The Dead Have Better Material

I try singing.

The first note comes out with something attached to it. I clear my throat and give the ceiling another chance. There is nobody here to stop me, which is one of the weaknesses of living alone.

I have decided to write a song about artificial intelligence. These are my own rotten lines, not something rescued from an archive:

The machine had read a million books,
And asked me for one more.
I told it I was fresh out of books.
It asked about next door.

Terrible. Still, I can sing it without hiring a consultant to explain what I mean. The rhyme has done most of the work. I am already considering giving it my job.

The trouble with a song is that somebody else gets to sing it. You can put all your intentions into the words and a man with a bad voice will come along and emphasize the wrong bit. He will change the line he can’t remember. He will make your sorrow sound cheerful. Your masterpiece goes into his mouth and comes out with his lunch on it.

I find this encouraging. It suggests a use for writing after the writer has finished being important.

Oxford has let OpenAI use digitized historical material from the Bodleian Library to train its models. Among the material scanned is a collection of ten thousand sixteenth-century broadside ballads: words and music once circulated on Tudor street corners. There are also historical dissertations. By June 2025, the library had shared 125,000 images from dissertations with OpenAI.

I will spare the dissertations my singing voice. The people who wrote them suffered enough getting them finished.

But the ballads catch me. A song goes from a street corner, through centuries of somebody bothering to keep it, into a machine. I reach for a pencil. My song needs somebody in the second verse, and the dead are conveniently unable to object to the part I give them.

They stole the song out of my mouth,
They sold it for a fee—

I stop there. A dead balladeer, apparently, has returned to deliver my opinion. I have not read his song. I have invented his grievance before finding out whether it belongs to him. All he has been allowed to bring back from the grave is a rhyme for me.

And what theft am I singing about? Oxford says this is out-of-copyright material, that OpenAI’s use is nonexclusive, and that the library retains the rights to the scans. The amount being digitized is modest, it says. The originals remain intact. The library plans to begin publishing the scans openly online within months.

I cross out the two lines. They were easy to sing. That is not in their favor.

If the scans let somebody read a page they couldn’t otherwise reach, the project has done something I want. Keeping a song difficult to find would be a peculiar way of defending the people who might sing it. I cannot spend a paragraph cheering the stranger who mangles a lyric and then demand permission from the dead before anyone uses it.

I could give the balladeer a speech about the company’s motives instead. Have him predict the entire business from a Tudor street corner. But I would have to put so much of myself into the poor bastard that he would need my trousers.

The partnership was announced in March 2025. That announcement emphasized digitization without stating the material would train models. Oxford rejects the suggestion that training was concealed and says staff were open about that use. Internal meeting records show concerns about reputation and environmental commitments. There are living people asking questions about this arrangement. They don’t need me to recruit a ghost.

OpenAI says it is proud to help its models preserve the world’s historical knowledge. Here I can find an objection without making anybody rise from the dead. Learning from a page is not the same as preserving the page. A scan can let me examine what was there. A trained model has learned from it; that doesn’t make the model an archive I can inspect.

My imaginary balladeer has just demonstrated the difference. I can manufacture a convincing complaint in an old man’s mouth without giving anybody access to what that old man actually said. I would rather have the page, even if the page ruins the song I was trying to write.

There is something worth singing about in the appetite, though. The web is filling with machine-generated text, making it less useful for training, and developers are turning to older physical collections for fresh material. The old words have become desirable partly because the new words are getting in the way.

I don’t need a costume for that one.

We filled the street with brand-new songs,
A thousand before tea.
Then went to find a dead man’s words—
He hadn’t heard of me.

That almost walks. I sing it twice. The second time I put too much weight on thousand and lose the tune.

I like the ugly arrangement it describes. We get the endless new prose. The companies go looking for human writing that predates it. Apparently the meal they want to cook for everyone is not the meal they prefer to eat. I am supposed to applaud the abundance while they search the shelves for something that escaped it.

An old song need not be good to be useful to a model. It need not be good to interest me, either. A lousy verse can show what somebody thought was funny, what he thought would sell, how hard he was willing to work to get a woman to rhyme with something. I would like to meet that failure in its own words instead of improving it before I have heard it.

When the scans become openly available, I could try. Until then, I have no business pretending my verses are a conversation with that collection. I have been arguing with a company and using a dead singer for atmosphere. The pencil has done some honest work crossing him out.

I still want another verse. No machine in this one. No spokesman. A landlord will do; I know enough about owing rent to supply my own embarrassment. I tap my foot until a line arrives, and try not to improve it to death.

My landlord knocked at half past eight.
I lay there counting ten.
I held my breath. He went away.
Then he came back again.

Source: Oxford lets OpenAI train its AI models on Bodleian library