubois explains. 鈥淲hen you have a constraint, what鈥檚 a clever way to work around this so we鈥檙e not taking a brand risk or a compliance risk?鈥�

Indeed, when it comes to keeping up with AI ethics in a rapidly advancing space, professionals can rely on the same methods they have been employing for years to solve ethical quandaries: human creativity. AI can provide information and context more rapidly than ever before, but ultimately, professionals themselves will be the ones relied upon to make sure AI is used ethically and responsibly.

鈥淎I is an idea generator,鈥� Greytok says. 鈥淭he solution comes from the human.鈥�

You can find out more about how emerging technologies are impacting professional services here

Relying on an entity that operates strictly within these binary constraints to make complex decisions affecting real lives underscores the necessity of human involvement in the legal system. Therefore, it is imperative that any technology integrated into the judicial process incorporates human oversight, which should then ensure that regardless of the extent of AI integration, human participation remains essential to respect these nuances and uphold the integrity of the legal proceedings.

The judicial process encompasses various roles, some of which could theoretically be replaced or supported by AI. And this is even more probable when considering generative AI (GenAI), and its ability to complete tasks that can often be done by humans with little to no assistance.

While clerks, paralegals, and court administrators might employ AI to direct individuals to necessary information more promptly and consistently, it is conceivable that technology could be developed in which judges, juries, and mediators could be replaced by algorithms capable of analyzing facts and rendering decisions. Further, case workers, forensic experts, and probation officers could utilize AI to standardize pretrial decision-making processes. Transcripts may also achieve higher accuracy if generated by AI rather than traditional court reporters. Of course, all of this raises the question of whether the benefits outweigh the risks.

…It is conceivable that technology could be developed in which judges, juries, and mediators could be replaced by algorithms capable of analyzing facts and rendering decisions.

While each of these options presents certain advantages, it is crucial to maintain human involvement in the judicial process, explains of Louisiana’s Fifth Circuit Court of Appeal. “The practice of law isn’t simply about applying rules to facts,鈥� he says. 鈥淪imilar to the intricate political maneuvering , it requires a nuanced understanding, careful navigation of precedent, and the ability to craft arguments that resonate with human experience. It necessitates what lawyers often refer to as the feel of a case 鈥� an intuitive grasp of the issues derived from years of experience and critical analysis.”

How many humans?

In the current economy and political climate, courts 鈥� like many other government departments and agencies 鈥� are faced with the dilemma of providing greater service with less fiscal ability. Quite simply, as the number of staff declines, the number of cases is remaining steady or, in some jurisdictions, increasing. Enter the shiny toy of GenAI, having been developed in private industry and seemingly ripe to help the public sector. As adaptations begin, the first question is a compound one: What ought to be done? And how many people are necessary to get it done properly?

The initial step is to ensure a fully staffed IT department and a well-funded budget to develop and implement an effective program within the court system. This requires allocating resources to develop technology that can function properly within government programs. An assessment of this sort will provide a clearer indication of the number of personnel needed to implement those programs that could benefit each system.

Where in the loop?

The application of GenAI in the legal sector is demonstrated, for example, by the chatbots, created by the People’s Law School in Canada. The chatbot can answer basic legal questions, directing a person to the correct statute, rule, form, or other resource in seconds. Although the People’s Law School is not a court system, it examines the functionality of the court system to improve user interaction. Human involvement in the development and testing of such GenAI systems could help ensure that AI supports legal processes while maintaining human oversight.

As the use of chatbots progresses, it requires human review and verification of the outputs generated by these bots. This may involve programming the bots to refer specific issues to humans for resolution and periodic human assessment of the chatbot’s output. This places humans at the beginning of the loop, allowing for control of the output.

In other instances, AI can be used as a research or drafting aid. In these situations, the AI acts as a paralegal or first-year associate, meaning the human in the loop is a more experienced or well-trained individual. Humans can serve as intermediaries in this process, and it is crucial that humans do not become complacent with the work performed by the system and must always rely on their own expertise with the justice system.

