Skip to main content

Artificial intelligence—who hasn’t heard of it? Today, you can encounter it practically everywhere. In a few previous articles, I’ve already mentioned that I work with it quite often. But how does it work from the perspective of digital accessibility?

The answer is: it depends. It’s also worth highlighting that my experiences with chatbots’ accessibility might differ significantly from the observations and conclusions of other blind people. I should also note that the tasks for which I used artificial intelligence primarily involved generating text. Of course, I am aware that models can be used for image and speech recognition, generating captions or music, and performing many other tasks. However, I mainly used chatbots, so that’s where I will focus.

Questions and Answers

Depending on the task type and my expectations regarding privacy, I used various generative models, both online and offline. Not all models are effective for solving every type of problem, and this is definitely noticeable during interactions.

I tested the effectiveness of many models on various tasks: from generating stories and translations to generating code and solving technical problems of varying complexity. I’ve come to the conclusion that AI can make many tasks easier, but sometimes it struggles with things that are obvious to humans. This was most apparent when solving logic puzzles and creating stories.

What do I use AI for most often?

Mainly for translating texts or generating simple content that I can later modify. Chats are also quite effective at checking grammar and spelling when I need help in that area. They can find information in lengthy PDF documents faster than I can or summarize a text. They’re also useful when I need to rephrase something, for example, to give it a more formal tone. While I can handle this task in Polish without AI, when creating content in English, I prefer to make sure that my tone is polite and professional. In such situations, a chat can be indispensable, although depending on the provider, the quality of the content often requires further verification.

Are there errors?

Besides the mentioned logical errors and noticeable shortcomings in so-called common sense, there are also mistakes resulting from the model misunderstanding ambiguous queries or exceeding the context length (more on that below).

Even the creators of these models admit that their products can make mistakes. Therefore, it’s worth knowing how to handle them. Below, I’ve collected a few thoughts I follow in such situations.

  1. I choose the right model for the task. Some are much better at translations or generating code, while others are more effective at data analysis or searching long documents. They also follow instructions to varying degrees, which affects the quality and accuracy of their responses.
  2. I create precise questions or prompts. The shorter and more specific, the better. Ambiguous or general questions reduce the model’s effectiveness. Highly specialized ones do too, as the model might lack certain information.
  3. When a conversation with a chatbot is longer, model errors may result from cut-off context. In this case, “context” can be understood as the working memory of the model. Once it runs out, various strategies are applied. The model may “forget” certain parts of the conversation, leading to frustrating mistakes. That’s why it’s worth breaking a bigger problem into smaller parts.
  4. It’s always worth considering what type of task the model is expected to perform and on what data. Therefore, for more sensitive tasks, I prefer to use offline models, provided the task doesn’t exceed my hardware’s capabilities.
  5. I always double-check. If the model suggests code, a translation, or a problem solution, it’s worth verifying its accuracy in another way. Although the data used to train chatbots is periodically updated, it’s not always the latest. We also have no control over how it was prepared.
  6. Models have limitations imposed by laws or the provider’s policy. This means many will refuse to cooperate when creating certain types of content by default. And while this seems reasonable, especially for online models, it’s not always the behavior I would expect. This is particularly evident when creating fictional content or discussing socially challenging topics. I don’t have to look far for examples. When preparing to write this article and asking an offline model for suggestions, the response contained hints that treated people with disabilities in a thoughtless manner. This definitely results from how the model was trained. You see, for me, being a blind person is completely normal. And the fact that it wasn’t always this way doesn’t change anything. Living independently without a functioning sense of sight might be unimaginable for some. But should it be reflected in how a model is trained? I don’t know, but I have my doubts.
  7. Some models are designed to make interactions pleasant. They sometimes imitate human emotions (sometimes in quite subtle ways). As a result, they will be polite, supportive, tactful, and generally always open to conversation. This can emotionally hook some people or isolate them from others, who can sometimes be rude, angry, or simply foolish.
  8. Some tasks I do faster on my own. If I were to estimate how long it takes to write a post myself and to correct a chat-generated one, I wouldn’t get a definitive answer. In many cases, generating and correcting content takes more time than writing it from scratch. I’ve also noticed that many chats are much better at creating content in English. Posts in Polish are sometimes bland and overly verbose.

What about Chatbot Accessibility?

As with any other application, it depends on the provider. Models generate text based on queries. Depending on the given prompt or built-in settings, they will do this in various ways. What’s important is how the application or website developer prepares the user interface. If done in a friendly and accessible way, communication with the model will be seamless.

Here are a few observations on user-friendly solutions I’ve encountered.

  • On some sites providing access to generative models, I’ve come across very clear information that the model is processing the query. This information is repeated periodically, and when the model finishes, the response is read aloud automatically. Upon request, the response can be re-read.
  • In some desktop applications, a short sound signal is emitted when the response is ready. This is helpful because I can do something else while waiting for the answer. When working with an offline model, this can take a bit longer.
  • The ability to copy responses to the clipboard or export the entire conversation to a file is also a big advantage.
  • On websites, a properly implemented hierarchy of headings or landmarks also helps speed up navigation between different dialogue elements.

Have I Tried Tuning or Training a Model Myself?

It would be a very interesting process, but I haven’t had the opportunity to try it yet. Would I like to? Of course. If I find enough time, I’ll allow myself to experiment. And I’d happily describe the results in another article. In the meantime, I wish you fruitful collaboration with artificial intelligence, because one thing’s for sure—it can be a fascinating process.

Barbara Filipowska

Barbara Filipowska

Audytor dostępności