Like today’s large language models, 16th-century humanists had techniques to automate writing – to the detriment of novelty

The Renaissance scholar and educator Erasmus of Rotterdam opens his polemical treatise The Ciceronian (1528) by describing the utterly dysfunctional writing process of a character named Nosoponus. The Ciceronian is structured as a dialogue, with two mature writers, Bulephorus and Hypologus, trying to talk Nosoponus out of his paralysing obsession with stylistic perfection. Nosoponus explains that it would take him weeks of fruitless writing and rewriting to produce a casual letter in which he asks a friend to return some borrowed books. He says that writing requires such intense concentration that he can do it only at night, when no one else is awake to distract him, and even then his perfectionism is so intense that a single sentence becomes a full night’s work. Nosoponus goes over what he’s written again and again, but remains so dissatisfied with the quality of his language that eventually he just gives up.

Nosoponus’s problem might resonate. Who has not spent too long going over the wording of a simple email, at some point or another? Today there is an easy fix: we have large language models (LLMs) to write our letters for us, helpfully proffering suggestions as to what we might say, and how we might phrase it. When I input Nosoponus’s intended request into GPT-4, it generated the following almost instantly:

Hey [Friend’s Name],

Hope you’re doing well! I just realised I never got those books back that I lent you a while ago. No rush, but whenever you get a chance, I’d love to get them back. Let me know what works for you! Thanks!

Nosoponus

But there was a solution in the 16th century, too. A humanist education on the Erasmian model could train its students to produce letters of any length, on any topic – quickly, easily and eloquently. The French humanist François Rabelais, a contemporary of Erasmus, appears to have understood these compositional techniques as automating the creating of text in a way that, retrospectively, looks a lot like how LLMs function. If we want to understand LLMs, and what they are and aren’t capable of, we can look at earlier versions of the same technology – like Erasmian humanism. We can also read authors like Rabelais, who is already thinking about automatic text-generation along these lines, as someone who appreciates the effectiveness of Erasmian generative technology, but at the same time sees it as vitiating the social force of language and, ultimately, ruining language as a tool for moral and political life.

Rabelais was a monk, doctor, personal secretary and spy, but today he is mostly remembered for his five-part literary fiction about Gargantua and Pantagruel, a family of (literal) aristocratic giants who navigate life in 16th-century France while being much larger than everyone else. His second book in the series, Gargantua (1534), is centrally concerned with the advantages and costs of a humanist education. Those of his characters who receive humanist training produce distinctive and sophisticated language, but when they speak or write they are always primarily communicating what a fantastic education they’ve had, rather than whatever they actually have to say. When we read a sentence such as: ‘No more just cause for pain can arise among men than if, from where they should in all fairness expect gracious benevolence, they receive travail and injury’ (my own translation of Rabelais) – or take its modern chatbot equivalent: ‘It is profoundly disheartening when those entrusted with our care or respect instead subject us to mistreatment’ (GPT-4) – we are more likely to be impressed by the linguistic capacities of the writer or the language model that generated it, than to reflect on the idea that it’s bad when someone who should treat you well treats you poorly.

Erasmus and Nosoponus are both writing in Latin, which, like any humanist, they learned by laboriously imitating the great Latin writers of antiquity (despite the fundamental difficulties involved in trying to match their native fluency and eloquence in a second language). Renaissance intellectuals argued fiercely over whose Latin they ought to imitate, although they actually disagreed very little. Cicero was the chosen scribe; his standing as the best prose stylist of antiquity was uncontested. The main debate, therefore, was between the Ciceronians, who believed in imitating their hero to the exclusion of all other models, and the eclectics, who believed in imitating Cicero along with other models. Erasmus was an eclectic. Over the course of The Ciceronian, his authorial stand-in Bulephorus gradually wins over Nosoponus, the titular Ciceronian, to the eclectic side. To understand the arguments he makes, and the stakes involved in Nosoponus’s conversion, you need to understand how humanist stylistic imitation worked.

Humanists read and reread the classical authors whom they wanted to sound like – not unlike how an LLM trains on a corpus. They internalised the qualities of their models’ prose as much as possible, and also kept a sort of verbal rainy-day fund in something called a commonplace book, which is to say that, as they read, they transcribed a stockpile of salient words, metaphors, turns of phrase and clichés – usually organised thematically – that they would then draw upon while writing. An eclectic writer would gather the high points of many different authors and genres, but a Ciceronian would copy only from Cicero. Writing with a commonplace book, then, would enable an eclectic to draw on the greater range of his reading to make his own prose more versatile, while a Ciceronian’s style could become only more homogeneously Ciceronian. Bulephorus points out that Cicero himself read widely among his own contemporaries: if Cicero were alive today, he argues, he would produce texts tailored to the modern world, and therefore quite unlike anything in the surviving Ciceronian corpus. Cicero, in other words, would be an eclectic.

