The Lump of Law Fallacy

You have never met a norm-maker

One standard answer to the worries about “AI taking all the jobs” consists in pointing to the “lump of labour” fallacy.

In its simplest form, the fallacy is as follows: there is a finite amount of “work” to be done, such that it can be redistributed by tweaking the parameters of who works how and on what conditions; it is also vulnerable to external shocks on the supply or demand of labour. Reduce, say, the number of standard work hours (as France did when moving to a 35-hour week) and, the theory goes, you can liberate enough work to fix unemployment.

The fallacy works because it is intuitively appealing. Many, if not most things in life can be shared in ways that are zero-sum: material objects, certainly, but also more intangible things like bandwidth or even “attention”. Yet assuming the same of “work” is fallacious, for the reason that “work” is a shorthand for a complex web of relationships and obligations between humans and institutions. Working more, or less, has an incidence on your income, your consumption, and then the work of someone else, in a dynamic relationship that has endless ramifications.

As such, the historical record has not been kind to the past Cassandras warning about external shocks (new tech, input of migrant labour, etc.) destroying or taking up all the jobs. And France did not beat unemployment.

Now, a very similar argument is being made, implicitly or explicitly, in the debate about AI and lawyers: legal work is a fixed quantity, some assume, and if AI can do instead of jurists, this will take the bread out of the latter’s collective mouth. Watch out, they say, as legions of junior lawyers are decimated by a simple AI routine; look out, they add, for the incredible competition LLMs will put on prices for everyone else. An entire industry is on the cusp of irrelevance, and good riddance, the hope goes.

But since I am on the record saying that we will have more, not fewer lawyers in the medium-term, picture me unconvinced. And a large part of my stance proceeds from the rejection of this “lump of law” fallacy.

Law as an equilibrium

The first thing to point out is that “law” (or “legal work”, for which it’s even more obvious) is, like “work”, a shorthand to denote a much greater, much more complex set of things.

There are, evidently, the positive norms that apply to any subject thereof, the stuff of laws and codes. Then other, regulatory norms designed and enforced by regulators. As well, “private laws” between parties governing their relationships inter se, or self-imposed norms. Soft law as well, I guess. And finally (without pretending to be exhaustive) everything that lives in the penumbra (or however you want to call it) of the former, norms of conduct, expectations, ethics, etc.

And just as with work, touching the legal framework from one side has consequences across all the others. Here are a few examples:

• The violation of one norm might engage other norms (e.g., “everything is securities fraud”);

• Norms can be escaped or evaded, through technological or human means (e.g., tax evasion), calling for more norms to patch the loophole;

• Relatedly, there are “loopholes” or “legal gaps” everywhere, if you listen to, I don’t know, anyone - decrying such gaps is the easiest way to become a TV anchor or commentator, and everyone to nod and applaud with their crab hands;

• Interpretation drift, whereby every application of a norm slightly changes its meaning, requiring work to keep up and adapt; or

• Jurisdictional layering, to ensure that one legal situation remains kosher in a “pluralist” world where you may, if you look hard enough, be subject to many, many different legal frameworks.

All these examples entail that whatever AI does to legal work, more such work is likely on offer, because labor laborem invocat. This stems, to some extent, from the fractal nature of the law, where - as we have discussed - there is no clear answer to what is “good enough”. But it also stems from the fact that several distinct forces - political, adversarial, economic - impact the dynamic equilibrium of the law at once, as we shall see.

The Jevons effect

The first vector is the typical Jevons effect, as applied to the cost of compliance.

Many norms nowadays are enacted through an enterprising norm-maker (a regulator, a legislator, anyone incentivised to create norms) meeting a latent demand (even if weak, fabricated or fantasised) for norms. This demand is, by default, maximalist: you want to regulate the heck out of it. But what prevents legislators from micro-managing everything ? To the extent they have some, common sense, but more generally, the feeling that a norm will be capable of being complied with.

Not that unworkable, uncompliable norms are never enacted; to the contrary,¹ many laws and regulations are maximalist by design, knowing full well that a modus operandi will eventually emerge that will be to the advantage of the norm enforcer (e.g., “we punish you now on this formality that is commonly flouted because we can’t pursue you on other grounds”).

But still, this remains bounded to an extent by what’s at least humanly conceivable in terms of compliance, what a good faith review would uphold and what a given legal community is ready to bear without complaints and/or evasion. The modus operandi itself must be viable; feasibility constraints are part of the political economy of rulemaking.

Now, add AI to the picture. It’s possible, even likely, that AI will reduce the costs of compliance, enhance the ease of complying altogether. For instance, where compliance takes the form of loads of texts and forms no one will ever read, LLMs are expert at writing (and later parsing) such text and it should be eminently feasible to engineer a pipeline that checks boxes that need to be checked.

What will that make of the equilibrium ? Will the norm-entrepreneurs go away ? Rather, they will have a new baseline, and concoct even further norms taking that baseline as granted. We see this every time a sedimented regime is up for “reform”, or “simplification”; EU readers may be aware that the European Commission is currently midway through enacting a set of reforms whose avowed goal is to simplify EU law and which, by all accounts, are still introducing, if not more norms, at least more legal uncertainty.

