Legal skills in the genAI age, Part 2: Should the 1Ls and juniors even try?

TL;DR: Generative AI is doing two opposite things to the younger generation of lawyers at once. It is stripping away the bounded work through which the profession trained, tested, and trusted them, while giving them new tools for analysis, simulation, and self-directed practice. Output can therefore improve before judgment does, and juniors may be asked to exercise judgment earlier than the old pipeline prepared them to. I argue that the accidental apprenticeship must become a deliberate one. Automate the grind whose value lies mainly in output, preserve the friction that forms the lawyer, and use AI to get closer to the clients, decisions, and consequences from which judgment grows. Firms still owe juniors a better training system, but juniors cannot assume one will arrive in time.

Let’s revisit the anxious law student from the beginning of this series. She obviously has some way to go until she has a senior's soft skills and intuitive judgment, so what does the future hold for her and those like her? As we saw in Part 1, her anxiety has a real structural basis. The work that the legal industry has traditionally assigned to juniors is often precisely the kind that generative AI eats up first, because such work is bounded, repeatable, and relatively inexpensive for a senior lawyer to self-produce.

So, should she and the other 1Ls therefore stop grinding through law school, or later stop grinding through the junior ranks?

I do not think so. But the object of that grind, and the way young lawyers learn from it, must change.

The junior job may survive. The apprenticeship may not.

Let’s begin with the apparently reassuring news. Contrary to many grim predictions, the US legal market, one of the markets in which genAI has made its deepest inroads into practice, remains remarkably strong. The class of 2025 posted a record-high 87.7 per cent employment rate in full-time, long-term law jobs (Karen Sloan, "Law Grad Employment Hit Record High Again in 2025, ABA Data Shows," Reuters, April 22, 2026, reporting American Bar Association employment data). But that class was roughly 7 per cent smaller, and the figures predate the full arrival of agentic workflows. More importantly, employment data tell us how many graduates found jobs rather than what is changing inside those jobs. That change, in my opinion, is that the junior position may survive even as the apprenticeship attached to it weakens.

Leaving aside fast-tracking through favouritism or nepotism, the idealised traditional pipeline performed four functions at once. Firms received useful work from juniors, juniors learned by doing it, seniors gained a checkable signal of who was ready for more, and the better performers gradually entered judgment formation, client interactions, and deal or litigation strategy. Proofreading contracts, reviewing documents, checking citations, and researching bounded issues mattered partly because a senior could evaluate them without redoing the whole assignment. The work produced value while buying the junior a chance to progress.

Notice where the most important formation occurred, because it was rarely in the bounded tasks themselves. An associate could bill three thousand hours reviewing documents and finish scarcely closer to becoming a good lawyer. What moved her forward was the access those hours bought. She listened on the client call, watched a senior lawyer quarterback a transaction with seven parties on the line, saw how counsel addressed the bench, or learned how a bad fact was confronted without a flinch. The grind was the ticket into the room. The room was where legal analysis met actual stakes, competing objectives, difficult personalities, and imperfect information.

None of this is a defence of training by osmosis. Neither I nor most of my peers found that ordeal especially enjoyable or effective. Many employers never designed even a half-serious training system. They relied on repetition, proximity, a little help from Fortuna, and the economics of leveraged junior work to produce competent lawyers somewhere along the way. Holmström and Milgrom’s multitask principal–agent model helps explain why. (Holmström & Milgrom, "Multitask Principal–Agent Analyses: Incentive Contracts, Asset Ownership, and Job Design," Journal of Law, Economics, & Organization, 1991) When a role contains several important functions but some are easier to observe and reward, attention concentrates on the measurable ones. Memoranda completed, documents reviewed, and hours billed were visible. The development of judgment, instincts, and client understanding was not. Professional formation limped beside the work that paid for it.

AI changes that arrangement at its economic root. A model can increasingly supply the memorandum, chronology, contract summary, or first-pass review without employing the junior who once produced it. The bots did not create the flaws in the old apprenticeship. They are exposing those flaws while removing much of the economic structure that allowed firms to ignore them. If firms need juniors less for the work that occupied their early years, it becomes less obvious what will bring those juniors into the parts of practice from which they can learn and grow into seniors.

