Template:M intro work Large Learning Model

LLM
/ɛl ɛl ɛm/ (also “large language model”) (n.)
Once upon a time, an LLM was a “Master of Laws”: the postgraduate mark of the sensei in the society of legal service providers — either of that, or of the indolence of one not prepared to strike out and put what she has learned into practice — but still: it spoke to perseverance, depth, comprehension and mastery, however pigeon-hearted its motivation.

If the thoughtleaderati are to be believed, now all one needs for that kind of expertise is a different kind of “LLM”: a “large language model”. Artificial intelligence rendered by a pattern-recognising, parallel-processing chatbot.

The legal profession is to ChatGPT, we hear, as poor old Chrissie Watkins was to Jaws.

But there have been contumelious prophecies of its demise before. In the manner of a blindfolded dartsman, Professor Richard Susskind OBE has been tossing them around for decades. Just by random chance, you would expect one to hit the wall at some point.

Is this big law’s Waterloo? Will ChatGPT do for our learned friends what the meteor did to the dinosaurs?

Or will the lawyers, like cockroaches, survive? Might they even turn this to their advantage?

Cui bono?

Who benefits, primarily, from this emergent technology? Experience should tell us that the first — and often the last — to benefit from legal productivity tools are the lawyers. Should we expect this time be different?

Now, it is a truism that she who has a tool uses it, firstly, to improve her own lot. A commercial lawyer’s “lot” is predicated on two things: (1) time taken, and (2) ineffability: the sense that what she does “passeth all muggle understanding”.

It is a happy accident that, generally, (2) begets (1): the more ineffable something is, the longer it takes, and the harder it is to work with. The longer it takes, the more you can charge.

Commercial legal contracts are like that. Long, and once they have calcified into templates, fiddly. For lawyers, this is a capital state of affairs. It is why no commercial law firm on the planet really cares for plain English. Oh, they all say, they do, of course — but come on.

This is, in itself, a neat “simplification defeat device”: if you make a contract template sufficiently convoluted, the one-off cost of simplifying it so vastly outweighs the cost of just “tweaking” and living with it that few clients will ever take that first step to simplify. Even though the the ongoing costs of not rationalising dwarf the one-off costs of doing so, the long-term savings are always over that hump.

And bear in mind it will be the lawyers who deploy LLMs as a tool, not their clients. Why? Because of that ineffability. An LLM is a pattern-matching device. It understands nothing. It cannot provide unmediated legal advice. It can only ever be a “back-breaker”: the “last mile” needs a human who knows what she is doing, understands the context and complicated human psychology at play in the cauldron of 1 negotiation. An LLM can draw pretty, impressive-at-a-distance doodles, but it cannot do that. Nor can it write legal opinions — well, not meaningful ones — and nor, unmediated by a law firm, does it have the insurance policy or deep, suable pockets for which a client is paying when it seeks legal advice in the first place.

An LLM can only be deployed, that is to say, by someone with skin in the game; who is prepared to put herself in jeopardy by accepting the assignment, which jeopardy she defends by the simple expedient of knowing what she is doing and checking her LLM’s output.

That someone will be a lawyer.

Now such a “last mile” lawyer could use an LLM to simplify documents, accelerate research and break legal problems down to their essences, thereby reducing the cost, and increasing the value, of her service to her clients. And, sure: in theory, she could give all this value up to her clients for nothing.

But she could, just as easily, use an LLM to further complicate the “work product”: to overengineer, to convolute, to invent options and cover contingencies of minimal utility: she could set her tireless symbol-processing engine to the task of injecting infinitesimal detail: she could amp-up the ineffability to a level beyond a normal human’s patience.

Which of these, realistically, should we expect a lawyer to do? Simplify, or complicate? Sacrifice time and ineffability, for the better comprehension of the unspecialised world? Or plough the energy this magical new tool bestows into generating more convolution and ineffability, racking up more recorded time, and building up the bulwark against the muggles?

She would do the latter with only the best intentions, of course; this is not lily-gilding so much as a noble outreach toward perfection: using the arsenal at her disposal to reach ever closer to the Platonic form.

Cynical, or just realistic? Foretellers of legal Armageddon must explain away some difficult facts: that the commercial-legal industrial complex has stubbornly resisted all attempts at simplification and disintermediation for a generation, notwithstanding the thought-leadership, regulatory prompting, appeals to logic and 40 years of enabling technology — Microsoft Word, mainly — which the world’s lawyers could have used, powerfully, to simplify and minimise the legal work product.

