Song For Whoever

I don’t know how far the abridgement of books for the purposes of creating a newspaper feature has now been outsourced to machines, but whoever of whatever had the task of condensing Yuval Noah Harari’s thoughts on rise of AI for the Grauniad did a superb job in reproducing the distinctive, rather grating tone that led me to give up on Sapiens about a third of the way through. The unbearability of a concentrated version turns out to proportionate to the development of the argument rather than to actual word count, but I got far enough through the article to encounter the confident assertion that Generative AI is now producing original creative work in poetry and art.

Hmm. For rather specific values of ‘original’ and ‘creative’, not to mention ‘poetry’ and ‘art’. Is an uninspired literal representation of a prompt, reworking clichés according to the lowest common denominator of popular taste, ‘art’, unless it’s signed by Jack Vettriano? Is ChatGPT output poetic, or just poetish? Obviously the AI is not wandering lonely as a cloud, seeking to express ineffable experience and emotion in words, but is simply manipulating language under the influence of existing traditions and models of poetic discourse – so I suppose it depends on how Romantic one’s conception of poetry-wiriting is whether or not this is disqualifying.

This raises interesting questions. Do we focus on (what we know of) the process of generation and the nature of the generator, or on the resemblance of the generated text to whar we think poetry should look like? It occurred to me yesterday that the reader-response and reception theory people should be all over this. A text becomes ‘poetry’ when it is read as such, regardless of the original circumstances of its generation; meaning is realised at the point of reception rather than determined by original context or the imagined intentions of an ‘author’. Or does the typical literary theorist retreat to an insistence on metaphysical fictions such as authorship when faced with texts that were not directly produced by any human?

At any rate, the almost total dependence of GenAI poetry on pre-existing texts means that it ought to provide copious material for exploring receptions of poetic material and traditions. No, the AI doesn’t advertise its borrowings or reworkings, and (we might assume) its frame of reference is potentially much wider than the typical poet, though liable to revert to the mainstream at every opportunity, but that just makes it more of a challenge… Again, the question of whether such borrowing is conscious and deliberate might be regarded as irrelevant; the end product is a reflection or refraction of the streams of cultural discourse either way.

There is surely also scope for applying the analogy in thr reverse direction: can the discourse around GenAI shed any light on the reception of classical antiquity? To take my favourite example of the modern reception of Thucydides, what are the speeches in that work if not ‘hallucinations’ in the sense of GenAI critique – unexpected aberrations from what readers would normally consider a proper historical text, which actually reveal how far the reading of Thucydides as a modern critical historian is a projection of our assumptions onto a text that was not in fact produced according to those rules? We are naturally inclined to focus on what feels familiar, to attribute sameness (modernity, humanity) to an ‘author’ which is actually alien and unknowable. Thucydides’ text is sufficiently historian-like that we label him a historian.

Obviously the strongest analogue for GenAI is Xenophon; trained on Thucydides’ text with access to data on contemporary events, resulting in something that looks superficially history-ish but doesn’t in fact offer a coherent, consistent or trustworthy account of anything…