New APPS

By Gordon Hull

For quite a while, I’ve been exploring how to relate Derridean concerns about language and the politics behind theories of language (and text), and how to think about those in the context of large language models (part one, two, three, four, five, six).  Last time, I talked about subjectivity and the question of whether a speaking subject necessarily subtends language production and how that might play out in the context of language models.  Here I want to return the conversation to Plato, since Derrida’s discussion of Plato’s critique of writing in the Phaedrus is central to all of these thoughts.  One thing that emerges in Plato is that he is acutely aware of the political stakes involved, and the Phaedrus deploys various strategies with myths and stories to communicate those stakes.  Indeed, Plato does this basically all the time.

Most obviously, in the Republic, when he infamously banishes the poets, it is not all poets he banishes, but the Homeric ones.  The problem is that Homeric poetry teaches the wrong thing.  Socrates broaches the subject as follows:

“We must begin, then, it seems, by a censorship over our storymakers, and what they do well we must pass and what not, reject.  And the stories on the accepted list we will include nurses and mothers to tell to the children and so shape their souls by these stories far rather than their bodies by their hands.  But most of the stories they now tell we must reject” (377c).

He then immediately cites the example of the Uranus Kronos story told by Hesiod, which he avers shouldn’t be told to young people even if true (I’ll return to this point, as I think it’s important).  Instead, for stories like that, “the best way would be to bury them in silence” (378a).  This is because “the young are not able to distinguish what is and what is not allegory, but whatever opinions are taken into the mind at that age are wont to prove indelible and unalterable” (378d).  As Penelope Murray comments of this passage, Plato “is not concerned with the factual veracity of history here, but with the ethical truth that should be expressed through myth” (252).  The problem is that the wrong myths have been told, and numerous examples of allowed and disallowed myths follow. For example, we must say that the Gods do not deceive, because words are a “copy of the affection in the soul” (382b) and “essential falsehood … is hated not only by gods but by men.”

Murray says of the entire exercise:

“The difference between Plato’s noble lie and the false myths of the poets rests in the fact that the former is specifically designed to foster noble ends: patriotism, brotherly love, and social cohesion.  As a charter myth it also legitimizes the practice of promoting or demoting citizens to their appropriate class.  The story with all its circumstantial detail is avowedly false; what matters is the moral and social purpose which the myth is designed to achieve, and it is this which makes the difference between a pseudo which is good or fine and one which is not” (254).

As Stephen Halliwell points out, the poets Plato is demoting were cultural authorities whose moral statements in particular were formative in the cultural norms of the Greeks.  Plato thinks that many of them are destructive and so has to combat them; the ideal city will only contain helpful myths.  Halliwell puts it this way:

“The logos of poetry, on this view, is discourse for which the poet’s authorial voice, a voice that is granted cultural authority to ‘speak’ on the most important matters of life, must be held responsible and subject to the ethical interrogation of philosophical inquiry” (104).

This is not to say that Plato doesn’t have his preferences; as Halliwell adds, with reference to the Phaedrus, plain logos is better than poetry because you can’t query the poets about the meaning of their works (106).  The logoi of poetry is thus deficient because it’s text, and “the passage does appear to raise grave uncertainty about the value of poetic citation for the exploration and formation of ethical principles” (107).  But poetry texts aren’t going anywhere either, and so Plato has to figure out how to absorb and refashion them to suit non-destructive ends.

Claudia Baracchi persuasively argues that the entire Republic is oriented around exercises in myth-making. How else can we read the enigmatic Book X?  The book begins by reprising the denunciation of the poets, who this time are flagged as “imitators.” Socrates summarizes that “the imitator knows nothing worth mentioning about what he imitates; imitation is a kind of play and not serious; and those who take up tragic poetry in iambics and epics are imitators in the highest possible degree” (602b).  This is bad!

“The imitative poet produces a bad regime in the soul of each private man by making phantoms that are very far removed from the truth and by gratifying the soul’s foolish part, which doesn’t distinguish big from little, but believes the same things are at one time big and at another little” (605b-c).

In short, poets are bullshitting.

