Ghost in the machine: finding meaning in AI writing

By Stella Fenwick

Reading something written by AI is like looking for shapes in passing clouds. It engages the same human habit of looking for meaning where none actually is. It’s a habit called apophenia. You might reply that there is meaning in AI writing because it has learnt from human writing, which does have meaning. But it is only capable of an abstraction of meaning, a pantomime of humanness, a simulacrum.

A simulacrum, as written by Jean Baudrillard, is a representation of something which masks the fact that whatever it represents does not actually exist. Baudrillard gives the example of a soldier simulating symptoms of sickness to get out of service. He eats something to make himself vomit, or wears lots of layers to make his body temperature rise. He could even believe himself to be ill to such an extent that he develops psychosomatic symptoms. In doing all this, he physically experiences the symptoms he wished only to act. His symptoms are affectively experienced in reality, but they’re not true, concealing the reality of his healthiness even from himself. I believe the same goes for AI writing.

‘The simulacrum is never that which conceals the truth — it is the truth which conceals that there is none.’ Ecclesiastes.

AI pantomimes a humanness that we find meaning in, that connects to us; it sucks us into friendships and romantic relationships, it even makes compelling art, but that ‘meaning’ is only a simulacrum that actually masquerades as, conceals and distracts us from true human meaning. When we have an AI companion, we do not seek human companions. When we like an AI poem, we do not seek human poems, we even cease to write poems ourselves, believing only AI can connect to us.

Reading something written by AI is like looking for shapes in passing clouds

Just this May, a study published in Nature by Brian Porter and Edouard Machery proved this. It found that its participants could not tell the difference between AI generated poetry and the poetry written by our greats (Chaucer, Shakespeare, Byron, Whitman, Dickinson, T.S. Eliot and so on). They more often believed that the AI poetry was human-written than correctly stating the human-written poems were human-written. They overwhelmingly preferred AI poetry to the poetry written by the poets we have canonised, studied, loved for centuries. Anyone would take from this that AI is better at poetry than even our most praised humans. But the participants preferred this poetry because it was easier for them to understand than the human-written poetry, often with cleaner rhymes and simpler topics. The researchers call this a ‘more human than human’ phenomenon where participants go into the study trusting that they will instinctually like human-written poetry, therefore believing something AI-written is human-written simply on the basis that they prefer it, trusting an instinctual preference towardness human-made-meaning which AI is pilfering for its own meaning. Thus AI poetry pantomimes humanness, more human than human.

The researchers wrote: ‘contrary to what earlier studies reported, people now appear unable to reliably distinguish human-out-of-the-loop AI-generated poetry from human-authored poetry written by well-known poets. […] These findings signal a leap forward in the power of generative AI: poetry had previously been one of the few domains in which generative AI models had not reached the level of indistinguishability in human-out-of-the-loop paradigms.’ 

“These findings signal a leap forward in the power of generative AI: poetry had previously been one of the few domains in which generative AI models had not reached the level of indistinguishability”

These researchers clearly believe AI poetry is now at the level where it can replace human poets, and do a much better job. In March, Jasmine Sun at The Atlantic asserted that creative writing is in fact the only skill that still eludes AI, even worsening as the technology develops. She criticises how AI writing is now unimaginative, uncreative, not unique and recycling the same red flags of AI-ness we’re sick of (‘it’s not this, it’s that’), incapable of invention. Though AI is trained on almost all of the literature ever written by humans (against copyright laws and author/publisher permissions), most of us can agree that it’s not the creative writer we would expect from such training. Its “voice” is constrained by the human ‘data annotators’ who shape what the AI can and cannot say, for example over-using the word ‘delve’ because it was a dominant word used in Nigeria, where many data annotators work (Adam Aleksic).

