Josephine Petrick, Chief of AI Products at BriefCatch – Interview Series

Josephine Petrick as Chief of AI Products at BriefCatch brings a rare combination of appellate litigation experience, AI product judgment, and hands-on legal drafting expertise to one of the legal industry’s most specialized writing platforms. In her role, she leads AI product strategy for BriefCatch, guiding the development of generative and retrieval-based tools designed to improve how attorneys and judges draft, edit, and refine legal documents. Her background includes serving as a partner at The Norton Law Firm, senior counsel and appellate litigator at Hanson Bridgett, and law clerk to Judge James L. Dennis of the U.S. Court of Appeals for the Fifth Circuit. That blend of courtroom, appellate, and advisory experience gives her a practitioner’s perspective on building AI tools that address real legal-writing challenges rather than abstract productivity problems.

BriefCatch is an AI-powered legal writing and editing platform built specifically for lawyers, judges, law firms, courts, and government agencies. Founded in 2017 and built around the legal-writing expertise of Ross Guberman, the platform helps legal professionals improve clarity, precision, persuasion, and credibility directly within their writing workflow. BriefCatch says it supports more than 300 law firms, including 45+ Am Law 200 firms, along with 80+ courts, and its newer BriefCatch Next product combines expert legal-writing guidance with optional AI tools intended to enhance—not replace—legal judgment.

You recently joined BriefCatch as Chief of AI Products after serving as an appellate attorney, certified appellate specialist, and builder of legal technology products. How has your experience at the intersection of appellate law, legal strategy, and technology shaped your vision for AI in the legal profession?

After majoring in linguistics at Cal, I was fascinated by AI and always wished I had gone into the field. I was drawn to its potential to expand human intelligence and help solve practical problems at scale. I became a lawyer for similar reasons. My mother was a constitutional lawyer, and from her I learned that the law could be used to help people efficiently, on a large scale. By the time I was practicing, I assumed the path to a career in AI had closed. So it is a little ironic, and deeply satisfying, that my legal career has brought me back to the questions that interested me in the first place: How do we use technology to make expertise more available, more useful, and more humane?

I’ve been experimenting with large language models since before they were packaged as chatbots, back when OpenAI offered its product as a sandbox. From the beginning, I could see the potential for practical problem-solving. Could this help with record review? Could it make local-rules research less painful? Could it help track pending California Supreme Court cases, estimate how long appeals will take, or streamline post-trial motion practice? Could it help someone get unstuck?

One of my favorite examples of the last question comes from my father, a retired number theorist who still works on open problems. I persuaded him to try a pro LLM account, and he has used it to prove theorems and make progress on problems he had been thinking about for years. One of the reasons LLMs are so effective at solving open math issues, like the Erdős problems, is that they have training in many areas of mathematics, and can transfer those techniques from one domain to solve problems in another (for example, using techniques from algebraic number theory to solve a problem in elementary geometry). That experience captured something important for me: AI can extend our creative reach, help us draw on a broader body of human knowledge and expertise than we can hold in our heads individually, and use that to make progress in human knowledge and endeavors that would otherwise have taken much longer.

I saw the same pattern in legal practice. As an appellate lawyer, the same procedural and strategic questions came up again and again. So I built databases, trackers, infographics, client resources, appeal-duration calculators, post-trial timing tools, and eventually helped build a legal rules database. The throughline was always the same: the knowledge existed, but it was trapped in PDFs, books, seminars, internal notes, or some expert’s head. I became obsessed with a problem AI is unusually well-suited to solve: how to make knowledge and expertise available at the moment people need it.

That shaped my vision for legal AI. Most legal tech products promise that AI will automate drudgery so lawyers can focus on strategy. That is true and valuable, but it’s only the beginning. What I’m most interested in now is activating expertise. A lawyer may have read the right book, attended the right CLE, saved the right post, or learned the right lesson from practice, but that knowledge is often dormant when the lawyer is actually drafting, researching, or advising a client.