In the same vein, norms beget meta-norms, such as the rules about technical/technological proficiency, which exist precisely because technology made lawyering more complex. Such meta-norms are meant to ensure the baseline is met so that primary norms are complied with at their level. And they won’t stay still: each generation of tools produces a new generation of expectation and duties to use them competently, which itself becomes the substrate for the next round.

Ease compliance, and the normload goes up.

The adversarial/Red Queen effect

A second vector is that law is a game played against other players, not a catalogue of rules waiting to be applied.

You see this in a lot of contexts: tax planning versus tax enforcement, claimant versus respondent, drafting around a precedent versus litigating to extend it, compliance teams adjusting to regulators who adjust to compliance teams, etc. Legal work is often shaped not only by what the norm says, but by what someone else might do with it.²

AI changes the level of that contest without changing its structure. AI as the proprietary edge of a sophisticated firm is a moat ; but when everyone has it, it is the floor. And as with traders chasing alpha against thousands of sophisticated alpha-seekers, lawyers will have to up their game (and increase their work) to stay level. We are back at the fractal point: demand for legal services can increase because one’s due diligence can always be more thorough, a contract can anticipate ever more contingencies (and balloon to inhuman length requiring AI to ingest), monitoring can embrace an ever-greater scope, etc.

In other words, AI in an adversarial context points to further work: the FTC scraping marketing copy at scale generates demand for compliance review of marketing copy at scale; or the use of AI in a litigation context will incentivise the use of ways to trip an LLM, which will require new defences in reaction to the offence, etc. In all this, the equilibrium amount of legal activity can rise rather than fall, because each side’s improved capability increases the stakes and the sophistication of the contest.

On this note, a lot of the “AI will replace lawyers” story misunderstands the nature and extent of legal demand: that demand does not necessarily and always match what’s needed to be legally secure, since legal certainty is asymptotic; instead, demand derives from endogenous constraints (ability and willingness to pay), partly depends on what the adversary can throw at you, and stops at the “good enough” level.

Like everything else, this demand will adapt to a world where everyone else has AI.

Enter Baumol

There is yet another aspect worth dwelling on.

It’s no secret, and hard to contest, that the supply of legal services has never really met the demand. Part of it is suppression of the supply through credentialism (the bar exam) or deliberate frictions. But another part might be a result of Baumol’s cost disease:³ law was, and still is, a field with little growth in productivity but where remuneration had to track other high-paying sectors; these higher costs, in turn, likely suppressed demand.

Assume now that AI reduces these costs: the story becomes, once again, one of swelling demand: there is every reason to believe the price-elasticity of legal services across the latent demand is high, and to expect many potential consumers of legal services at lower price points.

True, part of that demand will be met at the price point of the cheapest token, but many activities will still require an actual human lawyer, just like many consumers will prefer a professional - this is your typical “why pay a plumber when there are Youtube videos that tell you everything you need”.

Relatedly, lower legal prices will have an impact on unbundling: some legal services are gated by the fact that getting them requires committing to a whole bundle around them; e.g., you could not buy “ten minutes of judgment” from the senior lawyer without also buying the apprentice-hours and associate-hours and document-review-hours that came attached.

I am mentioning lower prices here, but this does not mean the actual danger is lawyers having to work twice as much to make ends meet. If anything, lower price for humdrum legal services will just entrench Baumol’s disease for the parts that remain the bottleneck: human judgment and expertise, the things a client won’t take from a machine, or won’t accept without a human name attached. Senior lawyers will not get cheaper because their associates got cheaper ; they’ll get more expensive, because they are now the constraint on whatever the cheap machinery is meant to produce.

But this time it’s different

The main counter-argument to all this is the same one made against the “lump of labour” fallacy: that however true this has been in the past, this time it’s different.

I don’t want to give the impression that this position has no merit. The advent of LLMs, and what they enable, is truly unprecedented, and I think we have only started to scratch the scope of what’s possible with advanced AI systems. There is the idea, fascinating in itself, that even if the progress in models’ ability stalls today, there would be enough for countless new applications, businesses, ideas, Benthamite utilities to spring forth in the short to medium term.

A lot will certainly happen, and that’s a position I am happy to share - one caveat to what precedes is that “not a lump” does not guarantee that particular lawyers keep their particular jobs; recomposition can be unfriendly at the individual level even when aggregate demand holds. We are in for bumpy times.

Yet, I am not seeing enough evidence that this means jobs will disappear, let alone legal jobs. To achieve this, AI would have to defeat all the mechanisms that make the law dynamic and responsive to new equilibria, and it would have to do so through people in charge of deploying it efficiently. In light of, you know, your standard human, this assumes too much.

And if AI reaches that point, if it can replace us end-to-end, then, and as I pointed out previously, this means I am taking the easy bet here; if I am wrong on (legal) jobs, I’ll look silly, but this will be the least of our (and my) worries.

---

(Programming note: I have enabled pledges for Artificial Authority. Regular posts remain free, and will always be; this is simply for those who find the Substack or the hallucination database useful and want to support it.)

1. I have mentioned already my notion of the “Beria model of the law”.

2. Consider that a lawyer’s answer is typically prefaced (and insured) by dozens of caveats, of which the two most important are (i) given time t and situation s (and absent sudden change of circumstances); and (ii) to the best of our knowledge having exercised due diligence in light of the context you provided. These caveats exist because someone is on the other side.

3. For an explainer.