Your argument reminds me of domestic cleaning. When we got machines to do laborious tasks, we raised expectations for basic livability and then raised the bar for what it means to be clean.
Within legal practice, there's a division between "products" for which lawyers' work is the principal input and products for which lawyers' work is a secondary or complementary input. For instance, lawyers are the primary input for the product "resolve a major commercial dispute," but only a small component of "launch a data center." For the first category, a positive productivity shock ought to generate more quantity. Ordinary Americans were wayyyy more litigious in the 19th century than today, and I could imagine a resurgence of cheap neighbor-on-neighbor lawsuits that were the stock in trade of lawyers like Abe Lincoln before Baumol's cost disease hit the legal profession. Personal bankruptcy and professional estate planning ought to come within the reach of more people, which is a good thing. But for secondary inputs, I would think that the "lump of labor" fallacy might come closer to holding. I don't see a profusion of data centers simply because lawyers cost less. (But, as a part-transactional lawyer, I already see your "adversarial" dynamic at play. Lawyers who might have raised 10 substantive issues with a document three years ago, now find 40 substantive issues to fight over in the same document.) Overall, I predict a welcome swing away from "papering high-class business transactions" and back toward the core business of providing legal services to ordinary people.
(Criminal law ought to be cheaper following a productivity shock, but I doubt the quantity demanded will vary much, since quantity of criminal law practice is fixed by the government, which, in some sense, determines the number of defendants and guarantees them all a baseline minimum quantity of legal representation.)
Super interesting, I was not aware (and in fact assumed the opposite) of the past's greater litigiousness. And yes, the distribution between B2B/B2C lawyering is likely one area where the transition will take place.
Paralegal employment has continued to rise with overall legal employment since the initial internet shock (say 1995) as some traditional paralegal work gets automated while the overall volume of work balloons and paralegals can take on even more complex tasks. In contrast, I believe legal secretarial employment has proportionately contracted, as lawyers and paralegals now do the work themselves (word processing, filing, scheduling) and share secretaries. The story may be similar with AI, which certainly does mean changes to the work mix and junior employment pipeline, but may not presage the death or even shrinkage of the legal industry.
While theoretically appealing, the author’s framework relies heavily on historical labor analogies that may fail to account for the fundamentally different nature of cognitive automation. A fair evaluation requires separating the volume of legal tasks from the volume of human labor required to execute them.
The author’s use of the Jevons paradox assumes that an increase in regulatory complexity (more norms) dictates a proportional need for human legal work. However, this assumes human labor remains the primary vehicle for compliance. If an AI system reduces the marginal cost of parsing and drafting compliance documents to near-zero, a 10x increase in regulatory complexity will likely be met with an increase in compute, not a hiring surge of junior associates.
The argument that lower costs will unlock vast latent demand—such as basic estate planning or small business contracts—is empirically sound. However, the assumption that this translates to more jobs for traditional lawyers is structurally flawed. Latent, low-to-mid-tier legal demand is highly susceptible to "winner-take-all" platform economics. A few highly efficient, AI-driven legal tech platforms could absorb millions of newly viable consumer cases using a skeleton crew of human overseers, effectively bypassing and hollowing out the middle class of standard practitioners.
The "Red Queen" effect posits that an adversarial escalation requires human lawyers to work harder to maintain an edge. A commenter rightly notes that lawyers who once found 10 issues in a document might now find 40. Yet, finding 40 issues instead of 10 is fundamentally a pattern-matching and generation task—the exact domain where Large Language Models excel. The adversarial arms race will primarily occur between opposing AI agents drafting and redlining against each other. The human lawyer's role narrows to a final review, drastically reducing the billable hours historically associated with adversarial escalation.
Determine which are adversarial, which are bounded, and which are unbounded, or a mix of these. The adversarial resist any efficiency gains, the unbounded respond to Baumol taking share from bounded, but leaving us with an approximation of full employment by default: https://substack.norabble.com/p/the-ai-jobs-blind-spot
The piece and several comments analyze legal work through a production lens - who does the work, what does it cost. But lawyers aren't just producers of legal output. They're liability sinks. When you hire a lawyer, you're buying someone whose name goes on the document, who carries malpractice insurance, who can be sued, who faces bar discipline. You're transferring risk. That function doesn't get cheaper when AI makes the drafting cheaper, because it has nothing to do with the cost of producing text. Until AI can be a legal person - and nobody is close to resolving who pays when it's wrong - every AI-generated legal document needs a human willing to put their name on it and accept consequences.
