Where Do Junior Lawyers Come From Now?
The apprenticeship that produced every lawyer alive was never designed. It was a byproduct of cheap junior labor doing necessary tedious work, and automation is collapsing the trade that funded it. A note from inside the room where the bottom rung is being sawn off.
In my second year of law school, I spent a long afternoon in the clinic office reading the same petition four times in a row. It was a guardianship case, and I was the student attorney on it. The petition had been drafted by someone else in a prior semester and I had inherited the file. The first read, I understood it. The second read, I noticed the language was different from what I had been taught in my civil procedure course. The third read, I realized the difference was not a mistake. The fourth read, I finally saw why the drafter had chosen what they had chosen. The judge in the case had a particular preference for a particular kind of phrasing, and the drafter had quietly accommodated that without ever saying so, and the case had moved more smoothly as a result.
Nobody taught me that. The supervising attorney did not sit me down and explain it. I learned it by reading the same document four times because I had time to read it four times because I was a clinical student and that is what clinical students were given to do. The work was slow. The work was repetitive. The work was, by any sane measure, beneath the time of a fully licensed attorney. And the work was the school.
I want to write about what happens when that kind of work goes away. Not in the abstract. The kind of work I am describing has, in many corners of the legal profession, already gone away. It is being done by software right now, in firms across the country, in the federal agencies I work with, in legal departments inside Fortune 500 companies. I help deploy some of that software. So I am not writing about this from the outside. I am writing about it from inside the change, as someone who has watched what gets formed in the apprenticeship and who is now in the room when the apprenticeship gets dismantled.
I want to be careful about one thing from the start, because it is the thing most of the coverage gets wrong. This is not, at bottom, a story about AI. AI is the instrument. The story is older and more general than the instrument, and if we let the current label stand in for the whole phenomenon we will misdiagnose it. What is actually happening is that a piece of automation has become cheap enough and good enough to absorb a category of work, and that category of work happened to be carrying a second function nobody had priced in. The same thing has happened before, with other instruments, in other trades. It will happen again with instruments that have not been built yet. The label on the tool is the least durable part of the problem.
The question I cannot stop turning over is the one in the title. We have a model for where senior lawyers come from. It is the model my grandparents would recognize and so would my great-grandparents. Junior lawyers learn by doing tedious work under supervision until enough of the patterns settle into them that they can do less tedious work under less supervision until eventually they can do the work themselves and supervise someone else doing the tedious version. The model is hundreds of years old. It is the model that produced every lawyer alive today.
And it is breaking. Not in some far-off future. Now.
So where do junior lawyers come from?
The work was the school
The legal apprenticeship, like every other apprenticeship that has worked, ran on a particular kind of trade. Junior associates and clinical students and research assistants did work that was tedious, repetitive, and below the hourly rate of the senior people who supervised them. The senior people did not assign this work because it was educational. They assigned it because it had to be done by someone and assigning it to a $200 an hour associate was cheaper than doing it themselves at $900 an hour. The educational function was a happy accident. The work was the school, and nobody had to be told that, because nobody had to design it that way.
It is worth being precise about the mechanism, because the mechanism is what the technology breaks. The apprenticeship was funded by an arbitrage. There was a gap between what the tedious work was worth and what a senior practitioner's time was worth, and the gap was wide enough that handing the work down the ladder was the rational move. The learning rode along on that economic decision for free. No partner ever sat in a compensation meeting and budgeted for the formation of the next generation. They budgeted for getting the document review done at the lowest defensible cost, and the formation happened in the margins of that calculation, invisibly, as a side effect. An apprenticeship is what you get when the cheapest way to do necessary work also happens to be the way that makes a person.
Collapse the arbitrage and you collapse the apprenticeship, whether or not anyone intended to. That is the whole danger in one sentence. When the marginal cost of the tedious work falls toward zero, the economic logic that put it in junior hands disappears, and the formation that was riding along disappears with it. Nobody decides to stop training the next generation. They decide to stop paying a person to do something a machine now does for almost nothing, which is a completely defensible decision, and the training was never on the ledger to begin with, so its loss does not show up anywhere a decision-maker would look.