The remaining work begins to split between two demanding ends. At one end sit poorly bounded responsibilities such as understanding what a client really needs, reading a negotiation, managing a legal project, and shaping litigation or deal strategy. At the other sit the workflow-aware tasks involved in evaluating machine output. Both require judgment, but juniors no longer have a clear route for acquiring it. Fewer “check this” and “look through that” assignments arrive. The assignments that remain are increasingly checks on machine output rather than blank-page research, and the layer that once buffered juniors from more consequential work has grown thinner.

The implication is uncomfortable. The old pipeline delivered professional formation as a by-product of employment. As machines absorb more of the work that generated that by-product, juniors must increasingly administer their own development. This sucks. It really does. The burden falls on the people with the least power to redesign the system. Yet the same technology that weakens the apprenticeship can also make a junior appear further along than she is. That is where AI’s gift becomes its trap.

A gift and a trap: when output outruns judgment

A 2025 University of Minnesota preregistered randomised controlled trial on AI-assisted legal drafting captures the duality. The authors called the result a “leveling effect.” AI improved weaker drafts while standardising or flattening stronger ones (Bednar, Cleveland, Erbsen & Schwarcz, "Artificial Intelligence and Human Legal Reasoning," Minnesota Legal Studies Research Paper 2026-21, 2026). Similar patterns appear in studies of law school exams, customer support work, professional writing, and management consulting (Choi & Schwarcz, "AI Assistance in Legal Analysis," Journal of Legal Education, 2025; Brynjolfsson, Li & Raymond, "Generative AI at Work," Quarterly Journal of Economics, 2025; Noy & Zhang, "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence," Science, 2023; Dell’Acqua et al., "Navigating the Jagged Technological Frontier," Organization Science, 2026). Across these settings, the least experienced or lowest-performing participants often gained the most. That is a genuine gift for a junior trying to close the distance between what an assignment demands and what she can yet produce alone.

AI can bridge that output gap on demand. It can provide a polished first draft, an instant synthesis, or a preliminary risk analysis that might once have consumed even a senior lawyer’s afternoon. But it does not thereby fill the experience gap that created the need for help. The junior can cross the output gap before she has crossed the gap in judgment, which lets the work product appear further along than the person producing it.

Picture a junior receiving an email from a partner and understanding only 30 per cent of it. The assignment involves a doctrine she has encountered once and a procedural context she has never encountered at all. Under time pressure and eager to impress, she asks a model to explain the assignment, identify the likely issues, and produce an answer better than anything she could have written unaided. If she tidies the answer and sends it back, she has completed the assignment. But she has not necessarily changed her mental model enough to answer the next one, and she has done little to show why the partner needed her rather than the same tool. The output demonstrates the model’s reach more clearly than the junior’s judgment.

To enjoy the gift without springing the trap, she must treat the machine’s answer as a diagnostic. What do I still not understand? What must I verify? Which view can I defend on my own? Where did the model’s analysis and mine diverge, and why? The machine should accelerate the assignment without replacing the learning. The line between assistance and borrowed competence lies in what the junior does before the work crosses into evaluation or reliance.

Skip that interval too often and the junior may plateau comfortably. Her work keeps improving because the models keep improving, while nothing in the polished output warns her that her own development has stalled. The Minnesota study supplies a sharper caution. Its strongest performers lost points when AI revised their drafts (Bednar et al., 2026). The study also offers narrow reassurance that AI use did not automatically erode short-term comprehension in that setting (Bednar et al., 2026). It does not establish that sustained overreliance is harmless, or that judgment continues to develop merely because immediate comprehension does not measurably decline.

The same levelling effect may appear in the labour market. Once AI skills become as commonplace as MS Office skills, producing a polished first work product becomes less scarce and therefore less differentiating. Each junior who secures a seat may become more capable even as firms need fewer juniors to produce the same volume of bounded work, or demand more from those they hire. None of this requires the legal profession itself to shrink. As discussed in Part 1, cheaper legal analysis may expand demand for legal services through a Jevons paradox. But aggregate demand for legal services is not the same as demand for junior labour, still less a functioning path from junior work to senior expertise. A profession can grow while the work that once nurtured its newest members migrates to machines.

The evidence therefore does not tell juniors to avoid AI. It tells them to distinguish acceleration from formation. Under the old model, the same assignment often produced value for the firm while developing the junior, however chaotically and imperfectly. As that coincidence disappears, the answer is not abstinence. It is deciding which work AI should accelerate and which work the junior must still perform because performing it changes her.

The sorting-hat question: which grind still forms a lawyer?