Closing the loop

The fear voiced by most people in this process is over the other instances in which AI can be used in the legal process. For example, one key fear is that AI can become the final arbiter of the case 鈥� although it鈥檚 not a likely outcome, it is one which runs contrary to what most judges want.

In the current economy and political climate, courts 鈥� like many other government departments and agencies 鈥� are faced with the dilemma of providing greater service with less fiscal ability.

Judge Schlegel , noting that every day, courts determine who raises children, whether someone is evicted, and who goes to jail or receives a second chance at life. These aren’t abstract data points or business metrics; rather, they’re profound decisions that demand empathy, experience, and the kind of nuanced judgment that comes only from years of practice.

To this end, we have to be careful with new iterations of AI, such as agentic AI 鈥� which operates autonomously, making decisions and adapting to changes, similar to a human employee, while performing tasks with minimal supervision. Indeed, we have to prevent agentic AI from taking over the final part of the litigation process. The growth of agentic AI alone necessitates an important for discussion around maintaining human oversight in AI operations.

Examples of agentic AI include autonomous vehicles, virtual assistants, robotic process automation, AI in gaming characters, industrial robots, and algorithmic trading systems. Although these programs are advancing, their involvement in courts remains a distant prospect, thus far.

Conclusion

There will always be a human involved in the judicial process. From technical support to referral attorneys, human presence is essential to verify the work completed by AI. Therefore, it is crucial to train attorneys not only in their legal professions but also as proficient users of new technologies. Indeed, developing new AI models must prioritize both clarity and user-friendliness 鈥� this is not optional, but rather it is imperative for an effective system.

You can find more about听how courts are using AI-driven technology听here

You can hear more insights from Judge Maritza Dominguez Braswell on here

Let’s start with a brief overview of the two posts.

In an essay entitled,, Amodei describes 鈥渁 country of geniuses in a datacenter.鈥� He envisions very powerful AI that is 鈥渟marter than a Nobel Prize winner across most relevant fields,鈥� skilled enough to 鈥減rove unsolved mathematical theorems,鈥� and capable of directing experiments and executing many tasks fully autonomously. Noting the current slow pace of groundbreaking discoveries in biology and medicine, Amodei lands at his central point: 鈥減owerful AI could at least 10x the rate of these discoveries, giving us the next 50-100 years of biological progress in 5-10 years.鈥�

He refers to this as the 鈥渃ompressed 21st century鈥� because the progress of the entire 21st century will be possible within a few years. According to Amodei, this progress includes reliable prevention and treatment of nearly all natural infectious diseases, elimination of most cancers, prevention of Alzheimer鈥檚, and the potential to double our lifespan. Amodei admits his vision is radical, but believes most people underestimate 鈥渏ust how radical the upside of AI could be.鈥�

Wolf then challenges Amodei鈥檚 vision. In his blog post, , Wolf argues that today鈥檚 systems are fundamentally constrained by their training data. He describes AI models as very 鈥渙bedient students,鈥� but not genius revolutionaries capable of true paradigm shifts. To create a data center of true geniuses 鈥� Einstein-level geniuses 鈥� Wolf argues, 鈥渨e don’t just need a system that knows all the answers, but rather one that can ask questions nobody else has thought of or dared to ask.鈥�

What does this debate say about law?

The Amodei-Wolf debate, while focused on the technological possibilities of AI, led me to a different question: beyond what AI can do, what will we choose for it to do in our legal system?

Even assuming Amodei is right 鈥� that AI models will become capable of true genius 鈥� do we want those geniuses deployed in all aspects of law? Amodei ponders whether powerful AI could 鈥渋mprove our legal and judicial system by making decisions and processes more impartial[.]鈥� However, impartiality alone doesn’t guarantee justice. Can AI, however brilliant, truly understand the human condition and context that shapes legal outcomes? And what relevance do Wolf鈥檚 observations hold in the legal context? Is there a need for more than just answers? Do we also need legal professionals who ask the questions that 鈥渘obody else has thought of or dared to ask鈥�?

