Making Art Accessible to All
Visual Descriptions for Every Artwork in the Collection
A visit to an art museum depends heavily on visual perception. For patrons who are blind or have low vision, this dependence can significantly diminish the experience. Descriptive textual representations of artworks help close this gap, giving blind and low-vision visitors access to the composition and content of a piece in ways that parallel how sighted visitors experience it. With that goal in mind, the CMA set out to write visual descriptions generated with the help of AI for the primary images of all 68,000+ objects in its collection, to be used both online and on-site. Explore any artwork on Collection Online and open its visual description modal to take a look.
Laying the Groundwork
In the summer of 2023, the CMA launched a project to create visual descriptions for the entire collection, building on the momentum of the museum’s accessibility-centric website that had debuted several months earlier. That site was developed with substantial support from Prime Access Consulting (PAC), an accessibility firm led by disabled experts. With a sharpened awareness of the needs of disabled patrons, the CMA redirected its partnership with PAC toward this new effort.
The first step was defining what makes a high-quality visual description. Over several months, the CMA worked with PAC to develop a comprehensive style guide tailored to the museum’s collection, one that outlines how descriptions should be organized, what visual details to prioritize, and what standards to uphold. From there, the museum’s team began implementing the guide by writing descriptions for collection highlights and a broader range of objects, then refining them based on feedback from PAC.
Support from AI
Manually writing descriptions for the remaining tens of thousands of objects in any reasonable time frame was not realistic, so the museum turned to AI to support generation at scale. The project went into hibernation for more than a year, however, as the available technology was not yet capable of producing results consistent with the style guide.
By 2025, advances in multimodal large language models changed that calculus. The CMA partnered with Google Arts & Culture (GA&C) to bring the project to completion, using Gemini, Google’s multimodal AI, to generate the remaining descriptions. After months of refining prompts through careful integration of the style guide’s rules and the use of manually written descriptions as reference examples, generous funding from GA&C helped offset the costs of text generation at that scale.
Bringing It All Together
Each visual description is attached to a specific image rather than an artwork record. This distinction matters because many objects in the collection have multiple views, and a description written for a front-facing image would be inaccurate for an alternate angle or detail shot. Attaching the description to the image directly ensures it only travels with the image it actually describes, wherever that image appears across the CMA’s platforms.
Constructing this system required significant changes across the museum’s technological infrastructure. The CMA’s digital asset management (DAM) system and collection management system (CMS) received new fields to store the visual description text and to indicate whether each description was written by a human or generated by AI. Both systems enforce logic so that AI-generated descriptions can be updated with manually written text, but never the other way around. The museum’s API surfaces those fields from the CMS, and the website draws on them to assemble the visual description modal that visitors see today, which includes a disclaimer for descriptions generated by AI. Examples of both versions are shown below.
Visual description for Statuette of a Woman: "The Stargazer"
Visual description for Adoration of the Shepherds
Before any description goes live, it passes through an automated quality evaluation using cosine similarity, a metric that measures how closely the generated text aligns with the image it describes. Descriptions that meet the threshold are published directly; those that fall short are flagged for human review.
On the website, descriptions power screen-reader-accessible alt text and appear in a dedicated visual description modal, making the content available to visitors of all abilities. Each modal includes a link to a feedback form where visitors can share thoughts on any description, whether manually written or AI generated, supporting the CMA’s commitment to ongoing improvement. The system is also in the process of being designed with future extensibility in mind, with the ArtLens Gallery and its corresponding app among the planned next deployments.
From the Cleveland Museum of Art
Digital Innovation and Technology Services: Jane Alexander (Chief Digital Information Officer), Tara Bobinac (Associate Developer), Andrea Bour (Collections Information Data Analyst), Delaney Marrs (Walton Family Endowment Fellow in Digital Innovation and Technology Services)
External Partners
Prime Access Consulting
Google Arts & Culture
Frequently Asked Questions
The central goal of visual descriptions is to make static media and objects accessible to those who are blind or low vision. While there are many different definitions of visual descriptions, we worked with accessibility experts at Prime Access Consulting to create a visual description definition specific to our collection and its description needs. We chose to write short visual descriptions, defined as brief descriptions of the key visual features of an object. Overarching guidelines for the descriptions include:
- Use up to 75 words
- Write in an active voice using complete sentences
- Detail only what can be seen in the image of the artwork being described
- Use language that can be understood by a general audience
Our goal is to allow visitors to imagine an artwork’s appearance and make interpretations on their own terms. Our descriptions focus on providing visitors with the information that helps them do so, including detailing key visual elements (medium, shape, size, color, and texture) and their spatial relationships. Further interpretation and context are for curators to expand on in the “Did You Know?” and “Description” sections of our art detail pages. Our guidelines emphasize that visual descriptions should not interpret the emotion or meaning of an artwork, nor incorporate outside context.
Our visual description style guide discourages censoring due to potential offensiveness. Blind and low vision users should be given access to the same information as sighted users. However, detailed guidelines in the style guide outline best practices for writing about sensitive content with respect and accuracy.
Bias is always a concern for both human- and AI-generated visual descriptions. We crafted our style guide to be robust enough to establish replicable standards. These standards outline what language is and isn’t acceptable and how information can be structured within a visual description to ensure that collection objects are effectively and consistently described.
Applying our style guide to all visual descriptions minimizes bias and makes it easier to identify when a description’s standards have been breached. We take care to find, check, and correct instances of bias that come to our attention. All issues are recorded so that we can better improve our style guide and future prompting for the AI generation of descriptions.