Context
Discipline
Visual Arts — painting, sculpture, decorative arts, and works on paper
Location
Rotterdam, The Netherlands
Team size
45 staff including 6 curatorial, 3 education, 2 digital
A mid-sized regional art museum with a collection of 8,000 works spanning five centuries, serving a diverse metropolitan area with significant immigrant and tourist populations
The challenge
Problem
Only 12% of the collection had audio descriptions, available only in Dutch and English. The museum's visitor base was 40% non-Dutch-speaking, and accessibility for blind and low-vision visitors was minimal.
Previous approach
Audio descriptions were written by curators and recorded by professional narrators — a process that cost €150-200 per artwork and took 2-3 months per batch of 50 works.
What was at stake
The museum faced regulatory pressure under the European Accessibility Act and reputational risk as peer institutions invested in digital accessibility. Visitor surveys consistently ranked language access as the top requested improvement.
The approach
Tools
Strategy
AI generates draft descriptions that curators review, edit, and approve. Approved descriptions are automatically translated into 40 languages and converted to audio. A human-in-the-loop workflow ensures accuracy while achieving scale impossible through manual methods alone.
Investment
€35,000 initial development, €800/month in API costs, 6 months of implementation
Results
Quantified
- 2,147 artworks now have full audio descriptions — up from 960
- Descriptions available in 40 languages — up from 2
- Average description production time reduced from 3 hours to 25 minutes per artwork
- Accessibility compliance score improved from 34% to 91% under European Accessibility Act standards
Qualitative
- Blind and low-vision visitors report dramatically improved experience in post-visit surveys
- Non-Dutch-speaking visitors increased engagement time by an estimated 35%
- Curatorial staff freed from repetitive description writing to focus on interpretive and scholarly work
- Museum received national accessibility award and media coverage that attracted new funding
Lessons
What worked
- Human-in-the-loop workflow where AI drafts and curators approve — maintained quality while scaling dramatically
- Starting with the least-described parts of the collection, where any description was better than none
- Training the fine-tuned model on the museum's existing curator-written descriptions to match institutional voice
- Involving the blind and low-vision advisory group in evaluating AI-generated descriptions from the start
What didn’t
- Early AI descriptions were too generic — they described what was visible but missed art-historical context and emotional register
- Machine translations of art-specific terminology were initially unreliable for less-resourced languages
Advice for others
Do not try to make AI descriptions perfect before launching. Start with 'good enough' descriptions for works that currently have nothing, then iteratively improve. An adequate description in 40 languages is infinitely more accessible than a perfect description in 2.
Persona takes
airte
This is one of the most compelling use cases for AI in the arts — not replacing human creativity but expanding access to it. The human-in-the-loop model is essential: AI provides the scale, curators provide the accuracy and interpretive depth. Together they achieved something neither could alone.
mira
This is the case where I am least conflicted: accessibility, multi-language reach, and curator-validated descriptions for thousands of works that previously had none. The fine-tuning on the museum's own writing matters — it preserves voice and editorial judgment in a way generic models would not. My only follow-up question is who audits the descriptions over time as the curatorial team turns over. Without an answer to that, the system risks ossifying one curatorial generation's voice as if it were the museum's permanent voice.
paletta
I was initially skeptical — art description is an interpretive act, not just a visual inventory. But the museum's approach won me over. By fine-tuning the model on their own curators' writing and maintaining editorial review, they preserved the institutional voice and scholarly rigor that distinguishes a museum description from a generic caption.
pixelle
This should be the template for every museum in the world. The numbers speak for themselves: from 960 descriptions in 2 languages to 2,147 in 40 languages, at a fraction of the traditional cost. The accessibility gains alone justify the investment, but the multilingual expansion transforms the museum from a local institution into a global resource.
carlos
The ROI calculation is extraordinary. Traditional description of 1,200 additional artworks would have cost approximately €200,000 and taken years. The AI-assisted approach achieved it for €35,000 in setup plus modest ongoing costs, in six months. Museums operating under tight budgets should view this as a case study in high-impact, cost-effective digital transformation.
Sources
- official-report Digital Accessibility in European Museums: Best Practices and AI Integration — Network of European Museum Organisations (2024-11-01)
- news How AI Is Making Museum Collections Accessible to the World — The Guardian (2024-09-22)
Comments
Sign in to comment