Which means what really happens when someone uses an LLM for legal work is that they are assuming all that risk themselves, and most of them don't know it. They think they're getting cheap legal work. They're actually getting uninsured legal work. The AI doesn't say "by using this output you are now your own lawyer." It just hands you something that looks professional. The people who can safely use AI for legal work are the people who already know enough to catch the errors - which means the people who already had the training. For everyone else, it's not democratization. It's the illusion of access with a hidden transfer of liability to the person least equipped to carry it.
It is a similar process to using LLMs to write computer code. It often looks good, and seems to work, but can fail spectacularly in unexpected ways.
When AI generates false quotations attributed to real cases, a new “verification” economy arises.
Of course, this secondary economy should not exist. Finding bogus computer output is not legal work at all.
It’s entirely possible that because output cannot be trusted, case-related legal AI is already dying — bought by early adopters but then ignored — just Alexa Skills died after billions of dollars invested in early adoption.
What happens to legal AI if the case-law-related products simply sit there, unused?
I think the false quotations are more of a function of how LLMs are trained- first on a massive general corpus of text, and only then refined to a specific task. In the first section, getting things "looking good enough" is a critical part of getting the training to converge, but the side effect is that hallucinations of facts can occur. If I were training an LLM from scratch to be able perform legal lookups and/or text generation, I would be far more strict on not allowing fictitious citations. There's nothing inherent to the LLM technology that implies that is an impossible task (though it remains to be seen if it's a practical one).
LLMs are next-token prediction machines. They have no internal state for "I don't know." When asked about a case that doesn't exist, the model doesn't experience a lookup failure - it generates what a citation would look like, with the same confidence as a real one. You can reduce the rate with retrieval-augmented generation and fine-tuning, but you can't eliminate it without building something that isn't an LLM anymore. On the well-trodden path (common case law, frequently cited precedents), the correct answer is also the most probable token sequence, so the output is reliable. Off the path, the model doesn't know it's left the path.
Your argument reminds me of domestic cleaning. When we got machines to do laborious tasks, we raised expectations for basic livability and then raised the bar for what it means to be clean.
Exactly !
Within legal practice, there's a division between "products" for which lawyers' work is the principal input and products for which lawyers' work is a secondary or complementary input. For instance, lawyers are the primary input for the product "resolve a major commercial dispute," but only a small component of "launch a data center." For the first category, a positive productivity shock ought to generate more quantity. Ordinary Americans were wayyyy more litigious in the 19th century than today, and I could imagine a resurgence of cheap neighbor-on-neighbor lawsuits that were the stock in trade of lawyers like Abe Lincoln before Baumol's cost disease hit the legal profession. Personal bankruptcy and professional estate planning ought to come within the reach of more people, which is a good thing. But for secondary inputs, I would think that the "lump of labor" fallacy might come closer to holding. I don't see a profusion of data centers simply because lawyers cost less. (But, as a part-transactional lawyer, I already see your "adversarial" dynamic at play. Lawyers who might have raised 10 substantive issues with a document three years ago, now find 40 substantive issues to fight over in the same document.) Overall, I predict a welcome swing away from "papering high-class business transactions" and back toward the core business of providing legal services to ordinary people.
(Criminal law ought to be cheaper following a productivity shock, but I doubt the quantity demanded will vary much, since quantity of criminal law practice is fixed by the government, which, in some sense, determines the number of defendants and guarantees them all a baseline minimum quantity of legal representation.)
Super interesting, I was not aware (and in fact assumed the opposite) of the past's greater litigiousness. And yes, the distribution between B2B/B2C lawyering is likely one area where the transition will take place.
Paralegal employment has continued to rise with overall legal employment since the initial internet shock (say 1995) as some traditional paralegal work gets automated while the overall volume of work balloons and paralegals can take on even more complex tasks. In contrast, I believe legal secretarial employment has proportionately contracted, as lawyers and paralegals now do the work themselves (word processing, filing, scheduling) and share secretaries. The story may be similar with AI, which certainly does mean changes to the work mix and junior employment pipeline, but may not presage the death or even shrinkage of the legal industry.
While theoretically appealing, the author’s framework relies heavily on historical labor analogies that may fail to account for the fundamentally different nature of cognitive automation. A fair evaluation requires separating the volume of legal tasks from the volume of human labor required to execute them.