Document review trained you to recognize the structure of an argument when it was buried in eight hundred pages of correspondence. First-pass legal research trained you to understand what makes a case actually controlling versus merely cited. Drafting boilerplate trained you to read other people's boilerplate with suspicion. Summarizing depositions trained you to hear what a witness is doing with their answer, not just what they are saying. Building a chart of authorities trained you to feel the weight of a doctrine before you could articulate why it had weight.
None of these skills were taught directly. They emerged from the repetition. You did the work, you made mistakes, someone senior caught the mistakes, you internalized the correction, and eventually you had absorbed enough corrections that the work felt different in your hands. This is what apprenticeship is. It is not a curriculum. It is a structure that produces a certain kind of person by giving them a certain kind of work to do for a certain length of time.
The legal profession built its entire pipeline around this structure. Law school is a credentialing prerequisite, not a sufficient training. Nobody comes out of three years of law school knowing how to be a lawyer. They come out knowing how to be the kind of person who can begin to learn how to be a lawyer, which is what the apprenticeship phase exists to convert.
I am being specific about this because I think the conversation about technology in legal practice has mostly skipped past it. People talk about whether the software will replace lawyers. That is the wrong question. The question is whether the software will replace the work that has, until now, served as the school. And the answer to that question is already in. It has. It is. It will continue to.
What the machine actually does
I want to get concrete about the mechanism, because the abstraction hides the part that matters, and because most people writing about this have never watched it run.
Take document review, the canonical first-year task. The old version: a corpus of emails, contracts, and filings lands on a junior associate's desk, and the associate reads through it flagging what is responsive, what is privileged, what is hot. The new version is a pipeline, and I have built versions of it. The documents are ingested and broken into chunks. Each chunk is converted into a vector, a long list of numbers that encodes its meaning rather than its exact words, so that a search for indemnification surfaces the paragraph that discusses holding a party harmless without ever using the term. Those vectors get indexed. When a question comes in, the system retrieves the chunks nearest to it in meaning, hands them to a model with a large enough context window to hold the relevant slice of the corpus at once, and the model classifies, summarizes, and flags. A human reviews a sample, or reviews the flags, and signs off.
First-pass research runs on the same skeleton. Retrieve the authorities that are semantically close to the question, synthesize them, rank them, surface the ones that look controlling. Deposition summary, the same. Boilerplate comparison, the same. The architecture barely changes from task to task. Retrieve, read, summarize, flag, hand a human the result for a check. It is, in the aggregate, very good at exactly the work that used to constitute the first two years.
Here is the part the architecture diagram makes legible that the discourse does not. In the old pipeline, the junior associate was the retrieval-and-first-pass layer. Being that layer, badly at first and then less badly, was the entire mechanism of formation. You retrieved the wrong cases and a senior associate told you why they were wrong, and the next time you retrieved fewer wrong cases, and the texture of controlling versus merely cited slowly precipitated out of a thousand small corrections. The corrections were the curriculum. The system has now taken over that layer, and it does have a learning loop of its own, but the loop runs in the wrong place for our purposes. The model's corrections accrue to the model. They are absorbed through evaluation and tuning into a product, owned by a vendor, that gets a little better with each cycle. They do not deposit into a person who is going to be a partner in 2046. The substrate that used to accumulate inside a human being now accumulates inside a piece of software. The learning did not stop. It moved. It moved out of the people and into the tool, and the people are who we needed it to stay in.
That is the precise shape of the loss, and it is not specific to law and not specific to this generation of models. It is what happens whenever the layer of work that formed practitioners gets relocated into a system that learns on its own behalf instead of on theirs.
What is different this time
The legal profession is not the first to lose its apprenticeship to a new instrument. Typesetters had their craft dissolved by desktop publishing. Draftsmen had their craft dissolved by CAD. Floor traders had their craft dissolved by electronic execution. In each case, the early-career work that had served as the school evaporated within a single generation, and the senior practitioners had to figure out what to do about it.
So the legal version of this is not unprecedented. I want to be careful not to dramatize it. People in other professions have faced this before and the world did not end. Lawyers are not uniquely entitled to a stable career path.
But there are two things genuinely different this time, and being precise about them matters.
The first is the speed. The radiologist transition, where image-recognition software started doing work that junior radiologists had previously done, played out over roughly two decades and is still ongoing. The legal version is playing out over three to five years. I have watched it accelerate in real time. The first deployments I worked on were proofs of concept that took six months to plan and a quarter to roll out. The most recent ones take weeks. The capability is improving faster than the institutions can absorb the implications of it, and the institutions include law schools, bar associations, partnership tracks, and the people who design first-year associate programs. The change is faster than the response, and that gap is going to do real damage to the people trying to enter the profession right now.
