This pilot tests how AI can be used to improve access to the collection while maintaining control, transparency, and professional standards.
The system generates image descriptions and supports semantic search, but all use is governed by defined principles and continuous evaluation.
The purpose is to make the collection more accessible, searchable, and inclusive, without compromising data security or artistic integrity.
Principles for responsible AI use
The pilot is based on a set of clear principles that guide both development and implementation.
Transparency
All AI-generated text is clearly labelled. Users are informed when content is generated by AI and how the system works.
Human control
Staff can review and edit generated descriptions. When a description is changed, new embeddings are generated automatically so that search results reflect the updated text.
Inclusion
Descriptions are written in clear and accessible language. In the pilot phase, alt text is tested with users, including people with visual impairments.
Artistic integrity
Descriptions are limited to observable elements. The system avoids interpretation, speculation, and assumptions about meaning or intention.
Continuous evaluation
Methods, prompts, and outputs are tested and adjusted throughout the pilot. Quality assurance is an ongoing process.
Data security
Images and metadata remain within the museum’s infrastructure. They are not shared with or used to train external models.
User feedback
Users and professional communities are invited to provide feedback. This is used to improve descriptions, search results, and overall experience.
How these principles are implemented
The principles are applied directly in the system.
Labelling and explanation
AI-generated descriptions are marked clearly in both alt text and search results. Users can access a separate explanation of how the system works, including a description of embeddings.
Editorial interface
Staff have access to a tool where they can review and edit generated descriptions. Changes are immediately reflected in the search system.
Prompt design
Prompts are designed to produce neutral, descriptive text. The system avoids interpretative language and focuses on visible features.
Example prompt:
“What is shown in this image? Provide a detailed description so that a blind user can understand the content. Describe colours and composition.”
Testing and validation
Descriptions are tested with users during the pilot phase. This includes evaluating clarity, usefulness, and relevance.
What we expect to achieve
The pilot is used to test whether AI can improve both access and understanding.
We expect to see:
- more relevant search results beyond traditional metadata
- improved discovery of connections across the collection
- descriptions that support accessibility and understanding
- a transparent and verifiable method for using AI
Limitations and responsibility
The system generates descriptions automatically. These descriptions may contain errors or omissions.
AI-generated text is treated as:
- a machine-produced description
- a supplement to existing metadata
- a starting point for further review
Responsibility for interpretation and final use remains with the museum and the user.
Next steps
The pilot is developed in stages.
- Test and adjust methods on a limited dataset
- Establish routines for quality control and evaluation
- Document how data is handled and protected
- Share experiences internally and with the wider sector
The aim is to develop a method that is transparent, robust, and suitable for long-term use.