The firm used to perform much of the sorting for juniors. Not because firms were nice, but because the same work produced revenue, trained the junior, and gave seniors a checkable signal of readiness. As those functions come apart, juniors increasingly need a sorting hat of their own. It is less magical than Hogwarts’s, but it asks the right question: was the value of this grind in the output it produced, or in the person formed by producing it?

The answer does not attach permanently to a type of task. The same assignment may have real formation-value for a novice and mostly output-value once the underlying skill has been learned. The practical question is whether another five, fifty, or five hundred repetitions will still change what you notice, how you reason, or where you exercise judgment. If you largely understand the procedure and the remaining value lies in the completed product, the machine should shoulder more of the load.

A whole class of traditional junior work reaches that point quickly. Figuring out Boolean syntax for a database, reviewing mountains of repetitive material, checking citations at industrial volume, summarising for the summary’s own sake, and adventuring with footnotes at two in the morning can teach the workflow. Their developmental returns nevertheless fall sharply with repetition. Letting those forms of grind disappear is not moral surrender to the bots. It is the same kind of triage that made trips to the library unnecessary once digital legal databases arrived.

Formation-value work is different. Reading cases closely, forming an independent issue map, writing an analysis before asking a model, preparing for client questions, comparing strategies, and making and defending risk calls matter for more than the document produced. The mental labour changes the person performing it. It builds the internal model through which a lawyer learns what is relevant, where uncertainty enters, and why an answer may be technically correct but contextually wrong for the client.

This work often feels unpleasant because it sits inside the zone of productive difficulty: hard enough to demand real thought, but not so hard that the learner drowns. AI’s frictionless ease can pull a task out of that zone before the work has done its formative job. Remaining inside it is now an active choice rather than a condition imposed by the assignment.

Some institutions have already begun making that choice deliberately. At least one American law school requires AI in certain assignments while keeping foundational first-year exercises AI-free (University of Chicago Law School, "AI Advances into the Law School Curriculum," 2025). There is no contradiction. Some exercises teach students to use the machine. Others build the unaided foundations required to supervise it. The rule follows the skill being trained.

The sorting hat therefore does not tell juniors to avoid AI. It tells them where AI should remove low-value friction and where human effort must remain. Once we know which friction to preserve, the next question is whether the machine can help us practise that work better rather than merely escape it.

AI can help rebuild the apprenticeship it is disrupting

Whether you came to law loving technology or precisely to escape it, engagement is no longer optional. Clients increasingly arrive with machine-generated answers, and senior lawyers increasingly expect machine-assisted efficiency. No one will poll juniors for their preferences first. There is no abstainer track.

The price of admission is nevertheless lower than it was during earlier waves of legal technology. The prerequisites are not coding or a deep technical background, but literacy, openness, and a problem-solver’s mindset. Think of the standard as the racing driver’s rather than the mechanic’s. You need not build the engine, but you must understand the machine well enough to recognise unsafe behaviour at speed. Coding remains optional, though I personally believe that learning a little pays off because it makes data and software less mysterious.

Operating the tool is only the baseline. The more interesting opportunity is to use the same technology destabilising the old apprenticeship to make practice more deliberate. Instead of asking a model to tell you what you need to know, ask it to interview you until it finds where your understanding falters. Have it role-play a sceptical partner, impatient client, aggressive opposing counsel, or critical judge, then practise defending the analysis under pressure. Ask it to generate a synthetic shareholders’ agreement with planted defects, then spot, explain, and negotiate them. Turn a difficult doctrine into a visual map, an audio lesson for a run, or a drill built around the mistakes you repeatedly make.

Another useful exercise begins with something you already know well. Write down, without assistance, the most consequential things experience has taught you about a clause, doctrine, or issue. Then ask the model the same question in several fresh conversations. The machine’s answers will often converge on competent, statistically central points. Yours should carry the marks of the files you have run, the client who surprised you, and the argument that failed before a particular judge. The comparison reveals which parts of your knowledge the machine already sells for pennies and which parts remain grounded in experience.

The model does not become the source of truth in any of these exercises. Synthetic practice cannot substitute for a live client, a real file, or the judgment of a good senior lawyer. It must remain anchored in actual authorities, known facts, and human feedback. Used carefully, however, AI can create more repetitions and personalise them around the learner’s actual weaknesses instead of merely completing the repetition for her.