BriefCatch sits directly in that opportunity. Legal writing is not just polish. It is where legal judgment, strategy, and persuasion become usable to another mind. I’m excited about a future when AI helps lawyers retrieve and apply the right rule, technique, strategy, warning, or strategic insight in the moment, while preserving professional judgment and accountability. I think of it as an expertise exoskeleton: technology that helps lawyers carry more knowledge, see more clearly, avoid preventable errors, and apply their best judgment when it matters most. If we build these systems well, society benefits because legal expertise becomes less trapped in expensive one-to-one channels and more available to the people, businesses, courts, and institutions that depend on it.

BriefCatch recently introduced RealityCheck, a tool designed to identify hallucinated or unsupported legal citations before filing. What did you see as the biggest gaps in existing legal AI tools that made this capability necessary, and how significant is the citation reliability problem today?

The biggest gap is that too many tools are still optimized around generation rather than verification. They can help produce text quickly, which looks impressive, but legal work does not end with plausible text. It doesn’t even start there. Our legal system fundamentally depends on judges and lawyers faithfully recounting and applying authority. Any legal tool must start with authorities that actually exist, must accurately describe what those authorities say, how they have been applied, and whether the proposition for which they’re being invoked is supported considering the context in which the authority was decided. Anything less than that is legal fanfiction. There are less polite words for it than that, but I’ll leave that to the judges who have dealt with it in their courtrooms.

That is why a tool like RealityCheck is so important. The problem is not limited to fully fabricated cases, although those are obviously alarming. There are subtler and more common failure modes: a real case cited for the wrong proposition, a holding overstated, a quote slightly altered, a parenthetical that smooths over an inconvenient distinction, or a citation that looks facially proper but does not actually support the sentence it follows.

These are not new problems in the practice of law, but they are now amplified. Lawyers have always been responsible for checking their work, but the scale and speed of AI-assisted drafting make verification even more important. The question is not just “Did AI hallucinate?” It is “Does this filed document contain legal assertions that are accurate, supported, and professionally defensible?”

I also think we need to be careful not to treat human-only work as the gold standard. Humans misquote, overstate, and miss things too, especially under time pressure. Regardless of the source of the text, the best practice is layered review: human judgment, source-grounded tools, deterministic checks where appropriate, and verification workflows that make errors easier to catch before they become an order to show cause.

Many lawyers remain skeptical of generative AI because of concerns around accuracy, ethics, and professional responsibility. What practical steps can legal technology companies take to build trust and encourage responsible adoption?

Trustworthy products should make their boundaries visible. They should show sources when sources matter. They should make it clear when a system is generating language versus verifying a citation or applying a deterministic rule. They should also be candid about limitations, because overclaiming is one of the fastest ways to lose lawyer trust.

Responsible adoption has to be built into the workflow. Lawyers should not have to become prompt engineers or AI risk officers to use these tools well. Good products should guide users toward responsible use by design: preserving confidentiality, supporting supervision, enabling review, retaining audit trails where appropriate, and making verification easy.

You’ve worked extensively with retrieval-augmented generation (RAG) and other techniques designed to improve the reliability of AI systems. Where do you believe current legal AI products still fall short, and what technical advances are needed to close those gaps?

Legal AI products are already impressive at drafting, summarizing, brainstorming, and retrieving. Some still fall short on an important metric: the exacting accuracy that legal practice demands. (Our justice system depends on it!)

I’d think about the reliability gap as a systems problem, not a single-model problem. Better models will help, but legal-grade reliability comes from several layers working together: retrieval and ranking, verification loops, memory, workflow design, and hybrid deterministic/LLM architecture.

The basic mechanics of retrieval are fairly well understood, but it takes a lot of testing and iteration to get recall and ranking right. Adversarial agents and agentic loops are becoming more commonplace, but the design challenge is making those checks principled rather than creating expensive, latency-heavy “double-check sprawl.” With LLM memory, the technical challenges include where the memory belongs (general memory versus a specific workflow; memory systems can feel pretty bolted on) and the form it takes to ensure the agent accesses it when appropriate. Workflow design still matters; if the system is given a single long, vague prompt, it will produce incomplete or generic results. Legal AI needs more structured workflows, breaking the task down into deliberate steps to ensure it has the context it needs and goes through all of the steps required to generate accurate, excellent legal work product. But this may become less of a problem as we develop bigger foundation models and improve compression.

