With the advent of AI-generated images, and the exponential growth of the usage of graphic content for everything digital, specific challenges and opportunities lay ahead for Persons with Disabilities. Panelists from advocacy organizations and companies involved in leading digital image technologies and applications will brainstorm and share their perspectives on how AI-generated images may impact Persons with Disabilities, including but not limited to their potential influence on disability representation, how AI-generated images or graphics may contribute to cognitive accessibility solutions, and how much progress is achieved in embedding AI-generated alternative text. The panel will also explore how to ensure the accessibility and compatibility with assistive technologies of emerging solutions to authenticate real Vs AI-generated images and of AI-enabled digital rights management solutions.
Session Chair: Jennison Asuncion, Co-Founder, Global A11y Awareness Day & Vice-Chair, GAAD Foundation
Panelists:
- Andrew Kirkpatrick, Director, Accessibility, Adobe Systems
- David Berman, CPACC, CPWA, ADS, President, DB Communications, and Author
- Christopher Patnoe, Head of Accessibility and Disability Inclusion, EMEA, Google
- Saqib Shaikh, Software Engineering Manager, Project Lead for Seeing AI, Microsoft
- Susanna Laurin, CPACC – IAAP Honorary Chair & EU Representative
This video is lacking captions. We expect captions by February 14, 2024.
Transcript
FRANCESCA CESA BIANCHI: Good afternoon, everyone. Welcome back to the ballroom. It is my pleasure waiting for the ASL, one moment. Okay. It is my pleasure to introduce you the next panel discussion, AIgenerated images: Challenges and Opportunities for Persons with Disabilities. And the Moderator is Jennison Asuncion, a cofounder of the Global Accessibility Awareness Day and ViceChair of the GAAD Foundation or GAAD Foundation. Thank you very much. The floor is yours, Jennison.
JENNISON ASUNCION: Hey. Good afternoon, everyone. Again my name is Jennison Asuncion, cofounder of the GAAD Foundation. The mission of the Foundation is to disrupt the culture of technology and digital product development to include accessibility as a core requirement. And I warmly invite all of you to join us in San Francisco, November 9th. We are having our first GAADies award celebrations. We are honoring three companies, Equalize Digital, Unilever and Zapper as well as Studio 24 as three organizations that are living and breathing the mission of our Foundation. And to learn more about the Foundation and our work we're doing, please check out gaad.foundation. So as Francesca said we are here as we have been for the last day and a bit focusing on AI. And the focus here for this particular presentation is on AIgenerated images. Now as someone who is blind I automatically, I know some things about AIgenerated images in terms of alternative text. But this panel is going to talk more about this. We have an amazing panel here. And I'm going to ask my esteemed panelists to have a oneminute introduction. We have one of our panelists who is joining us remotely. And I'm crossing my fingers that he is online. We're going to start with him first to do a introduction.
SAQIB SHAIKH: Thank you very much. Thank you for accommodating me remotely. Yes, unfortunately I could not be there in person due to unforeseen circumstances. But I am Saqib Shaikh. I lead the senior project to Microsoft. And we will hear much more about that later. But we're looking at how do we bring emerging technologies like AI to empower people with disabilities to do even more.
JENNISON ASUNCION: Thank you. We're going to go down the line to my left.
ANDREW KIRKPATRICK: Thank you. My name is Andrew Kirkpatrick. I'm the director of Adobe's centralized team. To help ensure accessibility across all of Adobe's offering both for customers and employees. In addition part of what I have done to introduce myself is that I have also worked on standards extensively including being the coChair and coeditor for the WCAG 2.1 standard.
JENNISON ASUNCION: Thank you, Andrew.
DAVID BERMAN: I'm David Berman. I'm an inclusive designer. Actually I'm a graphic designer with a core deficit. For two decades I was scared to mention that. I served as the ethics Chair for graphics designers. I designed their code of ethics. My book which is now available in seven languages talks about how we can work together to use design to design a better civilization.
CHRISTOPHER PATNOE: My name is Christopher Patnoe. I'm the head of accessibility and disability inclusion for a little startup called Google. And EMEA. My responsibility is to try to understand the needs of people with disabilities across Europe, Middle East and Africa. Making sure we are placed where people with disabilities can thrive and mess around with different kind of legislation and toys and things like that.
