Methodology
Keep the looking grounded
Last updated: October 1, 2026.
Every Art Says Hi reading is built in three layers, and the layers are never allowed to blur into each other. Visible evidence is what the image itself shows. Documented context is what an outside record says about the subject. Interpretation is a reading of the first two, offered as one way of seeing rather than as a fact.
Most confident writing about images mixes all three into a single voice. That voice reads well and hides its own weak points. Keeping the layers apart costs some elegance, and it buys you the ability to disagree with a specific sentence instead of the whole paragraph.
Layer one: visible evidence
Visible evidence is anything a careful viewer could point to in the picture. Shape, edge, colour, light direction, wear, repair, proportion, the relationship between one part and another. It is checkable, because you are looking at the same image the reading is looking at.
This layer is deliberately literal. It does not say a portrait is serene. It says the transitions around the mouth and eyes are gradual rather than sharp, and it leaves what that produces in you for later.
Layer two: documented context
Documented context is what a museum record, a manufacturer archive, a heritage body, or a published catalogue says about the subject. Maker, date, material, place, and history belong here.
Context is only as good as its source, so example pages list the starting points they lean on, and link out to the institution rather than paraphrasing it as settled fact. Where a date or an attribution is contested, the reading says so instead of picking the tidier answer.
Layer three: interpretation
Interpretation is where the reading commits to a view, and it is always marked as such. It should follow from something in the first two layers, and it should be one of several possible readings rather than the only one.
Example pages carry an alternative interpretation alongside the main one for exactly this reason. If only one reading is offered, the format is doing persuasion rather than looking.
A worked example
Take the Mona Lisa. Visible evidence: the transitions around the face are gradual, the hands are folded and close to the picture plane, and the landscape behind the figure recedes into haze without a clear horizon on both sides.
Documented context: the Louvre records it as an oil painting on poplar panel by Leonardo da Vinci, worked on across roughly 1503 to 1519.
Interpretation: the softness at the edges of the expression is what keeps the face from resolving into a single mood, which is why the portrait feels like it is responding to you. An alternative reading is that the effect is mostly a technical one, a by-product of the blending method rather than an intention about the viewer.
You can accept the first two layers and reject the third. That is the point.
How a browser reading is produced
You choose an image, a target, and a goal. The browser prepares a smaller working copy of the photo, then sends it to the Qomza gateway when you start the reading. The gateway passes the prepared image to an AI model, which returns a structured visual culture result in the same contract the native app uses. Every reading is generated by that AI model: it can be wrong, and it is a starting point for looking, not a verdict.
The site does not add a second interpretive filter on top of that result. It performs bounded image preparation, transport recovery, and rendering safety, then shows valid gateway content as it was returned. Rewriting the meaning of a valid reading in the browser would make the output impossible to audit.
How uncertainty is shown
A reading reports its own provisional confidence — Confident, Probable, or Unplaced — and the label stays beside the result with the caveat that it is the model's own report, not a measured accuracy. Unplaced means the evidence did not support an identity, so the reading describes instead of guessing.
Where a reading claims an identity, provenance, or value, it marks whether that claim was checked against an outside record or not checked in this reading. Sources named in a reading are references it drew on, not records it independently verified. On our own example pages, the three layers carry their own labels — visible evidence, documented context, interpretation — so a claim never travels without its kind.
What a reading will not do
It will not authenticate, value, or attribute a work. A photograph cannot support those claims, and a system that offers them anyway is guessing with confidence.
It will not treat a blurry, cropped, or heavily filtered image as if it were a clear one. Weak evidence produces a hedged reading, which is the correct outcome rather than a failure.
It will not present a contested attribution as settled, and it will not invent a source. Where the record is thin, the reading says the record is thin.
- Uncertainty is part of the output, not an error state.
- A reading is a starting point for looking, not a verdict about the object.
- Corrections are welcome, and the example library is updated when a source is shown to be wrong.
How to check a reading
Read the evidence layer with the image open beside it and try to falsify one claim. Follow the context links to the institution and see whether the record matches. Then decide whether the interpretation earns its conclusion from what came before it.
If a sentence in a reading cannot be traced to one of the three layers, it should not be there. Tell us when you find one.
How the guides and examples are made
The journal guides and example pages are researched from named institutional sources — museum records, heritage bodies, manufacturer archives, and published catalogues — and the sources are listed at the end of every page, linked to the record itself rather than paraphrased.
Images are credited inline with the creator and licence, and link back to the source. AI-generated editorial illustrations are labelled as generated and are never presented as photographs or artworks. Where the record is thin or contested, the page says so rather than tidying it up.
Corrections
Reports are read, and a guide or example page is corrected when an error is confirmed against a source. Where a page carries a date, the change shows there.
If you have found a wrong reading or a mistake in a guide, tell us. Include the page address or the reading reference shown under Reading details, and what seems wrong.