Another lever to pull is the mix between deterministic rules like regexes and probabilistic models. LLMs are incredibly useful in the prototype phase because they help you discover the shape of the problem. But once you have a good working prototype and understand the pattern, you may want to peel parts of the workflow back into deterministic logic. And although LLMs can be inconsistent, they are excellent at supplementing what deterministic rules are not flexible enough to catch, and at utilizing deterministic rules/code themselves. The strongest systems won’t be purely deterministic or purely LLM-based. They’ll use deterministic rules where the problem is understood, and LLMs where ambiguity, judgment, or context-sensitive language understanding is needed. The question is where the boundary sits: where do rules end, and where should the model take over?

When you start layering these techniques, you can make huge improvements in the reliability gap, but the tradeoff is engineering complexity, token cost, and latency. I expect these will improve with technical advances and as memory and compression become more effective and readily built into foundation models.

Legal writing often requires nuance, persuasion, and careful interpretation of precedent. Which aspects of legal writing do you believe AI can genuinely improve today, and which elements still require uniquely human judgment?

AI can already improve a lot of legal writing, especially when the task is well-defined. It can help identify unclear sentences, buried points, inconsistency, redundancy, weak organization, unsupported factual assertions, mismatched defined terms, citation problems, and places where the draft does not guide the reader well. It can also help lawyers brainstorm creatively, test arguments, and apply known writing and persuasion principles at the moment of drafting instead of relying on memory alone.

That last piece is especially exciting to me. A lawyer may know, in theory, that a brief needs a stronger introduction, a cleaner rule statement, or a more reader-friendly statement of facts. But when you are drafting under pressure, knowing and applying are not the same thing. AI can help bring expert writing guidance into the workflow when it can actually improve the document.

But good legal writing is not just craft or text optimization. It is also the ability to identify the most salient, interesting, and case-specific details, and then turn them into a theory that makes the reader understand why this dispute matters. In my experience, that is still where human lawyers are much stronger than LLMs. Maybe it is taste, maybe it is experience, and maybe it is something deeper about how these systems work. LLMs are trained to generate likely continuations of text, which creates a gravitational pull toward statistically ordinary-sounding text. In many legal contexts, that is useful. You often do not want legal writing to call attention to itself. You want the ideas, facts, and authorities to shine through.

But truly persuasive legal writing often depends on finding what is different. What is the detail that makes the case come alive? What fact changes the emotional valence of the dispute? What legal point will make a judge sit up and recognize the problem? As an appellate lawyer, that is often the thread you are trying to find. You are looking for the version of the case that is unique and compelling enough to make the tribunal care.

That is the part I would not want AI to bland-bias away. A system can suggest clearer phrasing, flag weaknesses, pressure-test citations, or remind a lawyer of persuasive techniques. But the lawyer still has to decide what matters, what to concede, how aggressive to be, what not to say, how to frame uncertainty, and how to explain the emotional and ethical heart of the case.

I am cautious about saying “AI will never be able to do this,” because those predictions have been wrong many times. AI already does many discrete tasks faster and more accurately than humans, and I don’t think it makes sense to feel threatened by that. It is a tool. But at least today, the most important parts of legal persuasion still require human judgment: selecting the right details, understanding why they matter, and building a story that is emotionally resonant.

As AI becomes more deeply integrated into legal workflows, how do you see the role of associates and junior lawyers evolving? Could AI fundamentally change how the next generation of attorneys develops legal writing and research skills?

Yes, and we should be intentional about it. Junior lawyers have traditionally learned by doing difficult work slowly: researching, drafting, revising, getting comments, and gradually developing judgment. AI will change that path, but it does not have to weaken it. Used well, it can make training more explicit and more effective.

The risk is that junior lawyers skip the struggle that builds judgment. If a tool produces a plausible draft too quickly, a new lawyer may not learn how the argument was built, where the weak points are, or why one framing is better than another. That would be a real loss.

But the opportunity is enormous. AI can function as a tutor, a simulator, a writing coach, and a source of immediate feedback. It can show alternative structures, identify unsupported claims, explain why a citation may not support a proposition, or compare a draft against principles of effective legal writing. It can also give junior lawyers more reps. Getting our reps in and getting feedback is how we learn and grow.