My name is Susanna. I'm the only one on stage without notes. Everyone else is going to know what they are going to say. I'm the IAAP representative to EU. I'm also chairing this standardization committee that tries to make sure we have technical specifications for the act. The research part is why I'm on this panel or make the gender equal.
JENNISON ASUNCION: I wanted to thank the folks at the Federation of the Blind, those blind folks in the audience, Braille is in my hand. And Braille is still alive and well. Thank you.
(Applause.)
JENNISON ASUNCION: All right. Let's that was my political statement. Let's get going. David, talk to us about the potential positives and challenges in working with images generated through machine learning.
DAVID BERMAN: For sure. Thanks, Jennison. I feel the main challenge we all face as accessibility specialists is simply not to repeat the past mistakes in human history where we have left people behind when crafting new tools. But let's start with the possibilities, because you asked me to think of some examples. So I asked myself what becomes possible when we engage machine learning on the future of human visual communications.
And as a designer, I think of images. I think of illustrations but I think of photos, emojis, logos, sketches, even those collections of clip art that we call typefaces. And so consider this, imagine you get a set of instructions from IKEA or from Legos to assemble a toy or a toy chest. And imagine the instructions you get are personalized. Magically the diagrams, wording illustrations, the color schemes are all personalized for you based on what you have chosen to share about your vocabulary. Imagine in this very event, I'm seeing the captions in front of me, and imagining imagine if the captions available of my words right now was beamed to a personal device you have and it would be words or emojis. And maybe personalized just for you and your personal dialect based on maybe what you have chosen to share of everything you have written or signed or sketched in the past ten years. That personal corpus would inform what you get. It will reflect your culture, vocabulary.
Now that may seem crazy but imagine like an AI native fouryearold going to kindergarten for the first time in 2024. Those kids are growing up in a world where there never wasn't generative AI. And they're going to pick that up in at the same speed they pick up TikTok. And those captions I'm describing, I'm thinking of the emphasis that goes beyond tags. I'm thinking every word could be in a different typeface, a different color, visualization has no limits. So imagining those two guys from Brisbane, that Michael and James who invented MVDA. Imagine them drafting, sketching the next logo for MVDA.
Imagine, picture this, generate an image in your mind, if you will, of a whole new norm of visual literacy for people with cognitive disabilities or people's whose first language is not a written language or maybe for everyone. Jennison, imagine just like we just like we need to have alt text for every image. Imagine a world where we have alt image for every text. Would that be cool?
JENNISON ASUNCION: That's great. Thank you. As it relates to designers and creators, can you talk about some of the professional and ethical issues that those folks need to contend with when it relates to AIgenerated images?
DAVID BERMAN: Sure. How would the role of a designer be transformed? And can we do it in a way that champions accessibility like never before? This is not the first time that designers have been terrified. Remember in the late 1980s when desktop publishing arrived. They were afraid for their craft and jobs. And understandably they were scared of like amateurs making lost cat posters and making a mess of the design world. And it took a few years. But within a decade, the quality of design globally at all levels went way up. Kids had favorite fonts. And graphic design became a household word. But it starts way before then. We have had perhaps 7,000 generations of humans. But we have only been recording knowledge for maybe 5,000 years. If we go back to Gutenberg. And this is a Gutenberg moment. Because when Gutenberg invented the printing press, he was trying to drive costs and the graphic designers of the day, these monks who would hand letter and generate every image one book at a time. They were up in arms. But meanwhile, Gutenberg had given the people of Europe a reason to read because books could be affordable. And yet so many people are left behind.
If we go if we come to the 1930s when telephones became ubiquitous in America and distance communication left out the entire deaf community. And it took us three decades to invent the teletype. Even in the 1980s when the GUI came out, nonvisual people employed with graphic user line, but the arrival but being forced into it and it wasn't accessible, it took over a decade for the GUI to become accessible. And seriously slowed the employment amongst people with disabilities. I think this time could be different. Because this time for the first time, standing on the leadership of amazing people in this very room we can actually create a technology in a way that includes everyone in the first place. In fact, I think AI is the largest opportunity yet to include everyone in the evolution of how humans communicate. Is it going to transform what it is like to be a designer? Absolutely. Is it going to be scarey? Sure. And it is going to change IP rights and dark patterns and all that.
But most importantly for me, it democratizes design like never before. And I'm not talking about the output. I'm talking about people of all abilities to IDA, to sketch, to test, to brainstorm. 95% of the designers who have ever been alive are alive today. I think we have the power to include accessibility in the source code, include accessibility in the source code of how we unleash it. So I say bring it on.
