I’ve written several times about my work on image descriptions for my Image Description Toolkit, My Photo Library Described and Quick Describe apps. Recently I added support for some of Microsoft’s on-device AI technology to all of these apps.
In adding the Microsoft technology I also learned that Microsoft has plans to add additional AI technology to Windows known as Aion Instruct. Today was a great opportunity for me to explore this new technology, and my The Chat Place app (TCP) makes doing these sorts of explorations a snap.
Using TCP I built a small prototype using the new Microsoft Aion Instruct, and ran the same set of pictures through it and through the model my apps use today so I could compare the two. I was a bit surprised when reading back the descriptions with a screen reader to hear unnatural pauses in the description text. For example, in the first picture I had described, there was an unnatural pause before the word “earth” in this text:
The image depicts a simple, serene landscape with a clear blue sky filled with fluffy white clouds. A bright yellow sun is positioned towards the top right corner, shining down upon the scene. In the foreground, there is a green, leafy tree standing tall, its trunk and branches clearly visible. A wooden picket fence runs horizontally across the bottom third of the image, separating the sky from the earth. The colors are vibrant and the composition is minimalistic, focusing on the contrast between the sky and the earth.
I was further confused by the fact that in braille, at that same place in the text, was some braille representation for a character I was not familiar with as far as what it meant. I resorted to a standard screen reading technique of reading by word and character to try and figure out what was going on.
JAWS, NVDA and Narrator all stopped on some representation of text before the word “earth”, as though there was text present. None of the screen readers said anything, and again in braille there was a character representation that I didn’t understand, and I’ve been reading braille since I was five years old.
For situations like this, most screen readers have a feature where you can get the phonetic pronunciation of any character and, if that’s not possible, also the Unicode representation of the same. I learned that one of my mystery characters was 8288, whatever that represents.
Claude, the AI engine I was using at the time, reminded me that Microsoft’s page for the technology I was exploring indicated that text from these models would have watermarking. So I’m presuming this text that is causing mysterious and incorrect pauses in speech and confusing, albeit likely accurate, character representations in braille is just that: AI watermarking.
There was one more surprise. The description above came from the Windows model my apps already call, not the new one. The new model’s picture descriptions in my test were all shorter and never reached the point where these characters appear, but its longer text replies contained the same characters. So I went back and checked older work. The longer descriptions my apps saved through mid-October don’t have these characters, but when I described some of those same photos again today, the ones that came back long enough did. Something changed in Windows on my PC in the second half of October. I run an Insider build of Windows, so I don’t know whether this has reached anyone else yet.
I understand that watermarking for AI-generated text is one of those emerging situations around AI. I’ll confess I’m not fully versed in where this stands on any legal front but, like several, have read the various stories covering the fact that this is either happening or on the radar for the AI world. I’d guess that the accessibility implications here are not getting as much attention. They need to receive more focus, because the experience I had, while not blocking, is at minimum distracting and confusing.
On the technical side, the mystery turned out to be a run of four characters that are meant to be invisible: U+200B (zero width space), U+034F (combining grapheme joiner), U+2060 (word joiner, which is the 8288 my screen reader reported) and U+200B again. In my testing they appear about every 400 characters, wherever the text happens to be at that point, which is how a pause ends up in the middle of a phrase. They aren’t always invisible on screen, either. In a quick test drawing the text in a few common Windows fonts, three of the four showed a small box before “earth”.
I did file an issue with Microsoft on my experience, and hopefully I’ll learn more as I continue to explore this technology. For example, if I know AI is going to do this, as an app creator, is it acceptable for me to strip out this watermarking?
I suppose it is possible that this mystery text isn’t actual AI watermarking. Even if it isn’t, it certainly shouldn’t be appearing. For the record, here is the original text from the AI technology, with the mystery characters left in. Note, it is entirely possible that the copy/paste and publishing process will alter the character representation, so you can also find the original text in this text file.
The image depicts a simple, serene landscape with a clear blue sky filled with fluffy white clouds. A bright yellow sun is positioned towards the top right corner, shining down upon the scene. In the foreground, there is a green, leafy tree standing tall, its trunk and branches clearly visible. A wooden picket fence runs horizontally across the bottom third of the image, separating the sky from the ͏earth. The colors are vibrant and the composition is minimalistic, focusing on the contrast between the sky and the earth.
Comments