Think of Brown v. Board of Education, for example. Thurgood Marshall鈥檚 triumph was far more than a mechanical application of the law. It was a strategic challenge to decades of entrenched precedent, driven by a profound moral imperative. This required more than information-processing, it demanded the uniquely human courage to confront the status quo and the ingenuity to forge a new path.

The inherent duality of the law

This is where the legal field reveals its inherent duality. On the one hand, our work involves the methodical processing of vast datasets. We collect data (facts, evidence, legal precedent), we analyze this data, and then we generate new data 鈥� a process that mirrors the capabilities of generative AI models like ChatGPT and suggests that certain aspects of our work are ripe for automation.

On the other hand, many of our core functions 鈥� deeply rooted in human qualities 鈥� decisively counsel against AI overreliance. For instance:

• • • Judgment & discretion 鈥� The law is not a rigid set of rules. Indeed, many of our balancing tests call for the inexact weighing of various factors, requiring the careful exercise of judgment and discretion.
    • Advocacy & persuasion 鈥� Legal practice is about representing clients and persuading others. This requires empathy, emotional intelligence, and the ability to connect with human decision-makers 鈥� a judge, a jury, or opposing counsel. AI might mimic human connection, but it cannot truly achieve it because of course, it is not human.
    • Adapting to novelty 鈥� New technological, economic, and social frameworks require legal professionals to think creatively and adapt to situations with no precedent. Consider the rise of social media and its impact on the collection of evidence. Or the need to adapt established legal and regulatory frameworks to entirely new financial instruments. Current AI is trained on past data and may struggle with the truly new.
    • Careful reasoning & societal attunement 鈥� While the legal system must not be swayed by fleeting whims of public opinion, it also must possess the capacity to evolve alongside shared norms. This adaptation requires careful reasoning about justice and fairness; and, in light of changing social structures and values, we need to ensure the law remains both grounded in principle and responsive to societal needs.

Thus, while many of our tasks are amenable to automation, much of our work demands a uniquely human perspective. As AI becomes more capable and integrated into our workflows, the defining question will not be, 鈥淲hat can AI do in the legal profession?鈥� But rather, 鈥淲hat should AI do in the legal profession?鈥�

The Amodei-Wolf debate is a helpful reminder that AI is advancing quickly, and our choices will be critical. We cannot resign ourselves to the inevitability of AI; instead, we must approach it with a clear-eyed understanding of our agency. By defining clear boundaries, establishing ethical frameworks, and carefully integrating AI where appropriate, we can ensure it serves as a powerful instrument for justice, not a force that undermines it.

You can find out more about how听courts are using AI to improve efficiency and access to justice听here

For example, the developed the chatbot Beagle+ to assist people with step-by-step guidance on everyday legal problems. , Digital & Content Lead at the People鈥檚 Law School, led the efforts to create Beagle+ with technical assistance from听, Founder of Tangowork. And the Alaska Court System (ACS) and听 using a grant from the NCSC to develop an AI-powered chatbot called the听Alaska Virtual Assistant, or AVA.听Jeannie Sato, Director of Access to Justice Services of ACS worked with , CEO and Founder of LawDroid, to develop the tool.

How courts can successfully experiment with AI

Jackson, McGrath, Sato, and Martin all offered their step-by-step guidance on how courts and nonprofits can experiment and use AI successfully within courts systems.

Step 1: Determine the problem

When starting a generative AI (GenAI) legal assistance project, it is crucial to first pinpoint the specific legal needs and challenges faced by your target audience. McGrath noted that he sees several common examples, including providing public access to legal information, creating internal resources like bench books for judges, and automating court document preparation.