My experience in music makes this intuitive. When I was pursuing music professionally, playing a set piece from beginning to end over and over produced little improvement. It often reinforced the bad habits formed during unfocused runs. Progress came from isolating the weak passage, slowing it down, varying the exercise, and drilling what was going wrong before returning to the whole. Legal development should work the same way. The point is not to accumulate hours, but to identify the skill that needs improvement and tighten the feedback loop around it.

That possibility becomes concrete through three loops in the life of an AI-assisted assignment.

The new junior grind: three ways to grow with AI when doing a task

The old pipeline bundled together the formation of judgment, the demonstration of that judgment, and progression into consequential work. The new junior grind must recover those functions deliberately. The three practices therefore sit at different moments: before the machine answers, before the work crosses into evaluation or reliance, and after the work meets the world.

None is as attractive as sitting on a patio with good pals after a gruelling week of court motions or data room preparation, and I am not suggesting that juniors turn every spare evening into recursive self-improvement. Nor does any of this absolve firms or law schools from building a better training system. But until that system arrives, juniors should act as though nobody will administer their legal formation for them. There are no supreme saviours.

Before the machine answers: form your own opinion

For work with formation-value, do not ask the model for its reasoning until you have formed at least a provisional view of your own. Map the issues, identify the important facts and principal risks, and write down your preliminary answer, your confidence in it, and what fact or authority would change your mind. Do this before the bot’s polished output overwrites your memory of what you originally thought.

Writing the view down matters even when it will never leave your desk. Writing does not merely report a finished thought. It is one of the ways thought becomes precise. The argument containing a hole, the sentence that wrestles with you rather than saying what you mean, and the discomfort of not yet having a clean answer are often where judgment forms.

Then compare the model’s reasoning with yours. Which issue did it see that you missed? Which fact did it overvalue? Which authority changed your view? Which conclusion still feels wrong, and why? Trace the divergence to the authority, fact, assumption, framing choice, or inferential step that sent the analyses in different directions. Ask the model to press your reasoning and defend its assumptions, but do not let it appoint itself judge of the disagreement. Test both positions against the authorities, facts, and, where available, senior feedback.

For recurring or consequential work, keep a divergence log recording where you and the AI differed, which view survived scrutiny, and why. It is not a scoreboard for proving that you were cleverer than the machine. Like a wrong-answer journal for the LSAT, it reveals recurring blind spots and tells you what you need to learn. Firms can reinforce formation by grading this loop rather than the polished output alone (University of Chicago Law School, "Rethinking Legal Education in the AI Era," 2026). Ask for the junior’s pre-machine view alongside the final work, then ask what changed her mind, which machine errors she caught, and what remains unresolved.

Before the work crosses a reliance boundary: make your work inspectable

The divergence log is inward-facing. This second practice is its external-facing counterpart, giving anyone reviewing or relying on the work a reasoned basis for doing so. Programmers routinely document scripts, modules, and projects so that others can understand what the code does. Lawyers, junior or senior, should adopt a similar practice for machine-assisted work because polished prose conceals which propositions were checked, which facts were assumed, where uncertainty entered, and where the lawyer rather than the model made the decisive call.

The depth of the record should rise with the stakes, irreversibility, and intended use of the work. An exploratory draft shared to expose uncertainty is different from a review-ready analysis a senior will use to evaluate the junior or build the next draft. Both differ from a court filing or advice on which a client will immediately act. Before the work crosses from exploration into evaluation or reliance, identify the authorities and facts carrying the conclusion and what you verified. State what remains unresolved, the strongest contrary argument, what would change your view, and which point still requires senior judgment or escalation.

The rule is not that every unfinished thought must be defensible before it is shared. Once machine-assisted work crosses a reliance boundary, into performance evaluation, substantive review, a downstream decision, or client, court, or counterparty use, the junior should be able to explain and defend the basis on which it rests. Work shared expressly for iteration can remain provisional, but its status, unresolved points, and the feedback being sought must be clear.

“Defend” does not mean claiming certainty or manually reproducing everything the machine accelerated. It means pointing to the key authorities and facts, explaining the reasoning connecting them to the conclusion, disclosing the material assumptions and uncertainties, and saying what would change your view. If the intended status is unclear, ask whether the senior wants an exploratory view, a review-ready analysis, or something capable of supporting downstream action. For consequential work, a short reliance note can identify the sources checked, material assumptions, unresolved issues, limits of confidence, and points requiring escalation.

