Two Drafts of the Same Silence: Reading Chinese Poetry Beside the Machine

There is a particular kind of quiet in my study at 6:40 in the morning, before the neighborhood wakes up and the radiators start their clicking. That is when I do my first pass on a poem. Lately I have been doing that first pass twice: once with my own hand moving across a notebook, and once with a browser tab open, a poem pasted in, and a machine offering me four lines in a fraction of a second.

I am not here to tell you the machine is bad. It is not bad. It is tireless, patient, and increasingly graceful, and I use it most days for exactly what it is good at — chasing down a botanical name, checking a dynasty date, confirming that a phrase I half-remember from a Song dynasty letter actually exists. The argument I want to have is narrower and more interesting than a verdict. It is about listening.

One Poem, Two Desks

I chose a poem almost every Chinese schoolchild can chant: Li Bai's four lines about moonlight at the foot of a bed, the suspicion of frost, the lifted head, the lowered head, the thought of home. It is short, famous, and deceptively simple — which makes it a clean test case. Nothing in it requires arcane vocabulary. Everything in it requires ears.

The engines, all three of them, handed me something close to this: moonlight before my bed, like frost on the ground; I raise my head and look at the moon; I lower my head and think of home. Smooth. Grammatically beyond reproach. A reader who encountered only that version would understand the poem's situation completely and feel almost none of its weather.

My own draft took six days and I am still not finished with it.

The Part That Does Not Survive Upload

Start with sound, because sound is where the gap opens first. The original rhymes. Two of its line endings land on the same vowel, and that rhyme is not decoration — it is the sound of a circle closing, a person leaving a room and coming back into it. The poem also runs on a strict tonal grid, a pattern of level and oblique tones that gives each line a slight forward tilt and then sets it down. Chanted aloud, it has a gait.

Machine output has no gait. It has correctness. And correctness, in a four-line poem, is a house with no floor.

I understand why. A translation engine is optimizing for fidelity to meaning across enormous distances, and meaning is the easy part to model. Prosody is not a fact about a sentence; it is a fact about a body reading that sentence. It lives in breath, in the ear, in the small physical pleasure of a vowel coming back around. You cannot average that out of a corpus.

The Allusion Problem Is Really a Memory Problem

Then there is the moon. In English, a moon is a moon. In the Chinese lyric tradition, that bright disc is a piece of shared furniture — it is the thing you look at when you cannot look at the person you miss, and it has been doing that job for a very long time. The word for home in that last line carries the weight of a place you have left, not a place you currently are. When the machine renders it as thinking of home, it is not wrong; it is simply reporting the surface, the way a good tour guide describes a cathedral's dimensions without mentioning that people have been married and buried in it for eight hundred years.

A human translator does not solve this. A human translator decides how much to carry and how much to leave in the original for the reader to feel as absence. That decision is the work. It is also the thing I argue with myself about at 6:40 in the morning.

There is one more piece of the poem that no engine flagged: an ambiguity in the very first word, a noun that can mean a piece of furniture or a well surround, inside or out, private or exposed. Scholars have bickered over it for centuries. The machine picked one reading and moved on, because a machine must produce a sentence. The poem's power partly lives in the fact that it never quite chooses.

The Silence Is Load-Bearing

Here is what I keep returning to. The poem's most important move is a repetition and a turn: head up, head down. Nothing happens. A person stands in a room and looks at the floor. The emotional event is in the space between the two gestures, and that space is silence — a caesura you can hear if you read it slowly, a held breath before the last line drops.

Machine drafts, by temperament, fill silence. They resolve ambiguity, smooth caesura, and close the gap a poet left open on purpose. It is not a failure of intelligence. It is a failure of restraint, and restraint is not something you can instruct a model to want.

So what do I actually do with the machine now? Three things. I use it as a hostile first reader, asking it to render lines flatly so I can see what the plain sense is before I start decorating. I use it to generate bad versions I can push against, the way a jazz player uses a metronome. And I use it as a reminder — every time it hands me something fluent and slightly airless, I get a free lesson in what my own job consists of.

What to Listen For Now

If you read poetry in translation and want to read it better in 2026, here is my short list. Read the same poem in two or three versions and notice where they disagree; disagreement is where the original is thickest. Read it aloud, even badly, and notice whether the English has a pulse or just a shape. Look for the places a translator has chosen an odd word — that oddness is often a signal, a hand raised to say the original would not sit still. And be suspicious of any rendering, human or otherwise, that is entirely comfortable.

I will finish my draft eventually. The machine would have finished it in November of last year, in less time than it takes to pour coffee. But the poem I want is the one with the floor in it, and floors take a while.