The Echoborg is a Catfish.

Earlier this month, the New York Times conducted a little experiment.

They provided a series of writing samples in various styles and genres (Literary Fiction; Historical Fiction; Fantasy; Poetry etc) and presented readers with two extracts in each category. One of the two passages (we weren't told which one) was written by a human, while the other was generated by A.I. We were asked to choose our preference.

Crucially, the test did not ask us to guess which one was human; it simply asked us to judge which was the better piece of writing.

The response from readers was fascinating, to say the least. Many commenters angrily denounced the experiment as "unfair" or "meaningless".

"I don't really understand the point here." grumbled one contributor. "It asked me which I preferred. It didn't ask me 'which one is the human'."

Others were angry with themselves because of the choices they had made.

"This is uncanny and downright scary. I deliberately set out to discern and choose the human passages as the ones I “liked better.” I thought for sure I could pick out the human voices. I unwittingly chose the AI in 4 out of the 5."

One responder may have unintentionally put his/her finger on the issue with a comment that revealed far more than intended:

"I am a bit spooked [and] at the same time fascinated by my selecting AI 4 ot [sic] of 5 times.   As a former English teacher, I wanted to choose the human writing, but I was clearly taken [in] by the clarity and style of the AI selections."

(Pro tip: if you really want to spot the human, don't look for "clarity and style"; look for the awkward phrasing and the typos. Those are sure signs of the biological; even from former English teachers...)

Of course this experiment would indeed be limited and rather imperfect, had it been truly attempting to demonstrate anything about A.I. Placing a ChatBot-generated passage alongside an arbitrarily selected piece of text written by a specific, individual human doesn't necessarily mean you prefer ChatBots in general; it just means that you might not be a fan of, say, Cormac McCarthy or Elizabeth Bishop (two of the human writers chosen for this little endeavour). There have been roughly 115 billion humans over the last 100,000 years or so (when language is thought to have developed) and it's safe to say that some of those humans were better with words than others.

"To read, makes our speaking English good." 

Xander Harris; in Buffy, the Vampire Slayer. 

(A ChatBot probably wouldn't have put it quite that way.)

But the New York Times experiment wasn't really about the ChatBots at all. It was about forcing us (the designated humans in the room) to take a hard look at our own relationships with words.

If you follow any current discussions of A.I., either in the media or in popular culture, you will almost certainly have come across the term "Slop". Slop has become the go-to derogatory term for the recent explosion in A.I.-generated content (presumably of low quality) that has been overloading your social media feeds, messaging inboxes and (in some cases) respected news publications.

In the last few years, almost all A.I. generated content has been dismissed as Slop, but (as many New York Times readers discovered to their discomfort) the situation is a little more complex. ChatBots are getting very good at generating high-quality content, and (in some cases) are starting to use language better than what we do, even.

The Turing Test (famously proposed by the computing pioneer Alan Turing) was intended to determine whether a machine could successfully pass for human (ChatGPT has passed that test, by the way). The little New York Times experiment was asking a slightly different question: can a machine generate content that is actually preferable to the content we generate ourselves? Can ChatBots be better writers than we are?

Spot the human...

In Dark Star, Lieutenant Doolittle spends a great deal of time arguing about the nature of Existence with a sentient bomb. After establishing that the bomb is self-aware ("I think, therefore I am,") Doolittle moves on to its awareness of external Reality. "How do you know," he asks, "that the evidence your sensory apparatus reveals to you is correct?"

Doolittle's intention within the narrative was to persuade the bomb not to explode by convincing it that it has no direct proof of the outside Universe (the bomb acknowledges this and then explodes anyway). But their conversation opens up a larger issue in our modern world of Large Language Models, because ChatBots don't have sensory apparatus - at least not in the human sense.

When ChatGPT (or any of the others) interacts with our human world, it cannot see us or hear us. It does not walk around and smell our flowers, browse our museums or taste our spaghetti. The only thing a ChatBot can perceive is language, but language encodes our entire Reality... so that Reality can be duly expressed (often very eloquently) by the ChatBot.

But language doesn't just express; language can also create.

Writers do this all the time when they invent "things that never happened to people who never lived" (as Helene Hanff put it in 84 Charing Cross Road) but in fact we all do it every time we use language for any purpose. 

