Artificial Intelligence & Accessibility: A Look into ChatGPT

Setting the Stage

In today's rapidly evolving digital landscape, where technology permeates every aspect of our lives, we are witnessing the advancement of revolutionary tools to help us in the workplace and at home. The more we integrate these technologies into our routines, the more critical it becomes to incorporate digital accessibility. 

Chat GPT’s Accessibility as a Standalone Tool

When interacting with ChatGPT (out of the box), it does not meet standard accessibility conformance. Its interface is simplistic, but the text area is inaccessible with standard screen reader technology. We’ll review what that process looks like in more detail.

If you’re one of the 60 million people who visit ChatGPT every day and you are relying on any technology other than a standard keyboard and mouse, you may find yourself encountering a few difficulties. To start, when tabbing through the home screen, it is difficult to find a way to access chat history. This implies that there are many clickable elements missing from the focus/tab order. This was before even turning on any assistive technology. 

Furthermore, with a screen reader on, the area to type a message into ChatGPT is simply labeled “send a message.” This is confusing for a number of reasons, but for a non-sighted user, it would be much more difficult to piece together that this is the place for the tool’s primary function. “Send a message” implies filling out a form, but the context is missing. This is the primary reason users would be visiting the site, and yet it has the least amount of explanation for a user with a screen reader.

Once a message is entered into the message bar and the unlabeled button to search has been selected, (the screen reader announces it as “button, group”) the screen reader experience becomes a bit confusing. Without a screen reader, the button turns into a load bar, showing a rotation of one, two, and three dots, until the message is finished being written by ChatGPT. The screen reader announced, “middle dot, middle dot, comma,” rapidly and repetitively. It did not stop until the message was complete, and in some instances, this went on for an extended period of time. (The length of the result written by ChatGPT determines how long this load bar is present and visible; the more robust an answer, the more times a screen reader announced the repetitive load message.)

A more significant challenge is that the screen reader was not directed to the new content once it finished responding to a request. This isn’t necessarily a must, but it made things more confusing in my next step. Attempting to use standard navigation to go through each element after initiating the request, the screen reader navigates away from the main content area entirely. With the position of the message bar at the bottom of the message history, a screen reader user must backtrack from the send button and the input message and listen line by line in reverse to the response ChatGPT has given in order to get to the response. This is a clear opportunity for improvement. 

We encountered these issues when running the most high-level screen reader testing. The goal in conducting this testing was to gain an understanding of just how accessible this new technology is. Based on my analysis, this tool is not fully capable of working for the masses in its native environment.  The standards and guidelines for making a purely text-based tool like ChatGPT accessible are already in place with WCAG 2.2. ChatGPT can leverage these guidelines to make enhancements for a more accessible experience.  

ChatGPT’s Accessibility as an Integration

Despite the limitations within the tool with regards to accessibility, ChatGPT shines in the flexibility of its API. We discovered integrations via its API have very few limits regarding accessibility. 

There are multiple examples of applications built on the ChatGPT OpenAI API that have vastly different structures and appearances. This means that the technology can be manipulated to be more accessible, both from a visual perspective as well as a functional perspective, leading to a more adaptive tool that is friendlier to assistive technologies.

The power behind the ChatGPT API is mostly contained behind the scenes, and developers are granted the freedom to decide what the user-facing content looks like. Front-facing content can be developed with HTML frameworks such as React and Angular, and developers have the flexibility to style their applications in any way they would like. 

The sole exception to ChatGPT’s accessibility is that the content itself can sometimes be inaccurate. This poses a lot of questions we’ll walk through in the second part of this series. One example of how this may be problematic can be seen when using ChatGPT to automatically apply Alt Text to images uploaded to your site. There are likely to be instances where ChatGPT outputs incorrect data. This will be a common theme with ChatGPT, and it underscores the importance of making sure users do not become fully reliant on its capabilities without double-checking for inaccuracies.

The key takeaway: Accessibility is in the hands of the developers using the ChatGPT API, as the API has no integration restrictions. For applications that generate articles or distribute content, developers can use WCAG 2.1 guidelines and other accessibility best practices as guideposts to integrate the ChatGPT API in a conformant manner. 

Summary 

ChatGPT is becoming a go-to in today’s digital landscape. Its API integration serves well for adaptability and allows for an experience that is optimized for many different technologies. However, out of the box, and in its most used state, ChatGPT’s interface can prove challenging for users with assistive technology. As the tool continues to be refined, and updates are made, we look forward to seeing how ChatGPT evolves as a user-friendly, fully accessible tool. 

Definitions

Digital accessibility refers to the inclusive design and development of websites, applications, and digital content to ensure that individuals with disabilities can access, understand, and interact with those experiences without barriers. 

Artificial Intelligence (AI) is a branch of computer science that enables machines to mimic human intelligence and perform tasks, make decisions, and learn from data with varying degrees of autonomy. ChatGPT is a new, advanced language model developed by OpenAI that uses deep learning techniques to generate human-like responses and engage in natural language conversations.