The firms and legal departments that do this best will not just give junior lawyers AI and hope for the best. They will redesign training around it. They will ask juniors to explain their choices, verify outputs, compare alternatives, and learn to supervise AI-assisted work the way they will eventually supervise human work.

You have advised clients across industries ranging from cryptocurrency and consumer technology to complex commercial litigation. Are there particular sectors where you believe legal AI adoption will accelerate fastest, and why?

I believe legal AI adoption will move fastest when

1. The organization has workflows that work well with LLMs,
2. The organization is open to innovation, and
3. The organization is motivated to make a change.

On the first point, legal is a natural domain for LLMs because law is, in many ways, human behavior and human values encoded in language. Legal work turns messy facts into categories, rules, obligations, arguments, risks, and consequences. That is exactly the terrain where language models can be powerful, provided they are grounded, verifiable, and designed around how lawyers actually work. So I think most organizations in legal are good candidates except, perhaps, those whose primary service model is human-to-human interaction. But even for organizations like pro bono clinics, LLMs can help with things like triage and knowledge management so I wouldn’t rule them out.

The second factor, openness to innovation, may be more of a gate. That is why I think in-house legal teams are especially well positioned. They sit inside organizations that are already trying to improve speed, reduce cost, and operationalize knowledge. They manage contracts, disputes, policies, regulatory obligations, customer issues, employment questions, product counseling, and risk assessments at scale. Those workflows are often repeatable enough to support AI adoption, while still requiring enough legal judgment that better tools can make a meaningful difference.

I also think litigation, appellate practice, regulated industries, fintech and cryptocurrency, consumer technology, insurance, and healthcare will be important areas. In litigation and appellate work, for example, AI can assist with record review, issue tracking, citation checking, procedural rules, preservation analysis, argument mapping, and drafting feedback. The stakes are high, so the tools need to be serious. But that is not a reason to avoid adoption. It is a reason to build better products.

The third point—motivation to change—is also a significant gating factor still. The way I see it, the legal industry overall should be motivated to change. The justice system is slow, expensive, and inaccessible to far too many people. The harder question is whether the profession is open to changing. I think it increasingly is. Some parts of the industry have been slow to move because the existing model was profitable and familiar. But lawyers are also problem-solvers. Many of us are justice-minded, reform-minded, and stubborn in the useful sense. We are trained to argue for change when the existing answer is wrong.

The private sector may move first because it is more nimble and better resourced. Government, legal aid, and pro bono organizations face different constraints and often have to focus on the urgent work directly in front of them. But those sectors should not sell themselves short. If AI can help make legal services more efficient, more accessible, and more affordable, the long-term opportunity is not limited to any one sector. It is relevant to everyone who cares about making the legal system work better.

Courts, regulators, and bar associations are increasingly scrutinizing the use of AI in legal practice. What regulatory or ethical frameworks do you expect to emerge over the next few years, and how should legal technology providers prepare?

Our existing ethical framework can already handle most of these concerns. Attorneys were always required to ensure that everything they represent to a court is grounded in the truth. They have to be candid when a case goes against them and transparent when they make mistakes. They have to handle their cases competently and supervise junior attorneys and nonlawyers who help them. They have to keep their clients’ sensitive information confidential.

The first wave of AI ethics does not require an entirely new moral operating system. We need to apply these familiar duties with new seriousness. The lawyer cannot outsource judgment and then disclaim responsibility. That part should not be controversial.

One thing we should start doing more of is technology-specific training so attorneys don’t get blindsided when they don’t adequately supervise or check AI output. California, for example, now requires technology competence as part of continuing legal education, and I think that’s the right direction. But one hour every few years is not enough. Law firms, legal departments, courts, and legal services organizations should be developing their own internal AI policies and mandatory training: not just “don’t cite hallucinated sources,” but when to use these tools, when not to use them, how to verify outputs, how to protect confidential information, and how to think about the allocation of responsibility between human and machine.

For legal technology providers, ethics should not be treated as a warning label pasted onto the product at the end. It should be product architecture. The best tools will make responsible use easier by design: source transparency, confidentiality controls, and workflows that facilitate human review.

