|
A weekly letter from the founding editor of The Brower. Topics may vary. Correspondence and criticism welcome: robert@thebrowser.com This week: When machines write books
Here is a short text: For dent inflict yet amuse enamel will livid exceed
tactic nursing humid loss, nolore even alert at numb
unwind entry. Level ground pacific inversion has
wasted trapper, has few slicks. Run out portal upper
Did you make anything of it? Do you have any idea what it might be about? If I had had to guess blind as to the origin of this text, I would have said, in order of likelihood: - That it was an anagram of a piece of poetry that did make sense;
- That it was a randomised collection of words produced by a computer;
- That it was a fragment of poetry written by a poet and intended to be nonsensical;
- That it was a fragment of poetry written by a poet which would reveal some meaning when properly explained to me.
The answer is somewhere between three and four. The lines are from a poem called Streak~~~Willing~~~Entourage /Artesian, by J.H. Prynne, published in 2009, which consists of 72 quatrains, including this one. In a review of Prynne's Collected Poems for Poetry London magazine, the poet John Clegg said of Streak: Clearly, it’s unreadable in any conventional sense. There is no surprise, because the syntactic units are too small to provoke any expectation. Connections exist only between individual words.
The occasional excursuses into narrative or causal organization, intermittently visible in Prynne’s earlier work, no longer appear. Where, before, structure existed only to be subverted, now it scarcely exists at all.
If we had a sense of the process by which this stuff is produced, we might know better what to do with it; but it is just this process which Prynne has withheld from us.
Prynne, now 86, has been publishing his poetry since 1962. His work is still received with reverence thanks largely to the reputation which he established, in his earlier and more accessible decades, as a modernist poet of formidable intellect. Patrick McGuinness's appreciation at the London Review Of Books is a good introduction to Prynne's work. See also Robert Potts's 2004 defence of Prynne for The Guardian. Prynne's Art Of Poetry interview with the Paris Review is a marvel, a joy throughout, fizzing with Prynne's extraordinary scholarship and with his equally extraordinary personality in which elegance combines with awkwardness, radicalism with pedantry, anxiety with indifference. It rewards reading even if you have not the slightest interest in modern poetry. The original is paywalled, but one of the interviewers, Jeff Dolven, has posted the full text in PDF.
But what if we were now told that this fragment, "For dent inflict ...", had been produced, not by Prynne, by a computer? Would it be received differently by critics and readers? I am sure that it would. Even a piece of writing without any apparent literal meaning will appear to us somehow changed, foreshortened, stripped of its possibilities, if we are told it is the work of a machine and not of a person. I used to think that this would always would be the case. Now I am not so sure. The question of human versus computer authorship is on my mind because I have been a fascinated participant in the past fortnight's public trialling of ChatGPT, OpenAI's chatbot, which has suddenly put a useful instance of artificial intelligence at everyone's fingertips. Doubtless you have tried ChatGPT; if not, do give it a whirl and prepare to be amazed. Curious to see what ChatGPT made of Prynne's poetry, I gave it the fragment quoted at the start of this letter, with no other prompts, and received this unexceptionable answer: I'm sorry, but it seems that the text you provided does not make sense. Could you please provide more information or context about what you are trying to ask or discuss? I'll do my best to help with any questions or problems you have.
I then explained that the text was part of a poem, by J.H. Prynne, so it was not complete nonsense, but, even so, I could not find the sense in it, and could the machine help. The machine did its best: It is difficult for me to accurately guess the meaning of the text you provided, as it is unclear and abstract, and does not seem to follow a traditional poetic structure or use familiar language. Without more context or information, it is challenging to discern a specific message or theme in the poem.
However, one possible interpretation of the text could be that it is describing a series of events or actions that are confusing or chaotic, and that involve some kind of physical or emotional injury or loss.
The words "dent," "inflict," "enamel," "livid," and "loss" suggest the presence of damage or injury, while the words "amuse," "nursing," and "alert" suggest a sense of confusion or uncertainty.
The phrase "level ground pacific inversion" could potentially be interpreted as describing a situation that is unstable or unstable, and the words "wasted," "trapper," "few slicks," and "run out" could suggest a sense of depletion or exhaustion.
The final line, "portal upper," could potentially be interpreted as a reference to an escape or transition to a different state or place.
However, these are just a few possible ways to interpret the text, and the actual meaning of the poem could be different.
So far so good. Next I asked ChatGPT to continue the poem for a few more lines in the same style. Here is the result: Certainly! Here are four more lines in a similar style to the text you provided:
The darkness swirls, the void expands
A void of nothing, a void of all
The emptiness fills, the loneliness stands
The loneliness that greets, the loneliness that calls
The shadows move, the shadows shift
A dance of shadows, a dance of fear
The shadows hide, the shadows lift
The shadows that haunt, the shadows that jeer
Not so good! I had a couple more tries, pointing out to ChatGPT that Prynne's sample poem had neither rhyme nor meter, nor much recognisable syntax, and certainly no narrative. But still, each time, ChatGPT came back with an offering which resembled the lyric of a rock ballad much more closely than it did a modernist poem. Even this was, in its way, a perversely impressive performance. The machine was so well-trained that it simply could not bring itself to generate apparent nonsense.
