A practical question: what does lower-impact AI look like?

Sustainable AI is not a label that can be added at the end of a build. It is a set of choices about scope, data, compute, deployment, and the lifespan of a product.

In practice, that can mean selecting a smaller model where it is sufficient, avoiding redundant data collection, measuring the real value of automated decisions, and designing for maintenance rather than novelty. The important point is to make these choices explicit.

I am interested in developing a stronger vocabulary and methodology for this work: one that treats environmental limits and social consequences as engineering constraints, not side notes.