MIKULÁŠ MITANA

Proxy

Spatial Design Studio

In your project, you used AI (ChatGPT) to create instructions for producing the final artwork. How did the process unfold, and what does the final set of instructions look like?
I began by creating a prompt that would have the AI write a truly usable manual for the physical realization of a sculpture – not a description of its shape, but real instructions on how to create it. This turned out to be the main problem. The AI could generate texts that looked like instructions, but they were too ambiguous. Phrases like “work with tension” or “leave space for quiet” created a framework for thinking but didn’t address specific questions. I revised the prompt over 11 versions. The turning point was manual number 6, which the AI structured like a musical score with tempo markings. The resulting objects were created based on this one.
But the final assignment arose from the clash between text and material. And that is the main outcome of the work: instructions can develop a possible framework, but a sculpture only comes into being through physical decision-making. I drew primarily on my experience with metal work and welding.

This reminded me of conceptual artists from the 1970s who simply wrote down instructions and anyone could theoretically create the work. What was the collaboration like with the non-human actor of artificial intelligence? Did you understand each other?
The parallel with Sol LeWitt is accurate. But I tried to flip the situation. It wasn’t that I, the artist, delegated the execution to others, but rather that the system delegated it to me. As the producer, I essentially swapped roles with the AI. 
And I think we understood each other – or at least the AI tried to. But the process was more complex: I created prompts for it, and it defined rules for my work, which is also a kind of prompt. I realized AI is a dataset like any other – it reacts to what it is given. So I began working differently for later manuals: instead of general descriptions, I gave specific artists as reference points. This network of limits ensured the responses were not too vague. 
It made sense when you knew the whole story of the prompts and the manuals. For someone without this background, it was confusing – instructions without context don’t work.

Was it possible to follow the AI’s instructions step by step, or did you encounter any obstacles or paradoxes that forced you to deviate from them?
The manual was so vague that I didn’t know where to start. Then I started running out of time, so I decided to take the latest manual and just get started. I followed the rules the AI set for me and created the object in about 4 hours. It was a matter of fairly random decision-making, but somehow it all fell into place. The main obstacle was that openness – there were no fixed points to build on.

photo: Polina Davydenko