The second difference is broader than the speed, and it is the one I think about most. In the earlier cases, the work that got automated was the visible craft. A typesetter's job was the typesetting; when it went, everyone could see the job had gone, and the new skill was named after the new tool. What is being automated now is not the visible craft of a lawyer. Lawyers still appear in court, still counsel clients, still make the calls. What is being automated is the invisible substructure underneath the visible craft, the slow formative work that nobody outside the profession ever saw and that even the profession barely registered as work. You cannot mourn the loss of something you never knew was load-bearing. The desktop-publishing transition was legible because the lost thing was the job. This transition is illegible because the lost thing is the making of the person who does the job, and that was never something anyone put a name on.
Which is why the honest answer about what to learn instead is so thin. When typesetters were displaced, the new skill was desktop publishing; the replacement was visible. When draftsmen were displaced, the new skill was CAD; the replacement was visible. For junior lawyers in 2026, we do not really know what the new skill is, because the thing that was lost was not a skill in the first place. It was the conditions under which skills used to form.
Some firms say the answer is AI literacy, but most of what gets labeled AI literacy is a ninety-minute training on a specific vendor tool that will be deprecated in eighteen months. Some say it is prompt engineering, but the techniques that worked in 2023 mostly do not work in 2026, and the techniques that work in 2026 will mostly not work in 2028. Some say it is judgment, which is the right answer in the sense that judgment is what survives the automation of everything else, and the wrong answer in the sense that you cannot teach judgment in a vacuum. Judgment is what gets built by doing the work you no longer do.
This is the part of the problem I do not have a tidy answer to, and I am suspicious of anyone who does.
What I see from inside the change
I want to be specific about what I have actually watched happen, because I think the value of writing this is mostly in being concrete.
The senior partners I work with mostly understand what is happening. They are not in denial. They will say, more or less openly, that the work they used to give to first and second year associates is now done by software. They will sometimes add that they themselves learned by doing that work, and that they cannot quite explain how their juniors are supposed to learn the same things now. They will say this, and then they will move on to the next agenda item, because there is a deal to close.
The junior associates I work with mostly do not have a clear picture of what they are supposed to be learning. They have absorbed the message that the technology is the future, and they are diligent about learning the tools. But the tools change every few months. The skill of using the current tool is not the same as the skill the apprenticeship was producing. The apprenticeship was producing something deeper, something that took years to settle. The juniors I talk to know, on some level, that they are being shorted on that, but they cannot articulate exactly what they are being shorted on, because the thing they would have gained by doing the slow work is precisely the thing you cannot describe to someone who has not done it.
The law schools are mostly behind. There are exceptions. There are clinical programs and individual professors who are doing serious work on this. But the institution as a whole is operating on a curriculum cycle that is slower than the change, and a hiring cycle for faculty that is slower still. The result is that most law students in 2026 are being trained for a profession that no longer hires the way they are being trained to be hired.
The bar associations are nearly silent. The state regulators are nearly silent. The CLE infrastructure is mostly producing content about how to use specific vendor products, which is useful in the short term and worse than useless in the long term, because it teaches lawyers to think of the whole shift as a series of vendor relationships rather than as a permanent change in the texture of the work.
The people who are doing well, from what I can see, are the people who are figuring out their own version of the apprenticeship from the inside. They are taking on work they do not have to take on. They are sitting in on calls they do not need to be on. They are asking the senior partner to walk them through why she made the call she made. They are reading the machine's output critically and forming a view about where it is wrong, which, done deliberately, turns out to be one of the few ways left to reconstruct the old correction loop by hand. They are, in essence, rebuilding an apprenticeship out of spare parts because the institutional version is not coming.
That is admirable. It is also not scalable. The people who can do that have time, mentorship access, and a certain disposition. Most early-career lawyers do not have all three. The profession is, in effect, producing a small number of well-formed senior associates and a large number of people who have a JD and three years of work experience and not much of the substrate that previous generations of three-year associates had built.
What is actually at stake
I want to name what I think is actually at stake here, because the easy versions of this argument get the stakes wrong.