Not only did they not do that, they used their tools to make everything more complicated. Boilerplate blossomed. Templates flowered. Even trivial contracts acquired wording dealing with counterparts, governing the form of amendments and excluding third party rights that weren’t there in the first place.

This is a perfect job for ChatGPT. Why should a difference engine designed to generate plausible-sounding but meaningless text be used do anything different?

You can see the effect is is having on legal work product. NDAs grow ever longer, increasingly riven with the same generic ornamentations that usually range between harmless and misconceived but which are now so prevalent — they recur as the LLMs hone their model — as to become hard for the meatware to resist.

The meatware, remember, has limited patience with NDAs, understanding in a way an algorithm cannot how much of a pantomime they are. Algorithms, on the other hand have unlimited patience and boundless energy. If negotiation comes down to who passes out first, we should bear in mind that LLMs don’t pass out.

Who’s client? Oh, right: she’s a lawyer, too.

“But, JC, come on. Be realistic. It is dog-eat-dog out there. Any lawyer keeps the bounty of the LLM from her clients will soon have her lunch eaten by others who won’t. You cannot fight the invisible hand. We are in a race to the bottom.”

But are we?

Ignoring how impervious to the invisible hand all other recent technologies have been, remember who the clients are. Consumers of high-end commercial legal services are not, generally, the permanently bamboozled muggles of common myth. Most are themselves lawyers, inhabiting weaponised legal departments mainly comprised of veteran deal lawyers. These are people also take pride in their ability to work with difficult, complicated things. This is how they prove their worth to their employers.

Lawyer and their clients, that is to say, have a common interest in convolution for its own sake. They are the jazz aficionados of text; cinéastes of syntax. They expect overwrought contracts: nothing says “prudent management of existential risk” like eighty page of 10pt Times New Roman.

Plain English is not for serious people.

Conservative motivation

Nor should we underestimate the overwhelming power of the lawyer’s intuition that what has gone before is sacrosanct.

Lawyers are the last great positivists: they understand instinctively that what has been already laid down by someone else — “posited” — is safer and than anything new that they might themselves contribute. The common law with its doctrine of precedent, after all, is to all intents a divine commandment: in times of doubt, to do what has been done before.

The more authoritative the source, the more sacred it will be.

Thus, lawyers will assiduously “track the wording of legislation” to ensure their drafting matches it with utmost fidelity, notwithstanding any private reservations they may have about how it was drafted. The more ambiguous, or just difficult the source text, the more assiduously should we expect lawyers to replicate it, because they fear it. They fear the limits of their own mastery.

This “positivism-through-fear” extends with equal force to established market precedents. It doesn’t matter how manifestly unfit for purpose it is, the resistance to change will be strong.

Literary theory, legal construction and LLMs

“What an astonishing thing a book is. It’s a flat object made from a tree with flexible parts on which are imprinted lots of funny dark squiggles. But one glance at it, and you’re inside the mind of another person, maybe somebody dead for thousands of years. Across the millennia, an author is speaking clearly and silently inside your head, directly to you.

Writing is perhaps the greatest of human inventions, binding together people who never knew each other, citizens of distant epochs. Books break the shackles of time. A book is proof that humans are capable of working magic.”

— Carl Sagan, Cosmos

“I think you underestimate the power of reading, Professor Sagan.”

—JC

Theres’ a hole in my model, ELIZA

Fittingly, the first chatbot was a designed as a parlour trick. In 1966 Joseph Weizenbaum, a computer scientist at MIT created the programme “ELIZA” to explore communication between humans and machines. ELIZA used pattern matching and substitution techniques to generate realistic conversations. You can try it out here.

By today’s standards, ELIZA was rudimentary, simply regurgitating whatever was typed into it, reformatted as an open-ended statement or question, thereby inviting further input.

Even though ELIZA was a basic “keepy uppy” machine, it proved surprisingly addictive. Weizenbaum was famously shocked how easily people were prepared to believe ELIZA “understood” them and contributed meaningfully to the interaction. When trying it out, Weizenbaum’s own secretary, who had watched him build the programme for months and knew how it worked, shooed him out of the room so she could have privacy with the machine.

This is, of course, how all “mind-reading” works: by asking the right kinds of question, the conjurer extracts from the subject all the information needed to create the illusion of telepathy.

LLMs as conjuring tricks

LLMs work the same way. Like all good conjuring tricks, generative AI relies on misdirection: its singular genius is that it lets us misdirect ourselves, into wilfully suspending disbelief, never noticing that it is we who are doing the creative heavy lifting to turn machine-made screed into magic. We are neuro-linguistically programming ourselves to be wowed by LLMs when the clever part is really happening inside our own heads.

