Beyond the Chatbot: Innovations in the Fast-Changing World of AI

by Jonathan Mosen

From the Editor: Jonathan Mosen is well known to Monitor readers, but he now writes in his capacity as Executive Director for Accessibility Excellence of the National Federation of the Blind. In this article he shares his insights about the rapidly proliferating artificial intelligence (AI) tools that help us manage our everyday lives, some of which are designed or have benefits for blind people. Jonathan, sometimes with help from other staff of our Center for Excellence in Nonvisual Accessibility (CENA), shares more insights, interviews, and technology news on the Federation’s weekly Access On podcast, and he and the CENA team are also primarily responsible for managing and developing relationships with key players in the technology industry, including some of the AI companies mentioned here, in order to share the expertise and lived experience of blind Americans. Here is what Jonathan has to say about some of the latest AI tools:

On November 30, 2022, a little-known San Francisco-based company, Open AI, launched a free research preview of a tool called ChatGPT. Within five days, it had attracted over one million users, an unprecedented rate of adoption representing a significant shift in public engagement with artificial intelligence. Users were struck by its ability to generate coherent, contextually-aware responses across a wide range of topics, from composing essays to writing code and engaging in thoughtful dialogue. For many, it represented a major leap forward—an AI system that felt not only useful, but accessible and intuitive.

Since then, generative AI has advanced at a remarkable pace. Models have become more capable, handling longer and more complex inputs, integrating real-time web access, image generation, and even multimodal reasoning that combines text, vision, and audio. This technology is still far from perfect. The abundance of AI-generated content requires us to be savvy, vigilant consumers. AI can still produce plausible-sounding but incorrect information—commonly known as hallucinations.

While AI has delivered tangible benefits to blind people, there are also risks. In the National Federation of the Blind, we know the truth about blindness. It is not blindness that holds us back, but rather outmoded stereotypes that translate into lack of opportunity. Generative AI is like a giant sponge, soaking up all the material on the internet. In that material, you will find plenty of inaccurate information about blindness and blind people. It is critical that we correct these models so they depict blindness accurately. AI is now finding its way into all walks of life, from tools to help make job recruitment selections to tools that help design public policy. So, while a chatbot dispensing inaccurate stereotypes about blind people is bad enough, there are many other far-reaching consequences of AI not being informed about blindness.

We are heartened by the commitment of all the major players in the AI industry to responsible, ethical AI when it comes to the depiction of blind people. In my personal use of AI, it has thankfully been some time since I have encountered an expression of sympathy from a large language model when I told it I was a blind person. Several AI companies are collaborating with services blind people use every day to obtain quality data about what it is like to be blind. The National Federation of the Blind is often at the table when these projects are being scoped, and I am encouraged by some of the products blind people will be using soon because of our direct involvement.

Large language models possess a lot of information. We extract that information by asking the AI a question or giving it some sort of instruction. Designing these inputs is known as “prompt engineering.” Giving the AI a clear, specific instruction can make all the difference to the quality of the output it produces. If you are interested in this topic, the Federation’s Center of Excellence in Nonvisual Accessibility recently produced a webinar on prompt engineering. You can hear or read the transcript of an abridged version of it in Episode 16 of the NFB’s technology podcast, Access On.

Many of us are already using blindness-specific AI tools regularly, particularly for scene or image description, including Be My AI® from Be My Eyes; Access AI from Aira, which offers the ability to obtain free verification of AI’s findings via a professional agent; and PictureSmart, which is built into JAWS. There are intriguing AI tools emerging all the time. Here are a few that illustrate how AI is impacting every aspect of our lives. My mentioning of these tools does not constitute an endorsement by the National Federation of the Blind. I draw them to your attention in case you are curious to try one of these services so you can determine if they prove useful to you.

First, let’s discuss AI’s ability to simplify complex research tasks. Conducting research is time-consuming for everyone, but it may be more frustrating for a blind person because some of the information we want might be on an inaccessible website. New technology offered by generative AI called “Deep Research” can circumvent those accessibility barriers.

The two most common Deep Research tools are from ChatGPT and Gemini. Simply visit the websites or use the app for the service in question. If you have access to the Deep Research tool, you will find a button near where you enter a query labelled as such. Simply press the button to toggle Deep Research on, type your query, and submit it in the usual way.

When you use Deep Research, your AI service of choice explores, synthesizes, and summarizes information from extensive, complex sources beyond surface-level facts. Unlike traditional search engines or basic AI tools, generative AI performing Deep Research can analyze large quantities of data, interpret context, and identify nuanced insights—such as hidden patterns, subtle trends, or meaningful relationships among topics. This process involves the AI going beyond straightforward keyword searches, instead delving into multiple layers of interconnected data, including academic papers, industry reports, historical records, and diverse media. The AI can then output coherent, insightful summaries or detailed explanations that previously would have required extensive human effort and considerable time. Deep Research also provides detailed citations, so you can consult the original sources if you wish.

It is still necessary to exercise caution, because sometimes the summarization process can yield inaccurate results. Yet overall, I have had outstanding results from Deep Research, particularly the one offered by ChatGPT. It is quite extraordinary, even a little eerie, to watch it “thinking” about the task you have set it: looking at sources, reasoning about whether that source is relevant to the tasks it has been assigned and growing in understanding about what it is finding. Sometimes, Deep Research takes a long time thinking about the task it has been set, but it is usually worth the wait.