To properly identify the problem, conduct thorough user research to understand pain points related to accessing and applying legal information. For instance, Martin suggests starting by speaking with court staff. “I think we sometimes get caught up in the excitement about wanting to throw AI at the problem and create a solution,鈥� Martin explains. 鈥淎nd there are many use cases, but I think the part that’s really important is to meet with your staff, meet with everyone who’s being impacted by the burden of work, and then determine, based on that, what is the best choice.”

Taking the time upfront to clearly define the problem will help ensure that any AI solution being developed is truly meeting a demonstrated need.

Step 2: Craft a vision

Shifting from problem identification to crafting a vision for the GenAI-powered solution is crucial. The People鈥檚 Law School鈥檚 Beagle+ chatbot illustrated this well. 鈥淏egin with the end in mind,鈥� says Jackson. 鈥淲hen you begin a project, keep in mind what you’re trying to achieve and what success looks like because that’s going to be different for each person.鈥�

Jackson further described how in 2018, the initial vision was to create a chatbot capable of intelligently answering questions about consumer and debt law in British Columbia. Today, while that vision is realized, the ability of GenAI technology to adapt and improve over time necessitates a continuous and evolving vision.

Step 3: Allocate realistic resources

Assessing available resources is crucial before embarking on a GenAI project, with a realistic evaluation considering such factors as existing legal content, technological capabilities, staff expertise and capacity, and budget.

It鈥檚 important to examine the state of the organization鈥檚 existing legal information, including its documents and web pages, to determine the quality and consistency. Indeed, conflicting information across sources often can confuse GenAI models.

For staff capacity, Sato explains how the ACS started with a small team of people, which included the court administrator, the chief technology officer, a webmaster, and two to three staff attorneys, who were necessary for content review, testing, and feedback. It is not uncommon for an initial project to consume about 30% of each team member鈥檚 time.

Technological expertise is also a key consideration in resource assessment. In fact, Martins says this underscores the importance of working with a technology partner that can help navigate the different choices and options available, including the need to understand options for AI model selection, vector databases, and embedding strategies. While some may consider using large language models (LLMs)to reduce costs, the expenses for setup and maintenance often outweigh the benefits compared to using established services like OpenAI.

Financial resources are also a consideration, of course; however, it is worth noting that the cost of OpenAI tokens is often surprisingly low compared to other project expenses. For the creators of Beagle+, for example, using OpenAI鈥檚 tool has cost no more than $75 per month, according to Tangowork鈥檚 McGrath.

Courts can explore the possibilities of AI tools in tackling their specific legal challenges by experimenting within

Addressing common concerns

Our experts say that two common concerns often arise when considering the use of GenAI to solve justice gaps: one is the need for multilingual capabilities; and the second is how to handle AI-generated inaccurate information, or so-called hallucinations.

鈥淎dvanced LLMs like GPT-4 demonstrate impressive multilingual capabilities and are able to understand and respond in numerous languages on-the-fly without requiring additional training or configuration,鈥� explains McGrath. 鈥淢ultilingual support is a key advantage of modern LLMs, enabling chatbots to serve diverse populations with minimal additional development effort.鈥�

However, hallucinations are a significant concern when using LLMs for legal applications. Fortunately, the combination of several advanced strategies can mitigate hallucinations:

• • • First, grounding responses in providing context through techniques like retrieval-augmented generation can help tether outputs to verified source material.
    • Second, careful prompt engineering and relevancy scoring can further constrain responses.
    • And finally, automated checks that compare model outputs to source documents can flag potential hallucinations.

At the same time, manual expert review by humans 鈥� known colloquially as human in the loop 鈥� remains crucial, even with automated safeguards in place. Therefore it is key to periodically sample responses for human verification and focus more intensive review on higher-risk conversations.

Creating a successful AI-powered chatbot for legal information requires careful consideration of the several steps cited above. By following these actions and staying up to date with the latest developments in AI technology, courts and organizations working to close the justice gap can create effective and responsible chatbots that provide valuable legal information to those who need it most.

You can register here for the upcoming NCSC webinar on March 19, which will explore the