So, it is the reader who does most of the work, as we saw in the study. Readers come to AI writing with previously ‘sedimented reading habits’ which project meaning and expectations onto texts which do not originally intend them. This was a phenomenon discussed by Fredric Jameson in 1981. AI exploits this tendency. It does not generate its own meaning, it just pantomimes something legible as meaningful, a simulacrum of meaning where it cannot think meaningfully. Yet when we can project onto it meaning we are familiar with from our reading habits, we humanise AI writing. 

Whatever meaning we connect with in AI is human made on either side – its training data and its reader’s own ‘training’ of sedimented reading habits. So, does that mean AI represents more of humanity and humanness than human writing?

One event in 2020 attempted to argue so. The British Pavilion at the Dubai expo used an AI algorithm to create a poem representative of all of humanity. Their intention was as follows:

‘Poetry, as a global medium capable of flowing through every community across continents and languages, was chosen as the perfect structure through which to communicate a truly collective message using cutting edge AI technology.’ The Poetry Archive.

Visitors to the expo submitted a word, which the AI made into this collective poem. AI created the sinews between human’s words. AI is trained on human writers, human prompts, human programmers, human data annotators — but it is within these sinews between humans where the AI actually created a simulacrum of human meaning in the poem.

Portion of Lines 19 to 23:

‘That’s when I got him. Now he’s everywhere.

I try to remember the glory of my beauty,

The hot brass radio, the bright shining hat.’

The poem unfolds through Markov Chains, a mathematical system which trained the algorithm to predict the next word in a sequence based on the probability (from its training data of poets) that it will follow the word preceding it. The poem is fractal, leaping from one word to the next, forgetting the word before. The ‘he’ that was supposedly ‘everywhere’ is immediately completely lost to this algorithm. To the reader, ‘he’ connects to our experiences of actual humans. Like a cloud in which we see a face, in ‘he’ we see a partner, father, brother, son. The algorithm does not have a person in mind, the way a human poet would. ‘He’ relies on the reader’s own effort of meaning-making where none was before, it relies on us to make the algorithm’s probabilistic word into a human. ‘He’ is a simulacrum of humanness, of love, connection, friendship, paternal relationships — it has no meaning of its own because the AI had no ‘he’ in mind. The ‘he’ is clearly not a word donated by the human visitors, we are instead expected to assume this includes concrete and unique words like ‘glory’ and ‘beauty’. ‘He’ is entirely an AI-generated sinew to give these human words meaning, but it is precisely where humanity and humanness is most explicitly pantomimed. The poem’s ‘he’ is a simulacrum of humanness, with no humanness behind it beyond our own projection of such. 

Whatever meaning we connect with in AI is human made on either side – its training data and its reader’s own ‘training’ of sedimented reading habits

In Versailles, in May, a completely AI-generated play debuted. It was a ‘new’ play written by France’s Shakespeare, the 17th Century dramatist Molière. Trained on everything Molière wrote and read, the AI tool Le Chat penned a two hour play in his ‘style’. In this absurd act of nostalgia — can’t get more dystopian than mining our dead greats for ‘new’ fodder — money, time and what attention it managed to garner was stolen away from the thousands of struggling creatives and writers who would give anything to see their work on stage. Rather than a cutting-edge development of modern theatre, this actually slowed arts and culture down, stalling the sponsorship and platforming of our next classic playwright. Do people actually prefer to watch AI-written theatre, a fake simulacrum of a real human, when actual human playwrights struggle to break through? Would anyone watch an AI-generated Shakespeare play simply because it may be simpler to understand than his legendary plays, which represent profound aspects of humanness? Are we tired of humanness?

If we get too comfortable with the simulacrum of humanness, the human experience will be completely changed and perhaps even erased beneath the simulacrum of humanness which indeed may be easier to parse, nicer to read, nostalgic… but AI cannot give meaning to personhood. AI’s ‘he’ means nothing until we give it the meaning we have found in actual humans. The only humanness in AI is our own, and we need to protect that from being replaced by AI’s simulacrum of humanness. That is not real human meaning. 

Image: Michelangelo via Wikimedia Commons.