What an AI Image API Should Never Store

The shortest version: do not store the things that would make the user feel betrayed if they saw them in your admin panel.

For an AI image API, that usually means prompts, source images, masks, reference images, generated outputs, and generation keys. Those are content-bearing objects. They can reveal private intent, identity, source material, client work, and final creative assets.

The API may still need durable records. It can store account IDs, model IDs, credit charges, task status, timestamps, error categories, capacity metrics, and content-blind operational events. These records let the business run without reconstructing what the customer created.

The difference matters. “A generation happened at this time using this model and cost these credits” is not the same as “here is the prompt and output.” A billing record is not automatically a content record. Keep that boundary sharp.

Source images deserve special discipline. They are often more sensitive than prompts because they can include faces, homes, client products, documents, or personal context. If the API stores them for convenience, it should say so. If the API claims privacy, source images should not leak into logs, dashboards, galleries, or debugging folders.

Generated outputs also need care. Synthetic does not mean harmless. Outputs can reveal the prompt, source image, product strategy, or personal use case. A hosted gallery is a retention feature. That may be useful, but it is not zero retention.

Keys are the quiet one. If an API uses encryption or sealed requests, durable key storage can undo the whole privacy model. A Generation Key Pair should be scoped to the task and live only as long as needed to process it.

The rule is not “store nothing ever.” The rule is “store operational facts, not customer content, unless the user knowingly asked for storage.”

If that sounds restrictive, good. Privacy is partly the art of not turning every useful byte into a permanent roommate.

Further reading: Unexposed data storage, How Unexposed works, and Why “We don’t train on your data” is not enough.