After the work meets the world: follow you work through

Most junior assignments are fragments of a larger matter. We research one issue, draft one clause, or prepare one memorandum, then hand it off before seeing the client decision, negotiation, or consequence that gives the work its meaning. AI may deepen that fragmentation by letting juniors finish their assigned slice faster and move immediately to the next one.

Use some of the time and credibility AI creates to move further along the life of the matter. Ask what decision the work supports and what happens next. Offer to present part of the analysis on the client call, own the post-meeting follow-up, or revise the recommendation after the negotiation. Then ask what happened after the memorandum went out. Did the client follow the recommendation? Which risk mattered? Which apparent red flag proved irrelevant, and why?

Access to this loop is not entirely within the junior’s control. This is why champions and sponsors matter even more as advice becomes abundant. A mentor or model can explain what good judgment looks like at two in the morning. Neither can place you on a consequential matter or persuade a partner that you are ready for more. Access and advocacy remain scarce.

Once you gain that access, seek feedback on the real-world effect of the legal strategy rather than only the doctrinal elegance of the work. A termination-for-convenience clause may look like an obvious red flag, but its significance depends on the deal, the parties’ bargaining positions, available alternatives, timing, relationship, and the counterparty’s objectives. AI can help rehearse that reasoning. It cannot supply the stakes or tell you which considerations moved the people involved. Judgment grows when you compare what you predicted with what happened and understand the difference.

This is probably be the hardest of the three-pronged new grind set to implement (definitely is for me personally), precisely because juniors were already told to do some version of it before AI arrived. We should be curious, ask to sit in on the call, follow up on what happened, seek more responsibility, etc. Whether that advice is realistic depends partly on personality, partly on workload, and partly on the culture of the firm and the generosity of the senior lawyers around you. Some juniors will be welcomed into the next stage of the matter, whereas others will have to ask repeatedly, and some will simply have fewer opportunities available to them. Such is the unfortunate price one must pay for having to interact with humans at workplaces.

But we must remember that AI makes the stakes of trying higher, not lower. If the technology compresses more junior work into discrete research and drafting tasks, then the learning that once came incidentally from spending longer on a matter must increasingly be pursued deliberately. So, it becomes ever more critical to understand how your piece of work entered someone else’s judgment, what changed once it encountered commercial reality, and what you would do differently next time.

Different ramps to the same destination for those who want to do law as law

After rebuilding the learning loops, using my suggestions or not, a junior will eventually ask what's next. The more forward-looking legal community, including the LegalQuants, increasingly speaks of a fork between a technical path and a more traditional lawyering path. I have no prescience about which roles will proliferate or what titles will survive, but I am confident enough to proclaim that for those who still want to practise law as lawyers, the destination has not changed.

That destination is understanding what the client actually needs, deciding what a satisfactory answer must accomplish, translating a messy problem into work, exercising judgment over the result, and putting one’s name to the advice with accountability. Some juniors will discover that designing legal systems is the work they prefer, and they may and should explore and pursue careers in legal engineering and systems architecture. These will increasingly become more real and valuable pursuits. Nonetheless, I am of the view that when the role no longer involves advising a client or answering for a legal result, it is aiming at a different professional destination. I by no means am stating this as some kind of demotion. Instead, I just want to make sure that a junior choosing between those different destinations should be do so deliberately rather than by passively drifting into one or another.

With a clear destination in mind, technical fluency can be a powerful ramp towards consequential legal work performed by lawyers. It can help a junior learn faster, understand how clients use AI, translate between lawyers and technologists, and contribute to the redesign of real workflows. But ownership of the workflow is not the prize. The prize is the access, legal insight, responsibility, and client or partner trust that the work can earn.

The profession has seen the danger before. When e-discovery arrived, some junior lawyers became excellent at the platforms and searches only to find that their expertise kept them adjacent to the case rather than moving them towards responsibility for it. The genAI equivalent is the associate everyone calls when a prompt fails or an agent goes crazy, but who is still absent when the client calls. Technical work has become a destination by accident rather than a ramp.

A non-technical junior can reach the same professional destination through legal depth, relationships, problem definition, sponsorship, and ownership of outcomes, so neither ramp is inherently superior for that purpose. Hence, what matters is whether the chosen ramp leads faster towards consequential legal work where judgment and accountability grow. The three loops I outlined above can be seen as my attempt of laying out a common foundation for building those ramps, meaning the specific one by which you convert them into more trust, opportunities, and responsibility may differ.