But The Ciceronian makes it clear that Ciceronians and eclectics do not just differ in the content of their commonplace books, but in how they use them. For eclectics, they are a kind of safety net: if Bulephorus can’t think of the right word, he can just consult his book, much as we might consult a thesaurus, saving himself all the time he would otherwise have spent trying to come up with it on his own. For a strict Ciceronian, however, a commonplace book is more like a security checkpoint. Nosoponus generates language without help, but feels compelled to laboriously verify the Ciceronian origin of every single word or construction that he wants to use in a commonplace book, which has, in effect, become an exhaustive concordance of the Ciceronian corpus. This insistence on maintaining conscious, executive control of the process of writing makes compositional fluidity impossible for him, to the point where he refuses to conjugate his verbs without first ascertaining whether Cicero used whichever particular inflected form he needs. No wonder a single sentence is a full night’s work.

The letter and speech sound good, but they are rather long and dull: there is something generic about them

Erasmus’s goal as a teacher was to train his students in just the sort of fluency that Nosoponus lacks. He did this by giving his students fictional or historical scenarios that required delicate rhetorical handling – in much the same way that we might prompt an LLM today – and having them compose letters or speeches ‘in character’. They might write a love-letter from Paris to Helen, for example, or they might write as Agamemnon, summoning his allies to make war on Troy, or else pleading with Menelaus to forget Helen and avoid the ghastly human cost of such a war. They could write as Menelaus vituperating Paris for stealing his wife, or perhaps forgiving Helen for leaving him. (Erasmus gives a slew of examples, including many of these, in his pedagogical manual On the Writing of Letters.)

When Rabelais’s humanist characters finally get the chance to put their training to practical use – attempting to avert a war with their neighbour, Picrochole – the texts they produce read as Erasmian schoolroom exercises. The giant Grandgousier sends his son Gargantua a letter summoning him back to defend the family land from his war-mongering neighbour Picrochole. At the same time, Grandgousier’s ambassador Ulrich Gallet also gives an anti-war oration urging Picrochole to desist from his aggression. Rabelais gives the text of both the letter and the speech in full. They sound good, but they are rather long and rather dull: there is something generic about them. They could just as easily have been produced by Renaissance schoolboys prompted with these scenarios, or by LLMs simulating them.

They also don’t work. Grandgousier and Gallet may decry the horrors of war, insisting on the moral obligation to avoid it, but both the speech and the letter actually accelerate the violence rather than preventing it. They fail to make peace because they are generic.

The automation of language, whether by Erasmians or by LLMs, depends on rejecting novelty: both work in identical fashion by decomposing apparently new situations and topics into familiar elements, so that those situations can be addressed with language that is already associated with those elements in the training corpus. What this means for Grandgousier and Gallet is that the humanist mindset that enables them to speak so well also makes them approach the conflict with a certain arrogance – with the assumption that they can anticipate anything the other side might conceivably have to say on the basis of what they have already read.

Indeed, the irony of their predicament is that Rabelais’s humanist characters harp on the necessity of establishing a dialogue with Picrochole, but prevent it happening with the very language in which they express the importance of hearing his perspective. This is language not as an olive branch but as a shield.

Grandgousier writes to Gargantua saying: ‘I have amiably sent to [Picrochole] multiple times, to understand in what, by whom, and in what way he felt he had been done wrong, but I have had no response from him but wilful defiance.’ But this isn’t true. Grandgousier has sent his ambassador to Picrochole only once, not multiple times – and recently enough that Gallet could not possibly have reached him already, let alone returned with a response.

The reason ‘unicorn’ and ‘raccoon’ appear in different contexts is that raccoons actually exist

Maybe Grandgousier is lying because he has no real interest in averting the war, and the letter sets up a paper trail to establish the geopolitical moral high ground. But what if we don’t think about the false claims he makes as lies, but as something closer to the so-called ‘hallucinations’ of large language models?

An LLM responds to a prompt with text that it ‘considers’ likely or appropriate in the context established by that prompt. Because it is so sensitive to context, it will tell me about unicorns if I ask for a fairy tale, but not if I ask about North American wildlife – but that does not mean that it has an internal representation of reality or of what it means for something to be real (although, of course, when prompted it can generate text describing such a representation). It does not ‘understand’ that the reason words like ‘unicorn’ and ‘raccoon’ appear in different sorts of contexts is that raccoons actually exist. This is fine as long as we ask LLMs for information that is common and consistent – like the number of teaspoons in a tablespoon. But when we ask for something obscure, or highly nuanced, they may just make up something that looks like a plausible answer – such as a fake legal precedent (where LLM performance is no better than random guesswork), or an NBA line-up with players from the wrong team.