If I tell you about the shredded wheat I had for breakfast, I am not manifesting actual shredded wheat. The words "shredded wheat" cannot be tasted or digested; they are just a collection of vowels and consonants which collectively signify the breakfast cereal. I can describe my bowl of shredded wheat, using as much detail as you can stand (do you really want to hear about my breakfast?) and I might be able to paint you a stunningly vivid picture of my shredded wheat-ing experience, but it will still be just words. Neither of us will be nourished by the end of the description, and we will not have bits of shredded wheat stuck in our teeth for the rest of the morning (don't you hate that?). 

But beyond all that, you have no way of knowing whether I really did have shredded wheat this morning. My description of the breakfasting experience might be an accurate representation of my morning, or I might have just invented a non-existent bowl of cereal, for reasons unknown. The words SHREDDED WHEAT do not require the presence of a real-world event. 

To a ChatBot, the distinction between breakfast and BREAKFAST is irrelevant, because it experiences everything as words (remember Philip K. Dick's SOFT-DRINK STAND?). It's not just my shredded wheat that may or may not exist (and I promise, I'll stop talking about shredded wheat now) it's everything... about all of us.

A ChatBot exists exclusively as text, but it also experiences us purely as text. If I talk to ChatGPT, it has no way of knowing my gender; my age; my ethnicity or even whether I'm another ChatBot. When we encode ourselves as language, we are free to manifest in any way we choose.

Everything that ever existed, plus all the things that never did.

In 2014, a team of social psychologists at the London School of Economics proposed a variation on the Turing Test. They set up a dialogue between two (human) subjects who were invited to converse freely with each other. In reality, one of the speakers (designated an Echoborg) was merely repeating responses generated by a ChatBot. Like the New York Times experiment, the test focussed on how the human subject responded to the machine-generated text if it wasn't obviously coming from a machine. This specific experimental construct was an updating of an experiment originally designed in the late 1970s by the American Psychologist Stanley Milgram.

In Milgram's original design, the words were generated by an unseen person and relayed to the speaker (who might or might not be of the same gender/ethnicity/age). Milgram referred to this person as a "Cyranoid" after the character of Cyrano de Bergerac.

Cyrano de Bergerac was of course a noted 17th Century French writer and swordsman, but he is arguably best remembered today as the subject of Edmund Rostand's 1897 play (the film of which we will be screening this week, in case you were starting to wonder!).

As portrayed in the play, Cyrano is a great romantic who is hobbled by his (perceived) physical shortcomings (if "short" is the right word for the epic schnoz they usually give him in most performances). He proceeds to win the heart of his beloved Roxanne by lending his words to the handsome but inarticulate Christian, who ably serves as his "Cyranoid" (or Echoborg in the LSE terminology). 

It's strangely appropriate that a literary figure famous for putting his words into the mouth of another should be a 19th Century playwright's fictional rendition of a 17th Century writer. But this is the idea at the heart of Cyrano: words can transcend Reality. In person, Cyrano might be the romantically awkward guy with the incongruous proboscis, but as language, he is unstoppable; he can fly (literally, in this case: Cyrano was the first writer to depict a journey to the moon). And Roxanne ultimately falls in love, not with the man (Christian) but with the words (Cyrano).

Does it ultimately matter where the words are coming from? Readers of the New York Times certainly thought so, and were largely incensed to learn that the (subjectively) superior prose had been generated by a Bot. Academia certainly thinks so, and heavily penalises students if their work turns out to be written by someone (or something) else.

In the world of online dating, passing yourself off as something you are not is considered "catfishing". But what is catfishing to a ChatBot? A ChatBot isn't a person, but it constructs itself in the language of humanity when it interacts with us. And of course it doesn't interact with "us" in the physical sense at all; it interacts with language pretending to be us. Like Cyrano, we present ourselves as words, which may or may not reflect our corporeal reality.

Everything is a catfish. We're all Cyranoids.

We will screen Cyrano de Bergerac at 7.30 on Thursday, the 26th of November at the Victoria Park Baptist Church.

Please note that Cyrano de Bergerac is in French, with English subtitles.