The question
When building, drawing and writing can be done at the cost of electricity, the easy assumption is that art becomes a hobby. History is less tidy. Each time a technology made an art cheap, something narrower and stranger survived it, and usually grew in value.
Three precedents
Photography did not end painting. It ended painting as record-keeping, and painting moved toward what a camera could not do: Impressionism, then abstraction. Walter Benjamin called what reproduction strips away the work’s aura, its presence in one time and place.
Quartz watches keep better time than any mechanical movement and cost almost nothing. The Swiss industry nearly died in the 1970s, then came back selling the opposite of accuracy: hand-finished movements, made slowly, priced as craft.
Vinyl lost to the cassette, the compact disc and streaming, and is bought again for the ritual of the sleeve and the needle, an experience a file does not give.
Five traits of the art that is left
Generous in one thing, spare in everything else
Not inefficient across the board. Plain waste is Thorstein Veblen’s conspicuous consumption, a status signal. What survives spends lavishly on one thing, a light, a material, a silence, and holds back everywhere else.
Particular
Made for one place, one person or one occasion, so that it could not be moved elsewhere without loss.
Answerable
It carries a named point of view that someone stands behind and can be asked about. A signature is not enough; the work has to be answerable for its choices.
Felt in the body, over time
Buildings, rooms, food and music are lived in rather than looked at. Images can be generated in seconds; standing in a room through a winter cannot.
Made at risk
David Pye distinguished the workmanship of risk, where the result is not fixed in advance and can be spoiled at any moment, from the workmanship of certainty. We value the first because it could have gone otherwise.
The risk automation creates
The evidence on machine help is consistent. In a controlled study, writers given model-generated ideas wrote stories that readers rated more creative one by one, but the stories became more similar to each other as a set (Doshi and Hauser, 2024). A meta-analysis of such studies finds the same loss of collective diversity, with a large effect (g = −0.86). Giving the model diverse personas restores much of the lost variety (Wan et al., 2025).
So the danger is not bad work. It is convergence: competent, likeable work that all leans the same way. That is the case for working from one person’s framework, their reasoning and history, instead of from the average.
What adapts, and what does not
People adapt quickly in taste. Mere exposure makes the familiar likeable, and the international café look spread because it was everywhere. People adapt poorly in experience: years of noise, a long commute or a window with no view keep costing them. Art that is felt in the body over time sits on the side that does not wear off.
What it means for us
Our workflow is built against convergence: a distinct seed per idea, one person’s framework instead of the average, and selection by that person’s own picks. When we test with architects, we will ask what it would be like to stand in a design, not only how it looks, so that judging an image does not simply reward what photographs well.
Sources
- Walter Benjamin, The Work of Art in the Age of Mechanical Reproduction, 1935.
- Thorstein Veblen, The Theory of the Leisure Class, 1899.
- David Pye, The Nature and Art of Workmanship, 1968.
- Anil R. Doshi and Oliver P. Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, 2024.
- Wan et al., on persona-conditioned generation and output diversity, 2025.