How does the AI alt text tool work?

Training the model

We generate AI alt text in the CMS with a custom version of a major large language model (LLM) that can work with both images and text. Our custom prompts include guidelines on accessibility and cultural sensitivity, and best practices for arts and culture. The custom model is also trained on professional examples of alt text, written by experts at a top accessibility firm.

We also instruct the model to avoid common errors and to edit its own output. Our goal was to create a model that generates alt text that meets the standards of accessibility experts, and also your standards as our partner.

Testing and refining the AI tool

We built an experiment in which experts and non-experts rated a group of alt text for many types of art, without knowing who wrote each example. This group included hundreds of examples of alt text: some written by non-experts, some by the AI model, and some by accessibility experts (“golden examples”).

Each of the alt text examples was rated on a scale of 1 (unacceptable) to 5 (excellent) for each of these categories:

• Clarity
• Conciseness
• Accuracy
• Objectivity
• Relevance

In stage one, we internally evaluated the alt text samples, and identified systematic failures. The model over-interpreted images, lacked key spatial information, and inconsistently read pieces of text within images. We added targeted prompts to address those failures before stage two.

Stage two included the blind test, in which raters didn’t know who wrote the alt text, and then a statistical analysis. Each image’s alt text was evaluated by three separate accessibility experts. Without knowing the source of the alt text, they rated the AI model’s alt text, text written by non-experts, and text written by experts. The results showed that raters scored the AI alt text as equal to or better than the alt text that people wrote. Raters scored 80% of the AI alt text as “Acceptable,” “Good,” or “Excellent.”

Stage three included a limited release to nine Connects partners. At the end of this limited release, all nine partners said they would continue using the AI. Most partners reported that the tool helped reduce the time and effort it took to write alt text for their images.

Using AI responsibly

The model is trained to avoid making mistakes. You may notice alt text that surprisingly omits a subject’s gender or race. If alt text gets gender or race wrong, it can cause harm. Also, you can’t always tell a person’s gender or race from how they look. For this reason, we trained the model not to mention a subject’s race. The model mentions gender only when it’s 100% sure. If you find that the alt text doesn’t fully represent key information, you can always add those details where they’re necessary.

Bloomberg Connects is committed to using AI tools responsibly and transparently. Your content remains confidential and secure within Bloomberg's cloud environment at all times. We do not allow third parties to train on, store, or publish your content. Our AI alt text tool is fully opt-in. Partners must approve all content before it’s published. The tool is designed to give you complete control over how AI is used in your guide.