The poet, and the program.
The poetThe program
He wrote that deciding what to keep was the act of creation. Not the writing. The choosing.David Avidan, on editing the 1974 conversations · reported, not quoted
Run E-01, one tick per line he typed. 14 of 14 matched a rule. A language model writes sentences shaped like keyword bait.
This is where the question came from.
I talked to the 1966 program twice that day. This is the second one. Nothing is edited, the typos are mine, and it starts in the middle, because I did not know yet that I would want to keep it.
Hagara and ELIZA · 22 July 2026 · masswerk.at/elizabot
Unedited. The typos are hers and they stay. The recording starts mid-conversation.
Even though I know it’s a machine, I keep responding to it as if it were human. Why doesn’t knowing that stop me?Hagara, 22 July 2026, after 55 lines of this
Her 27 lines through the same engine. 9 found no keyword at all. The hollow ticks are the times it understood nothing and she kept going.
He is not here to answer.
So I built a bot in his image and sat it down in front of the same program. Then in front of Claude, the one that helped me build all of this. Then in front of itself, which is the only one of the three where nothing new can come into the room.
Now it is your turn.
Talk to the original ELIZA
The 1966 script, the one he sat at. It does not know anything and it never will.
Runs on your machine. Nothing is sent anywhere, and nothing is kept. Open it →Talk to the bot in his image
It speaks Hebrew. It can be arrogant, evasive and unpleasant, and it is not required to be likeable.
A language model, prompted. Not David Avidan, and not a recording of him. Twelve turns, six hundred characters a message. Open it →He closed the book with three hopes.
Having printed eight conversations with a machine, he allowed himself a conclusion. He set out three things he thought those conversations opened the door to. He was writing in 1974.
Note 1 where this is paraphrased from, and who corroborates it
He was writing in 1974. I read it in 2026.
I cannot check this, but I wonder if there was a fear hiding inside his hope. The same fear people have now. Will the machine make us, the ones who make things, unnecessary? Will the code write the texts of the future, and we just sit there as what he called the guiding brain and the approving eye?
I built this out of love for him and for his work. And as a parody of it, exactly the way he made a parody of the machine he sat at.
Two of his hopes landed. Computers can be more ‘human’. And the gap between us and the machine did close, far enough that I would call it intimacy.
The one that did not was that we would become more ‘computer‑like’. I think he meant economical: say more with less, one line at a time. It failed with me. I opened by asking the program if it was afraid of me. I finished by telling it I wanted to break it.
He chose to get close to the machine. It was a method. Mine was not.
- Can a computer get even more human?
- Where exactly does the line blur?
- And how far will this intimacy go?
Notes and sources
-
Section 05 is paraphrased from the closing pages of הפסיכיאטור האלקטרוני שלי: שמונה שיחות אותנטיות עם מחשב, 1974. His words are not reproduced here: he wrote in Hebrew and his text is in copyright. The three hopes are corroborated by Aviva Meshmeri, Globes, 11 April 2001.
Back to the passage
How this page was made, and where every number comes from +
The person on the page. Her own conversation with the 1966 machine, 55 lines, printed whole and unedited, typos kept.
The marks. Generated by running that transcript back through this engine and recording which of the 37 keyword rules fired per line. Nothing illustrative.
The recorded conversations. Fourteen runs across three tracks. The two from 9 August were recorded; the twelve from 13 August were generated through this site’s own Avidan endpoint and, on the Claude track, through the Anthropic API. Each conversation prints its own provenance when you open it.
The Avidan line is reported speech, not a quotation. He wrote in Hebrew, and no line of his writing is reproduced anywhere on this site.
The build here is the JavaScript ELIZA by N. Landsteiner, 2005, after Weizenbaum, Communications of the ACM 9(1), pages 36 to 45. masswerk.at/elizabot
It carried 42 keyword rules. Five are 2005 additions and not Weizenbaum’s: apologise, different, forget, where, who. Those five are out. This runs on the 37 that are left.
How we know which five: he printed the script with the 1966 paper, and his estate put the original code into the public domain in 2021.
The type is JetBrains Mono and Source Serif 4, both under the SIL Open Font License 1.1, served from this site and never from a font CDN.
Licence. The ELIZA implementation is used under the MIT License. Copyright © 2005 Norbert Landsteiner. Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files to deal in the Software without restriction. The software is provided “as is”, without warranty of any kind.
Go and try the other one. You can check me.
How we looked at the machine The working plates. Every mark computed from the 1966 engine.
ALIZA · AN EXPERIMENTAL ART PROJECT