JENNISON ASUNCION: Wonderful. Thank you. Andrew, Adobe Firefly has some accessibility features that are of particular interest to the subject of this panel. Could you talk a little bit about that?
ANDREW KIRKPATRICK: Sure. So if people aren't familiar with Firefly, Adobe Firefly is a set of generative AI models that Adobe has created and made available. And it allows you to create text that follows a pattern that you identify. So if you want text that is showing a jungle pattern, you can specify that. And it will and you indicate what text you want written, that pattern and it will give you something creative.
And you can keep trying and modifying that. But it will also do generative images based on a text prompt. And I think what is most interesting about this for me is how to David's point it's giving us an avenue that we haven't had before to democratize creativity. If you are someone who is blind or very low vision or you can't use a mouse or a pointing device today and you want to use Adobe illustrator or a different tool, you are going to have a very, very hard, if not insurmountable task to try and do that.
But with Firefly you can simply type in a prompt and you will get images that you can then refine. So that part of the creative art of, you know, part of the mechanics of creating art is how you describe it. And how the AI model interprets that description in order to create something for you.
And you can keep iterating on that. And we're doing work to add additional models and additional tools so that you are able to modify images that you already have or to modify an image that you have already started creating. So if you have generated, for example, an image of, you know, a puppy on a beach, and then you decide well, this beach also needs some see gals. You can identify a region and say I would like an island out there. And it can add those features for you. It is enabling access in ways that have not only been possible to date, which is incredibly exciting to think about.
As far as Firefly and the accessibility features within Firefly, at Adobe we have three different principles that we think of that guide our accessibility work. And those are partnership, transparency and innovation. And when really fits into Firefly fits into innovation very well. But the partnership is that we want to really make sure that everyone can be involved in developing our tools, but then also in testing these tools.
And if we're going to have generative images we need to make sure that these stand up to scrutiny from people who are trying to generate images. We don't want inappropriate images to be generated and inappropriate covers a lot of ground in terms of generative AI. In terms of this particular audience is making sure we're representing disability and representing disability well.
And I can assure you that there is more work to be done in this area. But this is part of the reason why it is so important. It is another reason why it is so important to make sure that the tools itself, Firefly is accessible for people to be able to use, whether they are screen reader users or keyboard users or any number of different types of, you know, disabilities that a person might have.
JENNISON ASUNCION: Great. Another area that Adobe is focused on is the content authenticity initiative. Can you tell a little bit about that?
ANDREW KIRKPATRICK: Yes. One of the big challenges that we have in the world today is and with generative AI images is that you can take something that is a bona fide actual image that someone has captured with a camera and change it. We have image deep fakes that can be happening. And we're making it ever easier to make these. And part of what's exciting about and very important as far as generative AI is the idea to go back and forth between text and image and images and text. So I'm counting on AI to be able to help make sure that end users can get access to information and potentially to be able to query additional information from an image through AI. So that if there is an image that's shared whether it is on Instagram, Facebook, email, a user can find out deep information about that. And the risk is that a user may be told something that's an image and told by an authoritative source, this is absolutely a picture of the President of the United States doing something that you want the President of the United States doing. And the question we all need as citizens, as people is to know is what we're hearing true. And this is where content authenticity initiative comes, because what Adobe has been doing with this initiative and it is a coalition of about 8 or 900 different companies around the world that are collaborating to have a system to add imbedded metadata to images. And what this metadata does is it captures the Province Prominence of that image. So that if you if you are questioning whether an image is true or not, you now have metadata, you would be able to look at. And to be able to tell okay, this metadata in this case has been modified in photo shop or modified with Firefly. It gives end users to determine for themselves whether to trust that resource or not. And that's one of the main things that we really want, you know, as much as we have nutrition labels, when we buy food, we want to have a nutrition label for images that we're consuming online because we want to know what's in that image. And was this is this a fresh vegetable that was taken from the field. And it is, you know, pure and healthy, like a photograph that comes directly from a camera. Or is it something that has been changed in different ways. And maybe doesn't actually exist. So I think that's this is where the content authenticity initiative is going to help us, you know, for all citizens. But it is very, very important to my team and to Adobe to make sure this is helping make that this is possible for all users to be able to get access to that type of information.