The stake is not that lawyers will be unemployed. Most lawyers will not be unemployed, at least not in the near term. The stake is not that legal work will get worse. Most legal work will get faster and cheaper, and some of it will get better. The stake is not even that the entry-level jobs are disappearing, although they are.
The stake is that we are about to have a generation of senior lawyers who never developed the substrate that made the previous generation of senior lawyers senior. The substrate is the pattern recognition that gets built by repetition. It is the suspicion that gets built by being wrong and being corrected. It is the feel for a doctrine that gets built by handling it badly and then handling it better.
Senior lawyers do not just do senior lawyer work. They make calls. They make the calls clients are paying for. The calls are good because the lawyer has, somewhere underneath, twenty years of accumulated pattern recognition that lets them feel when something is off before they can articulate why. That feel is the entire product. It is what distinguishes the senior partner from a very confident first-year. And it does not come from anywhere except doing the work, slowly, for a long time, with someone watching.
If the work goes away and the apprenticeship is not replaced with something that builds the same substrate, then in twenty years the senior partner pool will be different. Not smaller. Not unemployed. Different. The pattern recognition will not be there. The calls will be made by people who have never had the experience of being slowly corrected on a thousand small mistakes, because there were no small mistakes to be corrected on, because the machine did the work and absorbed the corrections into itself.
This is the thing I cannot stop thinking about. The damage of this transition is not going to be visible for a while. The juniors who are missing the apprenticeship now will become the senior practitioners in fifteen or twenty years, and the gap will show up in the quality of the calls they make at the top of their careers. We will not be able to attribute it cleanly. We will just notice, vaguely, that the profession is not producing the kind of senior practitioner it used to produce. By then it will be late.
What we owe the next generation
I do not think the apprenticeship model is coming back. The economic logic that built it has been broken by the technology, and the technology is not going to retreat. The senior partners who say they cannot quite explain how their juniors are supposed to learn the same things are correct that something has been lost, and they are also correct that it is not coming back in the form they remember.
But the function the apprenticeship served still has to be served somehow. The pattern recognition still has to be built. The substrate still has to be laid down. If we are not going to do it by giving juniors work that is beneath the senior lawyer's hourly rate, then we have to do it some other way, deliberately, with intention, as a designed thing rather than a happy accident. The whole problem, finally, is that we are being asked to put on the ledger and pay for, on purpose, a thing that for a hundred years we got for free as a side effect. That is a harder institutional act than it sounds, because nobody has ever had to fund it before.
I have ideas about what that could look like. Radically restructured early-career programs that are explicit about the educational function. Simulation-based training that puts juniors in front of synthetic problems that develop the same pattern recognition the old work developed. Deliberate, structured critique of the machine's output as a formal part of training, on the theory that judging a first pass you did not write can rebuild some of what writing the first pass used to build. The deliberate inclusion of slow, inefficient work as part of training, with the understanding that the inefficiency is the point. Apprenticeship structures that move up the skill chain instead of being eliminated, with juniors doing what used to be midlevel work because midlevels are doing what used to be senior work.
I am genuinely not sure which of these is right. I am sure that none of them happen automatically. They have to be built. They have to be funded. They have to be sustained against the constant pressure of the billable hour and the partnership track and the temptation to extract the short-term efficiency of the tools without paying the long-term cost of what gets lost.
The people who built this transition are the people who owe the answer. I include myself in that. I help deploy these tools. I have watched the work I used to do, slowly, in a clinic office, four reads at a time, get done in eight seconds by a system I helped configure. I do not regret the work I do. I think most of these deployments are net good. I also think that the people building this owe something to the people who are coming up behind, and the something we owe them is more than a vendor training and a press release about how the future is here.
We owe them the honest version of what is happening, which is that the school is closing and we do not yet have a replacement for it. We owe them the engineering of a replacement, with the same care we have put into engineering the tools that closed the school. We owe them a profession that takes seriously the question of how the senior practitioners of 2046 are going to be formed.
I keep coming back to my four reads of that petition. The supervising attorney did not sit me down and explain what she had done. I learned it because I had time. The time is gone. Something has to take its place, or in twenty years we will all be paying for what we did not bother to build.
The work was the school. The school is closing. Someone has to figure out what comes next, and the people who built the closing are the ones who owe the answer.