By writing prompts, we create an expectation of what we will see. When the pattern-matching machine produces something roughly like that, we use our own imaginations to backfill, frame, filter, correct, boost, render, sharpen and polish the output into what we wanted to see. We construe the output to conform to our original instructions.

When we say, “fetch me a tennis racquet”, and the machine comes back with something more like a lacrosse stick, we are far more impressed than we would be of a human who did the same thing: we would think such a human a bit dim. But with generative AI we don’t, at first, even notice we are not getting what we asked for. We might think, “oh, nice try!” or “well, that will do,” or perhaps, “ok, computer: try again, but make the basket bigger, the handle shorter, and tighten up the net.” We can iterate this way until we have what we want — though note all the refining intelligence is coming from us — or we could just Google for a conventional photo of a tennis racquet. As the novelty wears off, that is what more and more people will do.

One of the more fantastic claims for Google glass was that it could read subtle clues in facial expressions to detect underlying emotional state of the subject:

The app detects emotion in the faces it sees using an algorithm generated by a machine learning system. This AI system was trained on large datasets of faces to decode the emotions from facial expressions.

But humans don’t need AI goggles to read each others’ emotions. They are really good at that — far better than any algorithm — already. Even dogs can read facial expressions.

So it is with generative AI. First impressions can be stunning, but the second look reveals an absurdist symphony. AI image generation struggles with hands, eyes and logical three-dimensional architecture. It is just as true of text prompts: on closer inspection we see the countless minute logical cul-de-sacs, bad guesses and non sequiturs from which a clever story miraculously emerges. (To be sure, many human authors write in logical cul-de-sacs, bad guesses and non sequiturs, but that is another story.) Again, where is the magic in this process? Inside the meatware.

Either way, the novelty soon palls and, as we persevere, we begin to see more and more of the magician’s wires. The first couple of passes are great: they get from zero to 0.5 of what we wanted. But the marginal improvement in each following round diminishes, as the machine reaches asymptotically towards an upper capability in “producing what you had in mind”, which we estimate, unscientifically, as about 75% of it.

As we go, we get a sense of how the model works: its familiar tropes, tics and consistent ways of doing things which are never quite what you have in mind. The piquant surprise at what it produces dampens at each go-round, eventually settling into an entropic and vaguely dissatisfying quotidian.

Now, as generative AI improves — assuming it does improve: there are some indications it may not; see below — the threshold of its “expectation-meeting capability” may move towards 100% but will never quite get there. And, again, don’t underestimate how important the meatware is in that refining process: many of its improvements will be down to how we learn to better to frame our queries. “Prompt-engineering” becomes the real skill, rather than the dumb, parallel pattern-matching process that responds to it.

Ours is the skill going in, and ours is the skill construing the output. The machine just does the boring bit. But this is what machines have always done: the bits that require strength, speed, reliability, and economy. Not ingenuity. This is the basic proposition of mechanising. Humans have been mechanising things since someone invented the wheel.

LLMs and literary theory

POLONIUS: What do you read, my lord?
HAMLET: Words, words, words.
POLONIUS: What is the matter, my lord?
HAMLET: Between who?
POLONIUS: I mean the matter that you read, my lord.

—Hamlet, II, ii

Now, in all kinds of literature bar one, “reading” is where the real magic happens. Construal. It is the emergent creative act and community consensus that renders Hamlet a timeless cultural leviathan and Dracula: The Undead forgettable pap. A literary work may start with the text, but it stays there barely a moment. The “meaning” of literature is necessarily personal to the reader: it lives between our ears, and within the cultural milieu that interconnects the reading population over the generations.

Call him post-modern — go on, do — but the JC doesn’t hold with Carl Sagan’s idea that a book teleports its author “inside our heads”. That would be to equate reading with symbol-processing. It absolutely isn’t, and that metaphor gravely underestimates the human brain when in construction mode. Nor, by the way, is this in any way to diminish Shakespeare’s towering genius, but rather to observe that, however impossibly brilliant it is, it is swamped by the flood of exposition, analysis, interpretation, re-rendition and performance, that has gone on since he published it, not to mention reader’s own role in “construction”.

