Harnessing AI for Better Web Accessibility: A Practical Guide to Alternative Text Generation

Introduction

Skepticism about artificial intelligence is healthy, especially when it comes to accessibility. As Joe Dolson wisely points out, AI can be a double-edged sword—capable of both inclusive and harmful uses. Yet, despite the risks, there are genuine opportunities where AI can create meaningful improvements for people with disabilities, particularly in the realm of alternative text (alt text). This guide will walk you through a practical, step-by-step approach to using AI as a tool for generating and enhancing alt text, while keeping human oversight at the core. Remember: AI is not a replacement for human judgment, but a force multiplier when used correctly.

What You Need

• Access to an AI image description tool – e.g., GPT-4 with vision, Microsoft Azure Computer Vision, or open-source models like CLIP.
• A content management system (CMS) or editor that supports alt text fields (WordPress, Drupal, etc.).
• Basic understanding of WCAG 2.1 guidelines for alternative text (Success Criterion 1.1.1).
• A sample set of images – at least 10–20, including decorative and informative ones, plus complex images like charts.
• Patience and a critical eye – AI outputs require review and correction.

Step-by-Step Guide

Step 1: Understand the Limitations of Today's AI Alt Text

Before you start, acknowledge what AI cannot do well. Current computer-vision models examine images in isolation, missing the surrounding context. They often fail to distinguish decorative images (which need empty alt text) from meaningful ones. Descriptions can be vague, irrelevant, or outright wrong—especially for complex visuals like infographics. By knowing these weaknesses, you set realistic expectations and avoid blind trust in AI output.

Step 2: Use AI as a Starting Point, Not a Final Answer

Prompt your chosen AI tool with the image and a request like: "Describe this image in a few sentences, focusing on elements relevant to web accessibility." The result will rarely be perfect, but it saves time—you can edit rather than write from scratch. For example, if the AI produces a nonsense description, use that as motivation to craft something accurate. Think of AI as a rough draft generator. Always review and revise to meet WCAG standards.

Step 3: Train a Model to Identify Decorative vs. Informative Images

One promising approach is to fine-tune a model on your own image library. Collect examples of decorative images (e.g., spacers, backgrounds) and informative images (e.g., product photos, diagrams). Use a tool like Azure Custom Vision or a simple Python script with transfer learning. The goal is to have the AI flag images likely requiring descriptions, and those that don’t. This speeds up human workflows: you only check flagged images, rather than every single one on your site.

Step 4: Implement a Human-in-the-Loop Workflow

Never publish AI-generated alt text without human approval. Establish a process where an accessibility specialist (or content author) reviews each suggestion. The loop works like this:

1. AI generates initial alt text.
2. Human reviews and corrects it.
3. Corrected text is stored and optionally used to retrain the model.

This continuous feedback loop improves accuracy over time, while maintaining quality control.

Step 5: Handle Complex Images Separately

Graphs, charts, and maps are notoriously hard for AI to describe succinctly. For these, AI might provide a basic caption (e.g., "A bar chart comparing sales over four quarters"), but you must supply the detailed data interpretation manually. Use longdesc attributes or provide a separate data table. AI can assist by extracting numerical data from the image, but the narrative explanation is a human responsibility.

Step 6: Evaluate and Iterate

Periodically audit the alt text on your site. Use automated tools like axe DevTools or WAVE to flag missing alt text, but also manually sample generated descriptions. Ask: Is it concise? Does it convey the same information as the visual? Is it free of bias? Adjust your AI prompts and training data based on these audits. Over time, you’ll build a more reliable system.

Tips for Success

• Start small: Pilot AI for alt text on a single section of your site before scaling up.
• Never rely on AI alone: Human judgment is irreplaceable, especially for nuanced content.
• Context matters: Teach your AI model to consider the surrounding text—headlines, captions, and page purpose.
• Keep the end user in mind: Ask people with disabilities to test your alt text. Their feedback is invaluable.
• Document your process: Create guidelines for your team on how to review and correct AI outputs.
• Stay updated: AI models improve rapidly. Re-evaluate your tools every 6–12 months.

AI is not a magic wand for accessibility, but it can be a powerful assistant. By following these steps, you’ll harness its strengths while guarding against its weaknesses—making the web more inclusive for everyone.