Access AI from AIRA Explorer

Last year Intelligence came to ‘Sight as a service’ with the launch of Be my AI – a new feature added to Be My Eyes and powered by GPT 4 from Open AI.

With BE My AI, users can upload images and have them described. Follow-up questions can then be asked about the same image and new images can be uploaded and questioned.
Be My AI can also be accessed via the iPhone’s ‘share’ feature, which is a great way of getting quick and easy detailed descriptions of cat photographs on social media sites like Mastodon and Threads. Other creatures are available. But why would you want to?

Now ‘Access AI’ is in beta testing. Access AI is a new feature in the AIRA Explorer app that is similar to Be My AI, in the same way AIRA Explorer is similar to Be My Eyes – except that using Access AI is also free, a fact that will become important at the end of this post.

I don’t know which AI model powers Access AI, but from the results, it seems to be one of the ‘frontier’ models – either GPT4 or one of the GPT4-equivilent models. But speaking of AI models, one important thing to remember is – these models haven’t been ‘adapted’ to generate image descriptions for blind and visually impaired people. The ability to turn an image into text is part-and-parcel of what the model has been trained to do. This ability is part of what is meant by the term Multimodal. This means these models will continue to be improved, because computer vision is mainstream and improved Accessibility for blind people is just a side effect.

The question has been, are AI-generated image descriptions truly useful? Are they a toy or a tool?

For me, the answer has been that they are far more than a toy, but still less than a tool. Or at least still less than a reliable tool, which is probably just a tool to a philosopher. I am only a student of philosophy, so I am not qualified to say…

The biggest difference between Access AI and Be My AI is that Access AI gives you the option to ‘verify’ the AI-generated image description with one of AIRA’s Visual Interpreters – this returns either an ‘OK’ message or a revised description is around 30 seconds. Obviously this is being done by a person, so it could take longer.

For me, this ability to double check what the AI is telling me is the final step that takes this technology from something that is good to have, nice to have, very useful, great for enjoying ‘Caturday, to a tool.

A useful tool that I could depend upon to do actual work. With the AI-generated descriptions, there is always the chance it isn’t telling you the truth, that you will look silly! I generated a copy of the famous ‘HOLLYWOOD’ sign, but with my name, CHARLOTTE instead. The AI described is perfectly, it sounded great. But when I sent it to a friend who had sight, she told me Charlotte wasn’t spelt correctly. When I went back and did it again, the AI still told me lies – until I challenged it with the truth, then it was very sorry and of course you are correct, the sign actually read ‘CHARLOTEE’ all along.

Now, with Access AI from AIRA Explorer, I could be confident that my image looks like I think it looks – obviously I wouldn’t want this when I am taking a selfie! But it does mean that I am now able to do actual work using ‘Intelligent Sight as a service’ for free, 24-7.

It has taken thirty-five years to get here, the Reading Edge was amazing back in pre-history (the early 1990s) but what we have now is so much better. What will we have next? What do we want next? What will part III of this series cover? The answer is obvious, a blind person could see it!

As we have seen here, a blend of human and machine is the most useful, the next step is to make this technology wearable. After that, I’m thinking implants – I am thinking Borg Queen!