Something similar might be going on with Grandgousier, who might inhabit the role of a good Christian-humanist prince in the same way that an LLM inhabits the persona its user assigns it, as a cluster of discursive and rhetorical features. After all, he is navigating a scenario straight out of Erasmian moral instruction: we know that he ‘should’ stop at nothing to avoid the war because we have read Erasmus’s extensive anti-war writings. And, indeed, his letter says everything that we, or he, might generically expect to hear from a virtuous prince in his situation (or from an Erasmian schoolboy imagining himself as one). And so, while Grandgousier is engaged in the highly trained and semi-automatic process of composition, he might write untruths out of a sort of discursive reflex – operating not contrary to but without reference to the truth – again like an LLM, which only approximates reality by saying the sorts of things that, based on its training, it expects its users will expect it to say.

Recall Erasmus’s argument against Ciceronianism, which is that no single author can offer the generic and linguistic resources for every possible discursive situation. But the difference between Erasmus’s eclecticism and the strict Ciceronianism he lampoons is one of degree rather than kind. On both models, the humanist can accommodate novelty only insofar as the tools he draws from a predefined corpus allow him to. But because the eclectic corpus is larger and more diverse than the Ciceronian one, its limitations are more subtle – and LLMs are more subtle still, since they are trained on corpora drastically larger than either. And yet Rabelais’s humanists display the very same solipsism as LLM users today, who, in thinking ‘with’ LLMs, merely refract their own thoughts through a body of other people’s words.

There is nothing wrong with this, but it is not communication.

Both the letter and the diplomatic speech are by definition communicative genres: they exist to produce an immediate effect on the particular other to whom they are addressed. The letters Erasmus’s students write are, by contrast, pedagogical exercises, which only the instructor ever reads. Gallet’s diplomatic mission fails because he treats it as an exercise. Although he is sent to get Picrochole’s side of the story, he merely simulates an attempt to do so through a series of rhetorical questions: ‘So what frenzy moves you now … ? Where is faith? Where is law? Where is rationality? Where is humanity? Where is fear of God?’ Somewhat ironically, he later reprimands Picrochole for failing to mitigate conflict through dialogue, as he is himself failing to do: ‘If some wrong had perhaps been done by us … you ought first to have enquired after the truth of it, and next rebuked us for it.’

This brings us to the another consequence of the solipsism of autonomous language, which is the degeneration of what the British philosopher J L Austin in the 1950s called its illocutionary force – that is, what it means for someone to have said something, as opposed to the (locutionary) meaning of what they actually say. The illocutionary function of an LLM’s language is compromised, most proximately, because there is no singular, social agent speaking: an LLM can’t get married, make a bet, or christen a ship (to use Austin’s examples). But even when one person uses an LLM to generate language to send to another person, there is a loss of social ownership over the language in question – which, much like an ambassador, is sent to address someone on someone else’s behalf.

Gallet sets his discourse in motion without considering its pragmatic effect upon its audience

Gallet’s embassy fails because he is blind to the pragmatic dimension of his language. His speech is wildly aggressive and insulting. He describes Grandgousier’s prior alliance with Picrochole as a ‘sacred friendship’; tells Picrochole that his military aggression demonstrates ‘that nothing is holy or sacred to those who have emancipated themselves from god and from reason, in order to follow their perverse inclinations’; says that his conduct has been ‘so far beyond the bounds of reason, so repulsive to common feeling, that it can barely be comprehended by human understanding.’ And yet, if we consider Gallet’s embassy as a social action – which is clearly how the deeply offended Picrochole takes it – it unambiguously signals a desire for war, even though its denotative content is pacifist. We can take this tonal mismatch it in two different ways. It could be a devious attempt on Gallet’s part to escalate the war while retaining the optics of victimhood. But then Gallet may also simply be failing to consider the illocutionary force of his speech, which he treats as a generic rhetorical exercise on the theme of the evils of war, rather than as communicative diplomatic action.

There is a simple ethical lesson to be drawn from all of this. If we wish to encounter – and make peace with the ‘other’ – we must avoid the example of Gallet, who sets his discourse in motion without considering its pragmatic effect upon its audience. We should, instead, imitate Picrochole himself, who greets Gallet not with a speech but with a question: ‘What have you got to say?’

In structuring the outbreak of the war around a series of humanist failures to ask, and to listen, Rabelais at once characterises those failures as moral, and locates them in the cognitive structures that delimit Erasmian rhetorical training. The accusation is this: that Erasmus gives his students a technology for producing language as an end in itself, but he doesn’t teach them how to communicate. Rabelais shows us that, when the production of discourse is automated, it becomes strictly monologic and loses its illocutionary social power. This sort of autonomous language is just like an ambassador: it speaks for us, but it cannot speak as us.