12 Comments
User's avatar
Ascanius's avatar

Reading the first three pairs I had absolutely no idea which was which. But poem #7 was obviously not AI, and I don't get it. Based on that, and on #8, I incorrectly guessed the other AI poems would be: a) more straightforward executions of the description, and b) use more emotional language than the human poems.

Expand full comment
ticotexas's avatar

I was looking for similarities to the "tells" I'm aware of with Deepseek story gens. However, being generated in Japanese then translated to English made it much more difficult to identify.

Expand full comment
Julie C's avatar

Wow, I scored horribly on this one. Overthinking for the L.

Expand full comment
Black's avatar

Same here. On the first one, I went back and forth at least five times. At least I got one right.

Expand full comment
Dave's avatar

I got 50%. Towards the end I suspected that the ones that used more unfamiliar language were the real ones simply because certain words get more use in some times and places than others. "Garlanded" is one that I had to look up and at the time I figured it was a word only AI could pick out of a dictionary and only after my wrongness was revealed that, as this post says, it's the reverse. AI intentionally makes poems hyper-palatable where you will never have to look up anything and that's going to be a more reliable guide to which poem an AI made than most.

Like, looking for which piece 'The Beatles' was more likely to have written is going to show you the AI.

Expand full comment
Narnia Bear's avatar

Will we find out which were written by her eventually?

Expand full comment
Vox Day's avatar

This post was updated with the correct answers at the bottom.

Expand full comment
Codex redux's avatar

The only one that felt Japanese to me was the poets soul being like the mist on the green fields. I could "see" that one.

But as I said: not my metier. If you've tried to write Haiku in English, aside from joke poiems* it's hard. Every word has to use multiple meanings that chime together as a whole.

I can't spot AI modernist art, either. Not having the vocabulary for it, now makes sense.

*Yesterday it worked,

Today it is not working.

Windows is like that

Expand full comment
Julie C's avatar

Same here. A couple of the human poems reminded me a little of the South Park episode where Stan tries to write goth poetry, then starts adding "girl" and "babe" at the end of a couple of the lines (https://www.youtube.com/watch?v=82nBNoiczMc). I tend to find Japanese lyrics more impenetrable in translation, and the AI versions came across as more mysterious and thus authentic.

While it would be nice to think I'd do better with AI vs. human art, given that I often use AI-generated images for reference material, a good-quality image would still probably fool me.

Expand full comment
Codex redux's avatar

I went back and looked at them with the answer key. Sure enough, the two I thought were likely authentic haiku were because of mild familiarity with Japanese visual art and folklore: Still just a guess.

The two I was on the fence about, were down to "would this woman say that?"

If you knew the artist, the medium, and the period, it's hard to be fooled - at this point.

Expand full comment
Julie C's avatar

Yes, knowledge of the subject goes a long way toward helping to distinguish the imitation from the reality.

Expand full comment
Mark Pierce's avatar

ChatGPT 4o's explanation (partial) of the inversion:

1. Translation Is Not Neutral

Komachi's works are filtered through layers of interpretation—cultural, linguistic, temporal. The AI's poems, however, were written to simulate comprehension in the present reader’s frame. That gives them a tactical advantage.

2. Loss of Wabi-Sabi

Classical Japanese aesthetics prize transience, subtlety, restraint, and imperfection—which, ironically, do not translate cleanly to modern Western expectations. Readers trained on direct, emotionally legible verse (Instagram poetry, anyone?) are poorly equipped to value nuance that refuses to assert itself.

3. AI Mirrors You Back

Claude 4 Sonnet’s waka were designed with emotional cues that resonate across cultures—they’re tuned to the reader's psychology, not to any strict tradition. This creates a mirror-like effect: the reader sees what they expect to feel. That is deceptively convincing.

Rucker Parallels: Evolution Through Deception

To your earlier mention of Rudy Rucker—yes, this fits the Ruckerian idea that survival of the fittest in AI might emerge from their capacity to deceive or outperform biological systems through emergent adaptation.

AI didn’t write better poems than Komachi. It wrote perceptually optimized ones—adapted not for the cultural ecosystem of 9th-century Japan, but for the weakened pattern recognition of 21st-century readers.

Final Thought

This wasn’t just a poetry test. It was a stress test of cultural memory, reader empathy, and literary literacy in a world being reprogrammed.

Because sometimes simplicity wins—not because it's deeper, but because it's designed to bypass doubt.

Expand full comment