Recently, I was in the market for a new smart scale after the company producing my previous scale went out of business. I gave the Deep Research functions of ChatGPT and Google’s Gemini the same extensive prompt, telling them that I was seeking a smart scale that had compatibility with Apple Health, an app accessible with VoiceOver on iOS, the ability to collect data well beyond weight such as body fat percentage and hydration, and if possible provided some sort of haptic or audible feedback when the weighing process was complete. Both Deep Research tools went to work and produced comprehensive reports, with slightly different recommendations. This process saved me considerable time. I didn’t have to perform repeated searches and try to keep all the information in my head or take copious notes.

In another example, I asked Gemini’s Deep Research to produce a guide on using the Ecobee thermostat that my wife Bonnie and I have in our new Baltimore apartment, mindful that I use VoiceOver on an iPhone. This can be a useful strategy for getting information about new appliances, because many online or printed manuals use graphical symbols when describing buttons to press or steps to take.

ChatGPT’s Deep Research feature is excellent, but there is no free tier available. If you subscribe to ChatGPT Plus for $20 per month, you get ten Deep Research queries per month. If you haven’t paid for Google Gemini, you have limited access to Deep Research. Google appears not to have published the specifics of how many Deep Research queries you are allowed on the free tier, but users suggest it is around five per month. If you subscribe to Google One, which includes access to Gemini’s advanced features for $19 per month, you seem to get many Deep Research queries with your plan.

Another useful research tool from Google is called NotebookLM, available at https://notebooklm.google.com. NotebookLM allows users to organize, summarize, and interact with personal research materials and documents, enhancing productivity and understanding. It has not always been the most accessible of sites, and we would like to see accessibility improve further, but the site has improved since its initial release.

Users begin by creating a digital notebook and uploading up to fifty sources—ranging from PDFs and Google Docs to websites and YouTube videos—with a combined limit of 25 million words. NotebookLM then uses artificial intelligence to summarize documents, generate concise explanations, and respond to queries based on the uploaded content. Let’s say, for example, that you are interested in accessible televisions. You’ve watched some YouTube clips, read a few blog posts, and looked at many product descriptions. Once you provide NotebookLM with all this content, you can ask questions about it as if you were talking to a full large language model. The answers it returns will be based solely on the material you provided. As an experiment, I created a new notebook and uploaded every Presidential Report delivered by President Riccobono. I was then able to ask questions about the Federation during the Riccobono presidency and get accurate results.

One of the most intriguing features of NotebookLM is its ability to generate podcast-style audio overviews. These take the form of two AI personas, who sound quite humanlike, discussing the subject matter in question. It’s easy to download this audio and copy it to a device such as a media player or notetaker. It’s a fun way to distill complex topics. The summary it produced of the work of the NFB was engaging and accurate. Google has also added the audio overview feature of NotebookLM to Gemini, so you can upload content such as documents and slides to Gemini, then press the “Generate Audio Overview” button.

Finally, let’s look at the impact AI is having on shopping, and a product specifically designed for blind people. Featured in Episode 8 of Access On, Innosearch is a platform designed to make shopping easier for blind people. It aggregates over 500,000 retailers, from large players like Amazon and Walmart to smaller niche stores. Simply enter what you want in the search field, and Innosearch will come back with matches from a variety of stores. This has accessibility benefits, because as we all know, some stores are more accessible than others. Innosearch prioritizes results based on the best price.

Another powerful feature of Innosearch is the ability to query AI about the products it finds. It will also describe pictures of the products, and you can ask questions about those pictures. For example, when shopping for certain audio cables, it was very important to me to know precisely what type of plug was on the cable. Innosearch was clear about this and ensured I got the correct cable the first time.

There are some downsides at the moment. While Innosearch is an accessible, efficient experience, you may not be able to leverage the benefit of store-specific programs like Amazon Prime. Amazon features like Subscribe and Save are not available. These are tradeoffs that you may be willing to make, at least in certain circumstances, for an experience that is very easy to use.

Customer service is excellent and responsive. Recently, I ordered a specific MagSafe battery for my iPhone that shipped from China. It took a long time to arrive. Innosearch was monitoring the purchase, noticed it was taking a while, and followed up with the retailer without me having to ask. That was exceptional service.

On the website, Innosearch uses its technology for two product categories—shopping and flight search. The company is considering adding new categories in future.

The most exciting use of AI from Innosearch is a version of the service that can be used over the phone. To use it, call toll-free 1-855-ShopGPT (855-746-7478). Shopping hands-free with your voice is beneficial to everyone. But this is a particularly compelling use of AI because most blind people become blind later in life, and some don’t wish to learn to use a computer or smartphone nonvisually.

This is an excellent concept in early development, with a few wrinkles at the time I am writing this article in April 2025. Sometimes, speaking the name of a product you’re searching for yields relevant, clear results. You can ask to hear more information about any of the choices offered, add one to your cart, and check out.

Other searches I tried were less successful. When I searched for the Zoom H6essential Handy Recorder, Innosearch thought I was saying “zoo” not “Zoom” no matter how clearly I enunciated. It also stripped part of my search, so it came back with “searching for Zoo h handy recorder,” which didn’t deliver the results I wanted. Hopefully this will improve over time, as it’s great when it works, and opens online shopping to a neglected market segment.

We’ll have more to say on new AI tools in a seminar at national convention in New Orleans. We hope to see you there. The National Federation of the Blind is monitoring developments and having input into much of this innovative new technology. For further information, subscribe to Access On wherever you get podcasts.