For two hours Sunday and via a chain of roughly thirty prompts, I asked an AI engine to write Shakespeare’s Antigone.
It’s not Shakespeare, not even Beaumont and Fletcher… but it’s not awful, either. It’s better than Gammer Gurton’s Needle.
The language is superficially right. It’s probably good enough that an untrained ear couldn’t tell the difference. It would pass as Elizabethan iambic pentameter, though not inspiring. That’s a low bar in a way, because any good actor working on a Shakespeare play learns instinctively to improvise in iambic pentameter. This is… a bit better. But not Shakespeare.
Yet.
I got the engine to put rhyming couplets in useful places. I convinced it to deliver an okay mix of short dialog and speeches. I prompted the engine to build multilevel characters that are at least as rich as Middleton’s. (That, to me, is also a low bar. Cf. The Phoenix, or the Hecate scene in Macbeth. Yeah. I’m not a Middleton fan.)
Two hours was as much time as I was willing to waste on this experiment.
But….
Let’s say I had invested another 20-30 minutes per scene refining and iterating. I would eventually have come away with a playable Antigone. Now add, say, four hours of prompts to turn the Chorus into multiple characters within a subplot rather than the Sophoclean speaker from which the engine borrowed. Note that I’m doing this all with prompts to an AI engine. I haven’t written a word of this myself. (I have, however, say, “Do better.” And it did.)
Those hundred monkeys are learning not just to type but to write.
Could AI eventually get good enough to create a lost play, such as Cardenio or Love’s Labour’s Won? Not this year… but it’s not as far away as we might think.
Let’s assume in eighteen months that the engines keep improving at the current exponential rate. I suspect they’ll by then have access to the various tools scholars use to analyze the likely authorship of specific passages, if they don’t already. In other words, the engine will be able to tell if the text passes tests of Shakespearean authorship, and modify the output until it does.
Were I then to do a PhD thesis in Shakespeare, I bet I could then use AI (and a small country’s worth of megawatts better spent in other pursuits) to create a text that might fool other scholars and would be enjoyable as a performance. That seems… scary.
There was one thing I couldn’t do, yet. I wasn’t able to find a way to inject Shakespearean humor, or even wordplay, into the text. Again, I spent only two hours here. I think that actual humor (but not wordplay) will remain out of reach for a few years.
I wonder if Will might have considered adapting the Antigone story, either early or late in his writing career. It might be fun to imagine.
Or, as I said, scary. (It’s not like Sophocles’ version of the story has a lot of humor, either. But the main characters – Creon, Antigone, and to a lesser extent Ismene and Haemon – are already complex and conflicted.)
PS: I used the version of Copilot included in my Microsoft 365 subscription for this experiment. I do not and will not use generative AI in my actual writing. This post was completely hand-fashioned.
PPS: A computer could have typed the PS. They lie and hallucinate, you know….