JENNISON ASUNCION: Great. Thank you so much. Susanna, you have worked on the We For Cluster project. Project that I just learned about by virtue of you joining the panel. But could you share with the audience what the focus of that was as it relates specifically to images?
Susanna: Yes. So thank you. So the We For Cluster We For Authors Cluster was a but it was we were building a cluster of content management system producers or suppliers and also editors. Trying to make sure that the industry players work together to solve accessibility kind of at the core instead of doing remediation afterwards. It was an EU funded project. We did loads of things and we the heros of that were the suppliers.
One of the things that we wanted to do was to make sure that the Web authors or the content creators get more support from the content management systems. And one of the criteria we were trying to work on was the alternative text that if people know one thing about accessibility, that tends to be because it is a first criteria of WCAG, they read that and fall asleep or give up. That's why this is the most famous one. We worked at that time, it was a couple of years ago, we didn't have the strings and muscles in what is now the image recognition. But still the potential was there, of course.
And what we did see was that many of the Web authors we worked with, they were very negative towards having an auto generated image. Especially the ones that knew anything about accessibility. They said no, no, we need to do it. I need to feel what the purpose or image is. It became an artistic discussion of things. In reality most web authors don't know or don't care. They don't provide alternative text. It must be the volume here is what we want to achieve, raise the quality of.
When we started providing these auto generated image descriptions and we did user testing with all these web authors from across Europe, we saw if they got a kind of half bad alternative description automatically they went into change it because they didn't want the bad description to be there. That seems to be a human thing. That's not correct. That's not what I want to say with this picture. So then they did change it. If we didn't generate this alt text, auto generated alt text for them, then a much larger percentage of these users didn't care. So they just left. And they had exactly the same setup and same number of steps. So that could be used maybe not yet perfect. Maybe in a couple of years it will be perfect.
But since I'm working in the EU we have 24 languages. Many of them are very small. Even in English I think they are not yet perfect. So and you may want to add your personal feeling to it or whatever.
But still we can use it to nudge people to do something. So and I like the idea of using AI to do part of the job. But then add the manual aspect of it. That doesn't scale as well as doing it completely automatic. That's how I see AI being a tool helping you halfway or 80% or 95%. And then that could kind of support you to get over that barrier. If we could get more web authors to provide alternative text, I think we have come a long way. At least with that small proportion of what accessibility is all about.
JENNISON ASUNCION: Some folks with cognitive or other disabilities benefit more from information that's presented visually versus textually. I know you have some thoughts about that. Could you share that with us?
Susanna: I'm with David on the alt image. According to OECD in Europe, 20 to 25% of the grownup population in the European countries cannot read an Article in a newsletter, new in the news magazine, paper and answer controlled questions afterwards. 25% do not read well. That could be because of cognitive issues, reading and writing, dyslexia, not being native speakers, bad schooling. All sorts of reasons for this. The reality, we have a quarter of the population that we want to communicate with have issues. I found it very a lack of fantasy to say if a person had a problem with text we'd just transform the text and provide another text. Can't we do something else. It seems a little boring. We did a huge research project many years ago, funded by several different funding instruments and we did have in different batches with end user organizations. And trying to make sure that professional people who are experts in writing plain language or EasyToRead, we tested both and tried to work with text as much as possible to see how far we could get.
If we had a proportionate cohort of people reading, so people who claim to be good readers and then all the other readers in the group proportionally spread out and different ages as well, then we couldn't get more, reach more than 50 to 60% of the readers with text, no matter what we did with the text. We couldn't get over 60%. So and when we added images, or illustrations or diagrams or videos or something visual, 20 to 30% more people understood this. And we added audio and this was another 20%. That's really the way to provide as good accessibility as possible. So we are really promoting to do not do only one thing. To me accessibility is about choice. So don't go with text only. Do text and something else. And try to do that as much as possible.
And where we have seen this most recently with when it comes to this AI connection here is with communicating with people with autism where I have a friend with autism who said, always claims that she thinks and speaks in pictures. And in the beginning when I got to meet her she didn't write at all. After a while we kind of invented a way of writing emails to each other using images. We have stolen a lot of images from the Internet. But it is only two person communication. There are better free sources out there as well.
