Now that Anthropic has turned on its digital watermark,
it’s time we decided who (or what) will write our best sentences.
The seeds were planted decades ago. First, it was the mythical homecoming queen my high school buddies and I created out of whole cloth. She made it to the final ballot and only bailed when the pageant officials demanded she show up in person for the gymnasium catwalk. Then there was the Wunderkind whose essays our English teacher swooned to read aloud. One class period we got so fed up of her fawning that when papers were handed up the row, the dullest among us swapped cover sheets with Mr. Write to finally grab the undeserved attention we felt we deserved.
What is it about mistaken identity that pushes us to such creative lengths to manufacture it? Shakespeare gloried in dressing characters as one another. Jacob awoke the morning after his wedding to learn he married the wrong sister. And now, it increasingly appears, we revel in pretending we wrote such tales ourselves.
I’m speaking of the power of Large Language Models to digest every stage play, every ancient story, every essay, poem, or book ever written so that we, for a modest or no fee at all, might pass any of them off as our work.
But no longer. Or so it would appear with the recent news out of Anthropic. To comply with European disclosure laws, the AI powerhouse recently agreed to insert a kind of digital ‘wordprint’ into its output, signaling to auditors which bits are machine-made. The watermark isn’t for the naked eye, but anything running the same algorithm can instantly yellow-card it.
I say, High Time!
I’m looking down at my feet as I write this paragraph. Over the weekend I visited a name-brand shoe store at my local mall. I had recently ordered what I was certain to be a pair of their very shoes—twice!—not from the brand’s online store but from a big-box aggregator claiming to ship them direct. The first pair arrived with a stain on its left toe. The right shoe in the second box came scarred with a six-inch scuff down its length. Tempted to cherry-pick a perfect pair from the four, I placed the unblemished remnants side by side to discover they hadn’t been painted from the same dye lot. Double-burned, I decided to return both boxes and revert to brick-and-mortar shoe-shopping. But when I rehearsed my sad adventure to the clerk, all he could offer was, “Never buy shoes online.” I instinctively knew what he was not saying. And after purchasing the authentic pair I’m now wearing, I wondered not for the first time whether I might one day express a gene or two from my cobbler double great-grandfather and learn to handmake my own footwear.
It is possible, I suppose, to defeat the inevitable AI watermark. As all the major players weigh not whether but when to follow Anthropic’s capitulation, we are stepping up to the decision line. Do we rewrite by hand the entirety of that AI-generated essay, poem, or novel we might otherwise be tempted to publish as our own? And if that’s the plan, why not just create the thing the old-fashioned way in the first place?
Modernists point to the mountains of time purists must scale to replicate the research already there for the taking. But to bypass the word-for-word making of a thing goes against what I often say about that most human of crafts: Write to remember; Write to get it right; Write to change. What is to be gained by turning over such a deeply personal obligation to the same machine that already lives too much of my life?