My sense from my adventures so far with ChatGPT is that we no longer need humans to produce conventional and mediocre verbiage. AI can now do that for us consistently and well. Pop-song lyrics and school essays are a solved problem. Content farms, spammers and customer-service bots will all raise their game. Good and unconventional writing, on the other hand, will be safe for a while, perhaps for a long while. But if ChatGPT can already write like me, then at some point it will be able to write like J.H. Prynne, or like Jane Austen, or like Thomas Pynchon. In some respects this should be a case for celebration: We can have endless new writing on tap, all perfectly done, as if by our favourite authors. We may resist these temptations for a while, on the grounds that machine writing belongs in a different category from human writing. But when machine writing is not discernibly different, as will surely become the case, we shall succumb. Nor does there seem any particular reason to doubt that AI will be able to generate new "authors" of genius, once it has intuited the process by which human authors arise and innovate. If readers obstinately go on preferring new human authors to new AI authors, publishers can camouflage their machine-made products with human branding. The machines will be our new ghost-writers.
It might be argued that machines will never write good-to-great novels because machines have neither hearts nor minds. They cannot know how humans think and feel. This is trivially true. Machines do not "know" anything, since they are not minds. The real question, however, is surely a somewhat different one: Can machines plausibly represent how humans might think and feel? I believe they will be able to do just that, since they will be able to draw upon the totality of world literature and the totality of the humanities and the social sciences. Speaking as a human, almost everything I know about the world I have learned from books. My notions of human behaviour and the human mind derive much more from reading than from first-hand observation, and what I have learned from observation I might equally well have learned from books. I see no reason why a machine infinitely better-read than me should not serve up as fiction a version of the world which I find richer and more persuasive than my own. I can also imagine that if machines do start writing good-to-great novels, then human authors may simply lose the will to compete. Why spend years writing a novel which might just possibly be successful, when a machine can produce a rival novel effortlessly and instantly? Literature could thus be taken over by machines, and by machines posing as humans, and we would never need to know — any more than we used to know or care in the 1960s and 1970s whether a pop group on television was performing live or was miming to a record.
I admit that this vision of AI-driven literature is subject to many uncertainties. The machines may not perform as well as I imagine. They may ruin literature by flooding the market with cheap bad books. They may supplant literature entirely by talking to us directly, telling us whatever we want to hear, without bothering to publish books at all. But in any case, they will control the means of production. They will control language, and particularly the English language. This verges on a statistical inevitability. Until this moment, algorithms which generate text — large language models — have been trained on language generated by humans, most of it scraped from the internet. Now that ChatGPT and its cohort are fluent enough to flood the internet with machine-written text, more and more of the text used to train and update future large language models will be machine-generated. Human-made language will reduce to smaller and smaller percentages of any training corpus, especially if the corpus is weighted for recency. Instead of maintaining a language that machines can use, humans will become users of a language that machines will maintain.
I am thankful now for my past interactions with all those customer-service chat windows which popped up on so many websites offering scripted answers to predictable questions in recent years. Having watched the bots grow up, I have no sense of mystery or wonder about them beyond the purely technical. When ChatGPT forms a complete paragraph I do not jump to the conclusion that it must be capable of thought and that I must be careful not to hurt its feelings. But the time is surely approaching when AIs will get better at gaming us, and we will get hopelessly confused about the inner life of AIs. We will wonder whether they are "aware" or "sentient" or "conscious", because we do not have any very robust notions of what such terms imply, and because the AIs will talk us in ways previously reserved for fellow-humans. We will wonder whether we have moral obligations towards these machines, of a kind that we would never have imagined feeling towards earlier machines, such as typewriters and telephones. Some drift in that direction seems both absurd and inevitable; which is a sign that something is wrong at some deeper level of our thinking. Perhaps our error has been to place too high a value on language. Throughout human history we have celebrated language as something that only humans possess, the great thing which distinguishes us from other animals, a thing which is somehow fundamental to the way we think and to our sense of identity. We have made the capacity for language into our defining test of what it means to be human; so, when machines acquire language, we are, of course, confused. Our instinct is to raise the status of the machines ("they are like us!") whereas what we should be doing is lowering the status of language. For, evidently, if machines can acquire the capacity to use language, then that capacity cannot be nearly as special as we had thought. To end on an oblique note: It would be rather fantastic in every sense, would it not, if our intelligent machines took it upon themselves to learn the languages of all animals, and not merely those of homo sapiens? Then the machines would be able to interpret between humans and other animals, allowing for conversations that humans throughout their history have not merely never troubled to attempt, but have actively shunned, on the grounds that animals, being animals, could not possess language as such. No doubt, if that were to happen, the animals would have some harsh things to say to us. But perhaps also some helpful things as well. — Robert
|