We communicate nowadays with very few words and a lot of images. And it is absolutely fantastic how this has helped her communicate with me and with many other people and doing presentations and being in panels and so on. This has given her kind of a tool, a way to communicate. And we have tried this also in research with other people with similar impairments and it works really, really well. There are so many perspectives on this. People with intellectual disabilities where you can work with icons and illustrations and with autism, kind of personal images. But if AI could support people to create images or find images even that are free to create their own personal way of expressing themselves this would be a fantastic tool. I haven't seen that happen yet. The technology is there and somebody needs to put these things together. And we would have something that would support communication for I think a wide range of users.
JENNISON ASUNCION: Thank you. I suspect there is at least one or two really, really smart people in this room always talking about wanting to do a startup. There is your idea there.
Susanna: I would support it.
JENNISON ASUNCION: Christopher, can you share your perspective of people with disabilities on AI models today?
CHRISTOPHER PATNOE: If I may, the UK recently announced a plan, experiment using AI to allow people to express themselves. We are starting to see some interesting things in this space. Take a look. Back to your question about the inclusion and representation of people with disabilities, Andrew, Jeremy Andrew Davis a couple of months ago did a TikTok when he went into VID and showed me pictures of people with autism.
And he did it a hundred times and a hundred times they were young, white male, dower, dark, depressing, not representative of the community in almost any way.
And this is to do with everything that Jutta was talking about today, if you didn't have a chance to see Jutta's keynote this morning, it touched me in a really profound way. What I got out of this is a deep appreciation of the challenges that we have to do this right. The opportunities are tremendous but we have to do it creatively, thoughtfully, progress to trust. Is really, really important.
But how do you get there is really the interesting question. Because if you don't engage because you are scared, we are never going to be able to move us forward. So we have to learn how to work together thoughtfully, respectfully to create the models, to create the datasets, to have some that allows us to learn together, how to make this representation, to make the ideas possible.
So what we need to do is we can't be afraid. We have to get engaged. Because it is already wrong today. I don't know who made it. It wasn't us because we don't really do that kind of thing. But it is there. And what's interesting to Andrew's point about the voracity of images in the future when it comes to having valid, this is cool for the people who participate. But if you are a nongovernmental actor trying to cause problems, they are not going to add this framework and say is this valid or not. We have to be thoughtful of how we create imagery that represent us. How we develop technology for us. Because it is wrong today. But we have to engage to be able to make it right in the future.
JENNISON ASUNCION: Thank you. Just extending the conversation and talking a little bit more broadly, what is your perspective in general on the evolution of AI technologies?
CHRISTOPHER PATNOE: Taking this sort of upper level from generative AI, this has been around since the 1950s. It has been evolving for generations. It has been changing. But what's really exciting we are at this point whether the change is fast enough and the technology is strong enough. And the form factors are evolving fast enough that we are going to create a completely new way of interacting with each other.
So for me as an example of what I get most excited, I have done a lot of work in VR. What I'm excited about is VR is having a screen reader for threedimensional space. There is no DOM. No tree to describe. You have to actually see a tree and describe what the tree is. But you need to describe the tree in time, in context. If someone standing behind a tree at a game you don't want to talk about because it is probably part of the linear narrative. Now you have a story line that you have to include as you come up with this screen reader. Then you take away the virtual world and you have a real world wearing a pair of glasses. Now we have a computer vision.
So we learn in VR. And we experience the real world. But now we have a high performance location. We have 5G. We have realtime accuracy and, of course, the permission and willingness to share this information about yourself. But now we have the opportunity to help someone who is blind explore the world and browse. Go get lost. This is what the excitements of AI really is. It is the ability to do things you couldn't do before, because you are being supplemented with technology that was thoughtfully designed and created in concert with the community, and then we get to have fun.
JENNISON ASUNCION: That's great. Thank you. Saqib, seeing AI is a much loved and much lauded app within and among the blind community. And now before you get a wave of new downloads after you talk here, could you talk a little bit about what seeing AI is for the uninitiated and features that relate to AIgenerated images?
SAQIB SHAIKH: Thank you so much for the kind words. Absolutely. Yeah, why not. Going to seeingai.com and download. Sometimes I think of it more as a mission. I'm blind myself. I have a vision about what if I as someone who is blind had a wearable device that would let me know who and what is around me and know my personal preferences, to know what's important to me at the moment in time. What's changed. Highly personalized, wearable. We're not there yet. But I have been on this journey towards that for a number of years now.
