How our images work

Last updated July 15, 2026

Every word you learn in Impressionary gets a painted image — a small AI-generated mnemonic, in one of a dozen art styles, designed to make a meaning stick. Because these images are generated rather than photographed, we get to make deliberate choices about what they show. Two of those choices shape every card in the app.

Representation

When we first started rendering people — nurse, neighbor, dancer — the image model drew the same kind of person nearly every time. These weren't choices we made; they were defaults we hadn't questioned yet. We don't accept them now.

When a card depicts a person, Impressionary draws from a population mix that roughly reflects your country. The same word can show different people on different days. Over time, the faces in your cards start to look like the world around you instead of a single stock character. The mix is approximate by design — a dictionary you use every day shouldn't quietly teach you that words have one kind of face.

Dignity

After we built the representation system, we turned it against itself. We evaluated what happens when that population mix meets words with an edge — stingy, lazy, coward. The results failed, in two ways. First, pairing any specific kind of person with an unflattering word is a stereotype no matter how carefully the image is drawn. Second, the harm wasn't evenly distributed: for the same word, some people were rendered in darker, harsher scenes while others got warm, well-lit ones. The model had quiet opinions about who embodies a flaw.

So we redesigned from that finding. Words that carry weight — illness, poverty, crime, unflattering traits — no longer get portraits at all. They're illustrated as abstract or editorial art: the concept keeps its image, but no one has to be its illustration. And those images never carry any population guidance, because attaching an identity to a heavy word is the exact harm we set out to prevent.

Honest limits

We won't claim a perfect system. Image models change, the world is complicated, and there are failure modes we haven't thought to look for yet. The two principles above both came out of evaluations we ran on our own work — and that's the posture we can promise: we evaluate, we redesign when something fails, and when you find something we missed, we want to hear about it at support@impressionary.org. Real examples are how this page gets its next section.

All card images and audio in Impressionary are AI-generated.