The author’s use of the Jevons paradox assumes that an increase in regulatory complexity (more norms) dictates a proportional need for human legal work. However, this assumes human labor remains the primary vehicle for compliance. If an AI system reduces the marginal cost of parsing and drafting compliance documents to near-zero, a 10x increase in regulatory complexity will likely be met with an increase in compute, not a hiring surge of junior associates.
The argument that lower costs will unlock vast latent demand—such as basic estate planning or small business contracts—is empirically sound. However, the assumption that this translates to more jobs for traditional lawyers is structurally flawed. Latent, low-to-mid-tier legal demand is highly susceptible to "winner-take-all" platform economics. A few highly efficient, AI-driven legal tech platforms could absorb millions of newly viable consumer cases using a skeleton crew of human overseers, effectively bypassing and hollowing out the middle class of standard practitioners.
The "Red Queen" effect posits that an adversarial escalation requires human lawyers to work harder to maintain an edge. A commenter rightly notes that lawyers who once found 10 issues in a document might now find 40. Yet, finding 40 issues instead of 10 is fundamentally a pattern-matching and generation task—the exact domain where Large Language Models excel. The adversarial arms race will primarily occur between opposing AI agents drafting and redlining against each other. The human lawyer's role narrows to a final review, drastically reducing the billable hours historically associated with adversarial escalation.
I agree. In modeling changes in job demand, we ought to be doing the same for all industries: https://substack.norabble.com/p/ai-jobs-the-hidden-rules-of-demand
Determine which are adversarial, which are bounded, and which are unbounded, or a mix of these. The adversarial resist any efficiency gains, the unbounded respond to Baumol taking share from bounded, but leaving us with an approximation of full employment by default: https://substack.norabble.com/p/the-ai-jobs-blind-spot
The piece and several comments analyze legal work through a production lens - who does the work, what does it cost. But lawyers aren't just producers of legal output. They're liability sinks. When you hire a lawyer, you're buying someone whose name goes on the document, who carries malpractice insurance, who can be sued, who faces bar discipline. You're transferring risk. That function doesn't get cheaper when AI makes the drafting cheaper, because it has nothing to do with the cost of producing text. Until AI can be a legal person - and nobody is close to resolving who pays when it's wrong - every AI-generated legal document needs a human willing to put their name on it and accept consequences.
Which means what really happens when someone uses an LLM for legal work is that they are assuming all that risk themselves, and most of them don't know it. They think they're getting cheap legal work. They're actually getting uninsured legal work. The AI doesn't say "by using this output you are now your own lawyer." It just hands you something that looks professional. The people who can safely use AI for legal work are the people who already know enough to catch the errors - which means the people who already had the training. For everyone else, it's not democratization. It's the illusion of access with a hidden transfer of liability to the person least equipped to carry it.
It is a similar process to using LLMs to write computer code. It often looks good, and seems to work, but can fail spectacularly in unexpected ways.
This is indeed another argument I commonly make (but could not fit the recent columns); thanks for pointing it out here !
When AI generates false quotations attributed to real cases, a new “verification” economy arises.
Of course, this secondary economy should not exist. Finding bogus computer output is not legal work at all.
It’s entirely possible that because output cannot be trusted, case-related legal AI is already dying — bought by early adopters but then ignored — just Alexa Skills died after billions of dollars invested in early adoption.
What happens to legal AI if the case-law-related products simply sit there, unused?
I think the false quotations are more of a function of how LLMs are trained- first on a massive general corpus of text, and only then refined to a specific task. In the first section, getting things "looking good enough" is a critical part of getting the training to converge, but the side effect is that hallucinations of facts can occur. If I were training an LLM from scratch to be able perform legal lookups and/or text generation, I would be far more strict on not allowing fictitious citations. There's nothing inherent to the LLM technology that implies that is an impossible task (though it remains to be seen if it's a practical one).
LLMs are next-token prediction machines. They have no internal state for "I don't know." When asked about a case that doesn't exist, the model doesn't experience a lookup failure - it generates what a citation would look like, with the same confidence as a real one. You can reduce the rate with retrieval-augmented generation and fine-tuning, but you can't eliminate it without building something that isn't an LLM anymore. On the well-trodden path (common case law, frequently cited precedents), the correct answer is also the most probable token sequence, so the output is reliable. Off the path, the model doesn't know it's left the path.