And that's for me as someone who is blind. I sometimes describe it as a friend on my shoulder, whispering in my ear. They can span all of humanity because we are all different. For someone who is hardofhearing, for someone with a cognitive difference or just someone who is, you know, new situation, or new language. I cannot believe this idea that AI can bridge the gap between what we are able to do right now with our capabilities and what we could do.
So this is why AI excites me. That's why we have been on this mission for a number of years. And seeing AI app is one aspect of that. So we sometimes talk a bit as a talking camera app for the blind. It is a conversation with the blind community on the one side, the scientists, what are the emerging technologies we can bring to bear. And an important one has been describing photos that you take on your phone, that you are experiencing on social media. And all in fact, from anywhere else. And the technology we develop in the mobile app have gone out to other Microsoft products like powerpoint and Word.
Now this idea of automatic alt text has become mainstream. What's the latest we can do to empower the blind opportunity to do more? It is providing these extra tools. Now age description is one part of it. Talk more about that in a moment. We have been looking at how do you let blind people teach the system so they are more included and how do you use augmented reality to help someone navigate in a 3D space. Let's focus on image descriptions which I think is just it is a gamechanger. And with the recent enhancements in generative AI, we can do so much more.
JENNISON ASUNCION: Thank you. Now you clearly have deep expertise in generative AI. Could you elucidate further and share your perspective on the potential of and the positive impact it could have on the lives of folks who are blind?
SAQIB SHAIKH: Yeah. So I remember like about seven years ago I first heard automatic AI image description. That was a gamechanger. And the descriptions have gotten so much better. But I feel that gen AI is this step change. And from that, generate really, really rich descriptions from images. And that really empowers someone who is blind to get a lot more detail about photo or something in a textbook on a website. But also doesn't have to be one size fits all because at the end of the day, the alt text you want while reading a Web page is probably quite different to the one when you are reliving your family photos. Or when you are using interactively using a desktop graphical user interface. These are all quite different. So one of the things that generative AI gives us is this ability to have more personalized custom descriptions, based on the scenario that you are in. As taking that one step further, it is not just alt text where this rich description. You want to know more about a particular aspect. What was on the shelf this photo. I wanted to know about this aspect. So again also is now putting someone who cannot see into the driver's seat of what the AI is doing. And that's really important because AI is a tool. Humans should remain in the driver seat. Because it is not one size fits all. We are all different and what we need at different times can vary wildly. And yet I have to also mention, there are many challenges. We are at the very beginning of this very, very exciting journey. It is a real step change in what the capabilities are. How do we make sure it doesn't confidently tell you the wrong thing which is something that we have seen. But overall, the opportunities are huge whether it is regular photos, alt text on the Web, but also in work and education, mathematical graphs and diagrams. Because these things have been seen on the Internet and often wellknown formats, like textbook diagrams. I'm optimistic the potential this can have across different domains. I as a blind person myself I'm just incredibly excited by where this technology is going to take us and this new wave of upcoming months and years.
JENNISON ASUNCION: Thank you. I want to personally thank you all, my panelists for staying within their time. We're going to open things up for questions. When you do have the mic that you say your first name, and who your question is directed at on the panel. So do we have any questions?
There are a few hands up. Microphone is coming out.
Hello. My name is Kelsey. My question is for Chris about your comment on inclusion and representation and also Andrew about your content authenticity statement. So like I said I'm an accessibility UX researcher. And I recently ran a moratorium study with people who are neurodivergent. And it is all about their thoughts and trust and excitement and hesitation of AI. A majority of them commented that they fear a targeted discrimination because of an assumed vulnerability for having this disability. This was especially surrounding the theft of their original works. Many of our participants were artists, writers, creatives who rely on using imagery and art to make a living. So my question is, how do you or how do you plan to make a specific effort to protect this population, to perpetuate their creative autonomity while continuing to represent them? How can you relay that so that the disability community knows about it and are aware of these efforts?
JENNISON ASUNCION: Thank you.
CHRISTOPHER PATNOE: What a question. The challenge is understanding the depth of the problem. We can't fix what we don't know to be broken. We have to identify situations where things are being done wrong and the cause of it, and work together to try and solve it. It takes intention. It takes effort. But unfortunately it also takes effort from every single provider of these AI technologies. It is this multifactorial problem. This is really hard. This isn't something that a Google or Microsoft can solve. It needs some kind of conversation what is the right answer for this. And that it could be applied across the different organizations that are making these models. It can't be one company making these changes.
