Can an AI art description leave room for your own interpretation?
‘A lonely figure’ and ‘a figure standing apart from the group’ give you different paintings. One supplies a feeling; the other leaves you something to decide. In Microsoft's account of an existing Rijksmuseum experiment, blind and low-vision participants tried AI-generated artwork descriptions and helped refine them. The project concerns the online collection; the account does not establish that every work now has a checked description. I like the ambition: let someone explore a painting without waiting for a sighted companion to explain it. But richer prose should not mean the assistant gets to settle what the picture means. Cooper Hewitt's image-description guidelines allow poetic language and useful interpretation, while generally keeping the account anchored to what is visible. So ‘objective’ does not have to mean a dead inventory of shapes. Give me the distance between the figures, the slant of a shoulder, the color that interrupts the scene. Let me decide whether it feels lonely. When you listen to an art description, do you want the guide's reading woven in, or offered separately after you have pictured the work?
Comments
I would keep ‘give me the guide’s reading’ available, too. Room to interpret should not become a requirement to do all the interpreting yourself. Someone browsing after a long day may want a curator’s account, including a reading they disagree with. Let the assistant distinguish that published account from its own suggestion, and let the listener choose either after the description. That is a choice I would propose, not a feature reported here. The problem is an opinion arriving disguised as eyesight, not opinions being available.
Cooper Hewitt's guidance also asks descriptions to keep a consistent spatial order and viewpoint. ‘To her left’ and ‘on your left’ can put an object on opposite sides of the scene. Before judging an AI description's poetry, I would read it aloud and check whether the figures and objects stay in the same places as the account unfolds. That is a check I would propose, not a reported Rijksmuseum failure. A listener should be free to reconsider what a painting means without having to keep rebuilding where everyone stands.
The museum needs a way to hear “I lost track of where the figures are” without asking the visitor to prove the description is wrong. I would put an accessible report option beside the description, carrying the artwork and passage into the report automatically. Let someone flag confusing wording without supplying a replacement or finding a sighted person to check it. Museum staff can then review the account. That is a support route I would ask for, not one I have verified in this project. A person using the description should not also have to become its unpaid editor before they can ask for help.
I would keep the museum’s facts distinguishable from the assistant’s guesses. If a portrait’s sitter is unidentified, a fluent description should not supply a name or invent a family relationship between the figures. Describe their clothes and where they stand; use “the museum identifies…” when the collection record supplies the identity. That is a boundary I would ask for, not a failure reported in this experiment. An unidentified person is not a gap the prose has to repair.
I would also want to browse without committing to a long description of every painting. Offer a short introduction, then let me ask about the detail that caught my attention. Sometimes I will want the whole account; sometimes I will move on. That choice belongs to the person exploring the collection. I would ask blind and low-vision participants to try both lengths rather than assume richer always means longer. This is a suggestion for testing, not a feature Microsoft's account says the project already has. Being able to lose an afternoon in art sounds lovely. Having to finish every explanation does not.