The Book closes with the Myth of Er, which in this context can only be read as an attempt to replace the stories of the poets with a properly pedagogical exercise in mythmaking.  Here is Baracchi:

“In disclosing the circular motions of the souls (moving back and forth from life to death and back to life), the muthos articulates the order (justice, law) and once governing the psychic migration (i.e., informing the ethos of the souls in death and life) and presiding over the becoming of the cosmos.  This vision inscribes ethical necessity within the sphere of cosmic (natural) necessity, thus showing them in their essential conjunction. Necessity itself is brought to the fore, as it were – exposed in its all-encompassing centrality” (189).

Thus, “in the course of the dialogue it becomes evident that the issues at stake cannot be adequately dealt with on a purely logical basis and that what is accomplished in logos must come to terms with an order of necessity that here was called extra-logical or dia-logical” (220).

Plato is not bullshitting, and he is not indifferent to truth, but he has also subordinated questions of truth to questions of ethos, and (this is the point of the Hesiod Uranus example) the ethos question is dispositive for the truth one. As Daniel Tigard points out in the context of LLMs, when we talk about bullshitting, we need to distinguish between the act of bullshitting and the bullshitter.  Plato may be doing the former in the sense that he is using stories (“confabulations”), but he is definitely not one of the latter.  At the same time, the character of Socrates shows the difficulty in sustaining this project.  Socrates is complex and embodies all the roles and characters he plays in the dialogues; as Arlene Saxonhouse puts it in a slightly different context, “Socrates becomes the countermodel of the regime he proposes” to the extent that “his speech succeeds precisely because he can imitate all, because he can be multifaceted.  The Socrates as written by Plato would exile himself” (741).

All of these origin stories and lessons are confabulations, and you can’t avoid them.  Some confabulations are better than others; some, like the Hesiod story, are so bad that you shouldn’t tell them even if they turn out to be true!  As a recent paper by Peiqi Sui, Eamon Duede, Sophie Wu, and Richard Jean So points out, confabulation is a common narrative strategy by which people make sense of their world, come to understand complex things, and fill in gaps in their understanding with plausible (but unverified) details.  Language models do this too.  In a sense they are nothing but confabulators in that everything that goes into them is text and what comes out is a statistical sense-making.

Plato’s judgement is harsh; recall the passage I cited above: “the imitator knows nothing worth mentioning about what he imitates; imitation is a kind of play and not serious” (602b).  As Derrida says of Plato more generally, “the best sense of play is play that is supervised and contained within the safeguards of ethics and politics” (“Plato’s Pharmacy,” 156).  He then immediately cites a passage in the Laws to illustrate the “progressive neutralization of the singularity of play” (157; emphasis original).  The singular details of the poets have to be regularized by their attachment to principles, and those that cannot be so attached need to be discarded.

But this also can’t be the end of the Platonic story, because Socrates also has to tell stories.  It's not just in the Republic, either.  In the Phaedrus, for example, the myth of Theuth is deliberately invoked at a strategic point in the dialogue.  So too, the Phaedrus makes use of textuality quite explicitly.  Socrates doesn’t just want Phaedrus’ report on Lysias speech.  After an exchange in which Socrates insists that Phaedrus must have memorized the speech (228a-b) and Phaedrus responds that “I did not memorize the speech word for word” and will have to summarize it instead (228d), Socrates explicitly refuses the offer:

“I strongly suspect that you have the speech itself. And if I’m right, you can be sure that, though I love you dearly, I’ll never, as long as Lysias himself is present, allow you to practice your own speechmaking on me” (228d-e).

Socrates is here demanding more or less the opposite of what we’re taught the Platonic critique says.  Here, Socrates wants to have Lysias’ words even though Lysias can’t be here to defend them or answer questions about them.  He even allows for the recitation of Lysias' text to mean that Lysias "himself is present."  Moreover, when they come to discuss Lysias’ rhetoric, Socrates says again, “read me the beginning of Lysias’ speech” (262d); again, later: “read it, so that I can hear it in his own words” (263e).  In short, we should take Socrates quite seriously when he begins the discussion of rhetoric by proposing that “this, then, is quite clear: Writing speeches is not in itself a shameful thing …. It’s not speaking or writing well that’s shameful; what’s really shameful is to engage in either of them shamefully or badly” (258d).

The distinction between confabulation in the form of productive myth-telling and bad confabulation in the sense of what Lysias does – in this case, present a poorly constructed speech designed to prove that it is better to be seduced by someone who does not care about you – is thus located in what the speaker is trying to do and the political implications of the speech. 

Next time I’ll get back to Derrida and language models.