Three asymmetries that favour juniors

Juniors will always begin far behind experienced lawyers in judgment, client understanding, and professional instinct. No AI will change that unless it can somehow rewire our neural circuits and accelerate the biological rate at which the brain converts experience into durable intuition. Thus, AI does not erase the seniority disadvantage. But because juniors and seniors enter the new era of genAI from different vantage points, I see three narrower asymmetries that juniors can exploit.

Juniors have less to unlearn

Senior lawyers may have decades of settled habits and workflows to reconsider, while juniors are forming theirs right here and now. This is good news because genAI can be learned and wielded in ordinary language rather than code or database syntax. Some evidence even suggests that less-experienced workers obtain larger immediate gains from it (Cui, Demirer, Jaffe et al., "The Effects of Generative AI on High-Skilled Work," Management Science, 2026). Juniors can therefore build habits with AI already inside the picture, giving them the unique opportunity to be the first wave of legal workers who can form the habit of coming up with an independent view before prompting, verify before relying on the output, and escalate when the context exceeds what the system can safely handle.

This is of course only an opportunity to be seized, not an inheritance to be received. Plenty of senior lawyers will experiment more seriously and use AI better than their juniors. Being young does not confer discernment by biological right. The advantage belongs to juniors who use their blanker slate to develop dependable habits before bad ones harden.

The same productivity gain can buy a junior more trust

A senior lawyer who becomes faster remains a senior lawyer with more or less established credibility and consequential files. For a junior, a better memorandum or preliminary analysis can change what others are willing to let her do. In the customer-support study, workers with two months of experience who used AI performed as well as untreated workers with more than six (Brynjolfsson, Li & Raymond, "Generative AI at Work," Quarterly Journal of Economics, 2025). They did not acquire four months of judgment, but they demonstrated useful performance earlier.

This distinction matters because the productivity gain becomes valuable when the junior executes the "follow your work through" grind and converts it into access to the next stage of the matter. So, do not spend every hour AI saves accepting a larger pile of the bounded work it made cheaper. Please use some of the credibility it creates to seek better files, client exposure, and responsibility. In a profession where experience and reputation are almost everything, to get these two assets compounding as early as possible can alter the trajectory of one's career.

New fields narrow the seniority gap

While a partner may possess twenty years of legal knowledge and judgment in M&A of energy companies and B2B tech startups, but the profession has not had twenty years to establish how agentic legal workflows should be supervised, where human judgment must re-enter, or how responsibility should be allocated when a system fails. The same is true of legal questions involving agent identity, synthetic evidence, autonomous physical systems, and machine-mediated organisational decisions. The underlying doctrines may be familiar, but their application is being worked out as you read this article.

Experience compounds from the point at which relevant repetitions begin. In a mature field, a junior often enters decades or even hundreds of years behind. In an emerging one, she may be in the game only months behind. Senior lawyers still begin with stronger legal instincts, but the gap in directly relevant experience is narrower. Juniors who enter early can accumulate specialised judgment while conventions remain unsettled and help shape the standards by which the field develops.

Emerging practice does not eliminate hierarchy. It moves the starting lines closer together. None of these openings changes the ultimate professional destination: the lawyer must retain enough judgment and authority to refuse the machine.

Knowing when to tell the machine no

AI remains an amplifier rather than a substitute. An amplifier magnifies whatever it receives, both signal and noise. Lawyers who feed it well-formed judgment can obtain judgment at scale. Lawyers who feed it vague prompts and disorganised context will obtain slop faster than ever. Clear systems thinking matters more, not less, because the lawyer must work with the machine without becoming a biological conduit through which it reasons upon the world.

In 2001: A Space Odyssey, HAL 9000 refused the human: “I’m sorry, Dave. I’m afraid I cannot do that.” The junior doing the new grind must prepare for the opposite scene. Clients will increasingly arrive holding confident machine-generated answers of uneven quality, and the lawyer’s own tools will place dazzling first drafts in her hands. The point of the deliberate apprenticeship is to become the lawyer who recognises the moment to answer: “I’m sorry, AI. I’m afraid we cannot do that.”

That is why the anxious 1L should still grind. She is not competing with the machine to read the most pages or produce the fastest summary. She is developing the judgment to decide which work the machine should perform, which conclusions another person may safely rely on, and when the technically available answer is wrong for the client.

This leads to the operational question I take up in Part 3: what separates basic AI competency from the deeper AI literacy lawyers need to understand these systems well enough to exercise better professional judgment in a world with increasing amount of machine-mediated work?

Stay tuned.