In the still of night
The moon casts its silver light
On the silent pond.
A bit misleading, that. Which is rather the subject of this post. There’s a new AI model called ChatGPT by OpenAI. Sure, it answers questions, but it also writes text on demand about almost any subject. It’s good. Maybe not human-level good in every instance, but surprisingly good.
Take the haiku above, for example. That wasn’t me. I asked ChatGPT to write a haiku in the style of Basho, the 17th c Japanese poet. That’s what it came up with, and it only took a few seconds. Or this below:
Golden slumbers kiss your eyes
Smiles awake you when you rise
Sleep, pretty wantons, do not cry
And I will sing a lullaby
Rock them, rock them, lullaby
Care is heavy, therefore sleep
You are pretty, therefore sleep
For when you wake, you shall have all
The world’s delight, and all the world’s
Fashion, and all the world’s desire
Sleep, pretty wantons, do not cry
And I will sing a lullaby
I told it to write a poem in the style of W.B. Yeats. That’s all I told it. Proper poets may take issue, but I think it did all right…except for the fact that it totally cheated. It just copied a poem from Thomas Dekker (1570-1632) and changed a few words. Something we still have to be aware of, though as AI models improve, it may be less of an issue. For all I know, the Basho was also copied, though I haven’t been able to find a close analogue, so maybe it didn’t. If anyone can point me to an original, I’ll take it down. For something like this, it’s harder to tell.
Which is kind of the point, so I had to do a little experiment. As AI models get better at imitating humans, we’re looking at a generation of students and essayists and perhaps even fiction writers never having to write another word if they don’t want to. Is it ethical? Of course not. Just ask the artists now raising the alarm over painter AI models like Dall-E2 and Stable Diffusion. You tell them what you want in the picture, and they draw it. Sometimes really well. Sometimes it’s the stuff of nightmares. But, like GPT, they’re getting better.
So, where AI is the problem (or our lack of ethics, fully debatable), AI might be the solution. Some in the AI field are working on models that can distinguish between human and AI generated work. I tried one of them out with a simple test: I showed it a paragraph written by ChatGPT on my prompt and then a snippet from the work in progress. I’m not showing the paragraph from my prompt, because I strongly suspect it was at least partially ripped off from an existing series, because the name of the kingdom was from a known series. It also wasn’t very good, though I’ve seen worse. The detector thought it was fake. I hope it’s right. Here’s mine below:
Nailed it. As the caption above states, the model needs a certain number of “tokens” (word breakdowns) to be reliable, 50 or above to have a chance of being right. That is, the longer the piece, the more chance it will get it right.
Here’s the rub: I also showed it both poems above. Failed both. Granted, the pseudo-Basho only produced 15 tokens, so not a surprise but the pseudo-Yeats? 93. Of course, most of it was written by a human, and the detector apparently doesn’t take plagiarism into account, but still…
We may be in trouble, long term. Short term? Try passing this stuff off as your own work and you might be the one in trouble.
Richard Parks 