So none of the problems we’re currently seeing are what I would call “new.” But there is a truly unprecedented legal ethics question on the horizon. Right now, almost everyone agrees that AI should assist lawyers and judges, not replace them. I agree with that instinct, given where the technology is today. But what happens if AI systems become faster, less expensive, more consistent, more objectively reliable, and less biased than humans at legal reasoning tasks? Will we still say that a human must always remain in the loop?

That is a strange and uncomfortable question, but I think we need to be honest enough to ask it. I do not think we are close to replacing judges or lawyers in high-stakes human disputes, and I would be deeply skeptical of any system that claimed otherwise today. But this is already happening in discrete settings: AI is already resolving certain disputes with a major ADR provider and on X, Grok frequently settles fact-checking disputes.

We usually assume human judgment is the ethical backstop. Often it is. Humans bring moral responsibility, institutional legitimacy, empathy, practical wisdom, and an understanding of context that machines do not yet possess. But humans also bring delay, cost, inconsistency, fatigue, bias, gatekeeping, and unequal access. If AI-assisted resolution becomes more accurate, more affordable, and more accessible than the status quo, the ethical question should not be answered by reflex. The question should be, “What process produces the fairest, most accurate, most transparent, and most accountable result?”

If automated or AI-assisted systems eventually handle more routine disputes, human review may become most important at the edges: the unusual cases, the uncomfortable failure modes, and the places where rigid consistency starts to look less like justice and more like old-fashioned legal formalism.

That is how I think legal technology providers should prepare. Build for the current ethical rules, but also for a future in which the profession has to compare human and machine performance honestly. The goal should not be to preserve every existing ritual of legal practice. The goal should be to preserve what those rituals are supposed to protect: fairness, accuracy, accountability, legitimacy, and access to justice.

BriefCatch has built its reputation around improving legal writing rather than simply automating legal work. How do you balance productivity gains from AI with the need to preserve critical thinking, legal reasoning, and professional accountability?

I do not think productivity and critical thinking have to be in tension. In fact, when a lawyer is more efficient at routine tasks, it creates more room for creative, critical thinking. So the question is what kind of productivity we are creating. If AI just helps lawyers generate more words faster, that is not necessarily progress. Legal work already has plenty of words.

The better goal is to help lawyers produce clearer, more accurate, more persuasive, and more defensible work. That is why BriefCatch’s focus on writing matters. Legal writing is not the decorative layer on top of legal analysis. It is where the analysis becomes precise, testable, and useful to the reader.

The way to preserve accountability is to design AI as an expertise layer, not an authority substitute. The tool can suggest, flag, verify, compare, and pressure-test. It can help apply writing principles, identify weaknesses, and catch problems. But the lawyer has to make the final decision and remain accountable for the work. To paraphrase my favorite Peloton instructor, “The AI makes suggestions. You make decisions.”

My own view is that AI is most valuable when it makes the lawyer more engaged. A good system should sharpen the lawyer’s attention. It should ask the question the lawyer might have missed, surface the authority that needs checking, or suggest a clearer way to express the point. It should not lull the user into thinking plausible text is the same as finished work.

Looking ahead five years, what does the ideal relationship between lawyers and AI look like? Do you envision AI primarily serving as an assistant, a collaborator, or something closer to a trusted legal co-pilot?

I prefer “expertise exoskeleton.” The ideal legal AI system should help lawyers and judges do what they are already responsible for doing, but with better reach, better recall, better consistency, better creativity, and better verification tools.

Five years from now, I hope lawyers will have AI systems that are deeply embedded in their workflows and much more specialized than today’s general-purpose chatbots. A lawyer drafting a brief should be able to draw on writing expertise, case law, record materials, procedural rules, citation checks, internal knowledge, and strategic feedback without leaving the workflow. But the system should also be transparent about what it knows, what it checked, and where the lawyer needs to exercise judgment.

The ideal relationship is not delegated understanding or judgment. It is calibrated trust. Lawyers should know which tasks the system is excellent at, which tasks require supervision, and which tasks belong squarely to human judgment.

If we get this right, AI will make expertise more available. It will help lawyers write better, reason more carefully, avoid preventable errors, and spend more time on the parts of the work that require intuition and judgment. I find the idea of a legion of “Mecha Lawyers” to be much more interesting and optimistic than simple automation.

Thank you for the great interview, readers who wish to learn more should visit BriefCatch.