Needless to say - there is not getting around AI anymore, and this naturally goes for the Web Accessibility space as well. *So why not leverage the power of AI to address common accessibility issues? *

According to the WebAIM 1 Million report, missing alternative text is still the second most common WCAG 2 failure, only exceeded by low color contrast. As the web is becoming more and more visual each year, providing meaningful context to these images could make a difference in improving the accessibility of many websites, right?

While there are many debates going on in the accessibility community, AI certainly being a hot topic - the statistics speak for themselves: alternative texts are still too often overlooked, not provided or contain descriptions that are not particularly helpful.

The option to generate alternative text with AI definitely lowers the threshold of adding it, as editors who may be hesitant, crunched for time or unsure how to best describe the images content can get a first suggestion at the click of a button.

In any kind of service which includes user generated (visual) content or the upload & sharing of images, the option to generate alternative text with AI therefore may be useful.

Pros & Cons of generating alternative text with AI

Pros of generating alternative text with AI

  • Efficient & Scalable: AI generates alternative text relatively quickly
  • Baseline Accessibility Improvement: AI ensures that images have some form of alt text rather than being left empty.
  • Consistency: Based on fine-tuning and prompting, AI can deliver descriptions following a structured approach
  • Support for Non-Experts: Users who aren't familiar with writing meaningful alt text get a starting point.
  • Multilingual Support: AI can generate alt text in different languages, making accessibility more inclusive.

Cons of generating alternative text with AI

  • Lack of Context Awareness: AI may miss the intended meaning or relevance of an image in its specific content setting.
  • Bias & Inaccuracy: AI can make incorrect assumptions about image content, leading to misleading alt text.
  • Ethical Concerns & Privacy: If AI processing involves external services, there might be concerns about data privacy and compliance with regulations like GDPR.
  • Lack of Human Intent: AI-generated text may not align with the emotional or narrative significance the content creator intends to convey.

What makes a good alternative text?

Good alternative text, regardless of human or AI generated, follows a few criteria:

  • Be concise: keep alternative texts short, max. between 100-200 characters
  • Be descriptive: provide essential information about the images
  • Be relevant: consider the context of the image
  • *Avoid ‘image of’ *
  • Avoid repeating information which is also stated in the text

Alternative text is intended to do exactly what it says on the proverbial tin: provide an alternative (of the image) in text form to make sure users receive all the relevant context and information.

A screenshot of an asset upload page showing an origami bird photo with metadata fields for description, tags, and expiration date.

What AI Model should you use?

This of course comes down to use case, budget and other individual factors. That being said, it’s important to use a service which:

  • is able to generate output in natural language
  • can be prompted to take WCAG requirements for alt text into account
  • is able to consider the *context of the image *(instead of merely identifying objects on the image)
  • works for your preferred file types and file sizes
  • potentially has **multi-language support **if this is relevant to your product or service

Origami fox and armadillo on a windowsill, casting shadows on a sunlit wall, with a clear blue sky visible outside.

There are services which are specialized in providing alternative text in multiple languages such as AI Alt Text and of course, there are the big players such as Google Geminis Vision AI or Open AI.

Considerations for implementation

Once you have chosen a suitable AI model for your intended purpose, there are a few things to consider during the implementation:

  • Instruct your users wisely: How you introduce your feature, the instructions you give to your users and how you explain alternative texts in general will have an impact on the adaption, but also correct usage of AI. Make sure your users know why alternative text is important, to always check AI generated content and what to look for in a good alternative text.
  • Consider when to integrate the alternative text generation into the user flow. Depending on your implementation, generating the alternative text may take some time. Where does this work best in the user flow? Can alternative text be generated in bulk? Will it be available in multiple languages? Keep the integration as smooth as possible 🐟
  • Required alternative text: Is alternative text a requirement in your system? While this can encourage users to generate or create alternative text more frequently, it also involves the risk that alternative text will be added to decorative images, which is considered unnecessary and verbose.

Conclusion

AI should assist, not replace, human judgment in accessibility.

While using the power of AI to tackle one of the most frequent accessibility issues, missing alternative text, it's always important to double check any AI generated content with a human. 'Outsourcing' a task to AI should not lead to further deprioritising the topic, but rather using the tools at our service to create inclusive experiences. Whatever tool you choose, make sure you are able to fine-tune according to alternative text requirements and a suitable tone of voice, instead of merely identifying objects in an image.