Digital Accessibility Index: Learn where the world’s leading brands fall short on accessibility.

See Report

New GPT-4 Model Can Reportedly Describe Images Accurately

Apr 20, 2023

OpenAI has announced a new version of its popular GPT (Generative Pre-Trained Transformer) artificial intelligence tool — and a powerful new image recognition feature, which may have enormous implications for web accessibility. 

“GPT-4 can accept a prompt of text and images, which [...] lets the user specify any vision or language task,” the OpenAI team writes in a blog post announcing GPT-4. “Specifically, it generates text outputs (natural language, code, etc.) given inputs consisting of interspersed text and images.”

In plain language, that means that users can submit an image and ask the artificial intelligence (A.I.) questions about the visual content.

The technology could improve experiences for users with vision disabilities. Instead of relying on content creators to write accurate alternative text, those users may be able to use A.I. to understand the context of visual content. 

The implications for accessibility are enormous — with a few important caveats.

No A.I. model is perfect, and chatbots make mistakes

Alternative text (also known as alt text or alt tags) provides important functionality for people who use screen readers (software that converts text to audio or braille output), people with low vision, and people who browse on slower internet connections. 

If A.I. can generate alt text for images automatically with a reasonable degree of accuracy, that would help to address a major accessibility barrier. 

In an analysis performed by WebAIM (Web Accessibility In Mind), 55.4% of the internet’s top 1 million websites had missing alt text for images on their homepages. If automated tools could generate alternative text for all of those images, the internet would be a much more inclusive space. 

But currently, no A.I. model can describe images perfectly. That’s one of the reasons that image recognition is frequently used for human verification on web forms. 

Related: Will Artificial Intelligence Help Accessibility?

A.I. models are prone to hallucination

Hallucination refers to the tendency of A.I. to confidently present false information. A.I. chatbots are particularly prone to hallucination, since their training data consists of internet content (and unfortunately, not everything you read on the internet is true). The chatbot’s goal is to speak with authority — not necessarily with accuracy.

That’s an issue, because if an A.I. model presents incorrect context, web accessibility suffers. For example, let’s say that an A.I. model describes an image as “a green apple.” If the image actually shows a pear, the alt text isn’t useful.

Based on OpenAI’s sample images, GPT-4’s output seems accurate. However, the Web Content Accessibility Guidelines (WCAG) specifically requires text alternatives to serve an equivalent purpose to non-text content. For A.I. tools, that’s a high bar.

If A.I. fails to recognize context that would be obvious to a human user — for example, the image of the apple is on an orchard’s website and shows crops from a recent harvest — it isn’t truly equivalent. 

And even if A.I. tools are perfectly accurate, they may create issues for users. The tools may not be able to determine whether images are decorative (purely decorative images don’t need alt text). They may write excessively long descriptions for incidental images, which can impact the experiences of screen reader users. 

Related: 8 Common Image Alt Text Mistakes to Stop Making

A.I. is the future of accessibility, but human experts will always play a crucial role

In the near future, A.I. solutions will provide users with new tools for understanding content. If a website contains untagged images, improper color contrast ratios, keyboard traps, or various other accessibility issues, A.I. tools may be able to quickly address those problems.

However, when you’re developing content for real people, you need to rely on the judgment of real people. The best accessibility solutions combine powerful A.I. with expert human guidance, which allows for effective remediations at scale without sacrificing accuracy. 

The bottom line: A.I. tools will be a powerful asset in the fight for digital accessibility, but you shouldn’t wait for new technologies to take the first steps. By prioritizing inclusive design as early as possible, organizations can create content that reaches more people — and provides better experiences for everyone.  

To learn more about the benefits of digital accessibility, read The Business Case for an Accessible Website

Use our free Website Accessibility Checker to scan your site for ADA and WCAG compliance.

Powered By

Recent posts

Understanding WCAG: What Does "Accessibility-Supported" Mean?

Jan 12, 2024

Does Font Size Matter for Web Accessibility?

Jan 8, 2024

What’s the ROI of Web Accessibility?

Jan 3, 2024

Not sure where to start?

Start with a free analysis of your website's accessibility.