“Construal” and “construction” are interchangeable in this sense: over time that cultural milieu takes the received corpus of literature and, literally, constructs it into edifices its authors can scarce have imagined. Hamlet speaks, still, to the social and human dilemmas of the twenty-first century in ways Shakespeare cannot have contemplated. to be clear: a reader is no simple “symbol processor”, decrypting text to reveal a one-to-one assembly of the author’s intention in her own head. Literature is no instruction manual, recipe nor a computer programme.

Now there is one kind of “literature” that is like a computer programme: where the last thing the writer wants is for the reader use her imagination, to construct a meaning: where clarity of authorial intention is paramount; where communicating and understanding purpose is the sole priority: legal literature.

Rather than ceding interpretative control to the reader, a legal drafter seeks to squash all opportunities for improvisation and stomp out all ambiguity. Just as there are no atheists in foxholes, there are no metaphors in a Trust Deed.

Legal drafting seeks to be as finite as it can be. It strives do to readers what code does to hardware: to reduce them to mere symbol-processing machines, extracting the author’s single incontrovertible meaning. But, in a natural language that is constructed out of dead metaphors, this is very, very hard to do. That there is such a living to be made conducting commercial litigation shows that.

It is one reason why legalese is so laboured. It is compelled to chase down all blind alleys, previsualise all phantoms and prescribe outcomes for all logical possibilities. To remove all possible ambiguity and render the text as mechanical, precise and reliable as it can be. There are no metaphors in a trust deed. Where normal literature favours possibility over certainty, legal language bestows certainty at the cost of possibility, and to hell with literary style and elegance.

Where literary language is, in James Carse’s sense, infinite, legal language is finite.

Now: the punchline. Given how integral the reader and her cultural baggage are to the creative act in normal literature, we can see how, in that domain, a large learning model, which spits out text ripe with interpretative possibilities, begging for someone to “construct” it, is a feasible model for that kind of language: to move from a model where most of the creative work is done by the reader to one where all of it is, is no great step.

There is enough bad human literature in existence already, that is is no great stretch to imagine doing without the human altogether. Who cares what the text is meant to say, as long as it is coherent enough for an enterprising reader to make something out of it?

But that does not work at all at all for legal language. Legal language is code: it must say exactly what the parties require: nothing more or less, and it must do it in a way that leaves nothing open to a later creative act of interpretation. Legal drafting is as close to computer code as natural language gets: a form of symbol processing where the meaning resides wholly within and is fully limited by the text.

But unlike computer code, the operating system it is written for is not a closed logical system, and even the best-laid code can still run amok. You can’t run it in a sandbox to see if it works. You have to test in production.

This is not to say that a large language model can’t be used to generate legal boilerplate: it just can’t do it by itself, and the process of working with it will be a lot more labour-intensive than the first round of generation suggests. An LLM will silt towards an intended meaning asymptotically, getting progressively less efficient as it goes.

We have seen it suggested that one might invert the processes instead, and use the machine to critique drafts prepared by humans, to identify potential errors and omissions. But this is to get the division of labour exactly backwards, using expensive, context-sensitive meatware to do the legwork and a dumb machine to provide the “magic”.

Without wishing to seem all John Connor about it, there is a question of basic human dignity going on here. We need to stop subordinating ourselves to machines. We should not compromising just to optimise for machine processing, to make it easier for machines to manage. There is a question of self-belief — self-respect — here.

And besides, there is a design flaw in any legal process which supposes that the risk in a legal contract is distributed evenly throughout its content, and that therefore the legal proposition is one of handling volume.

Boilerplate is boilerplate for a reason. It is pinned down; done, it takes the goes without saying out of the equation. The quotidian is a utility, not an asset. There is nothing to be gained from having a large language model drafting it from scratch each time. The important part of the legal drafting process is not assembling the boilerplate, but getting the deal-specific bespoke bits right.

Meet the new boss —

We don’t doubt that LLM is coming, nor that the legal industry will find a use for it: just that there is a useful, sustained use for it. It feels more like a parlour trick: surprising at first, diverting after a while, but then the novelty wears off, and the appeal of persevering with what is basically a gabby but unfocussed child wears pales.

The traditional legal model faces existential challenges for sure, but they are not presented, and will not be addressed by random word generators.

Coda: is ChatGPT getting worse?

In other news, scientists are concerned that ChatGPT might be getting worse. Studies indicate that its accuracy at tasks requiring computational accuracy, like playing noughts and crosses or calculating prime numbers, is rapidly diminishing.

Perhaps ChatGPT is getting bored, or might it have something to do with the corpus increasingly comprising nonsense text generated on the hoof by some random using ChatGPT?