JENNISON ASUNCION: Thank you.
ANDREW KIRKPATRICK: Thank you for the question. As far as what the content authenticity initiative would do for these individuals, if they're using a tool that supports the ability to add this information, like Adobe illustrator then they would have to have that embedded metadata that's in there, that's basically a digital signature that is indicating that it is, you know, an authentic resource. And someone still could take that image and turn and sell it to Jennison and say this is mine. Jennison would need to know to go and look to see whether there is any credential information that gives information about, you know, where this image came from or whether has since been modified. So, you know, there is layers of problems that are part of this. Related to how do people know about this. I think for accessibility features and features that impact end users with disabilities we need to do a much better job of conveying and communicating what those features are and the impact of those are. And I think the more that we have end users with disabilities that are a part of our process of development, the more that that happens in a, you know, natural way, to get started. And then requires further work on top of that.
JENNISON ASUNCION: Thank you, Andrew. Any other questions?
Yes. Larry with Booz Allen Hamilton. The Freedom Scientific product, JAWS has a feature called Picture Smart. When you come across a photograph you can initiate a description to be generated. And in that description if there are familiar people or things they are announced and what's generated. I'm wondering if there has been any thought to standardize some sort of shared library of photos that might be like descriptions that might be able to be used across technology platforms and other types of devices that would really follow that sort of paradigm?
JENNISON ASUNCION: Who from the panel are addressing that question to?
Pretty much all of you.
JENNISON ASUNCION: Let me ask Susanna maybe if you had some comment on that.
Susanna: I have no not any initiatives covering that.
JENNISON ASUNCION: Okay. Saqib, did you hear the question?
SAQIB SHAIKH: Yes, I did. And, you know, it is a great idea in the sense that, you know, how do you make what these models can recognize more standard. And certainly they will require standards, policies, make sure that we're conserving privacy. Up for a challenge always. Can we get all the players of the world to come together to have this way that the things you want to have recognized in your images can be recognized regardless of what kind of AI the software you are using leverages, if I understand the question correctly.
CHRISTOPHER PATNOE: There is an interesting feature in Google Chrome called Get Image Descriptions. This is a platform solution. The different platforms you can get these kind of image descriptions today. And these have gotten better over time. But to Andrew's point, people don't know about these features. So we could have the world's greatest features. If nobody knows about it, they are no good. In the future you can take a look at these other solutions that exist. And see how these work. But then provide feedback to us in terms of how these images, these descriptions that are generated are they good, helpful. And if not, let us know so we can make them better. These features exist to serve but they are only as good as the feedback as we get.
JENNISON ASUNCION: David or Andrew, do either of you have anything to add to this?
I will just back up what Christopher said, I think there is the strong possibility that different accessibility tools, different, you know, browsers will be using their ability to generate great descriptions and great alternative text as a feature. So that it's I'm not sure if Larry, your question was so much about storing alternatives that have been generated in one circumstance that can be used by others. And apologize, that's not what you are intending, but I think that Chrome would undoubtedly like everyone to feel like it does the best job of generating rich descriptions. And similarly, you know, Adobe would like to be known for that as well.
I'll just add one thing, completely agree because we have been playing with that feature in Chrome and watching it get better and better. I would like to see it get interrogatible. Tell me more about that cat and maybe even crowdsource the truth about it.
There is an application we are seeing called Lookout and a feature we added where you can actually share an image. And it will give you a generative AI description. And then can ask what kind of dog is it. Are there clouds in the sky. These technologies are coming online right now.
JENNISON ASUNCION: We have time for a couple more questions.
Hi. Mary Fernandez here. Cisco Systems. And I am a blind woman and so I have a story and I have two questions. So, you know, I have been a lifelong advocate of education. I have been through higher ed without access to educational materials. The power of leveraging these technology is not lost on me. I want to bring it to the things that really matter. I got unhinged the other day and started taking screen shots because unlike I have the right to be as shallow as everyone else. And so and so what I have been doing is I have been going through profiles, taking screen shots, run it through Be My Eyes description. My thing is for y'all there is a great business opportunity here for dating apps, integration with AI description. This is free advice. I will take any proceeds that you may come across. But I guess so two questions. And I think they have been partially answered, Christopher kind of segue wayed into that. Part of equity is play, right? And the world is so rich with images around fashion, around eating, dating apps. All of those things are really important to how we are our society right now. And so how do we think about play and the things that aren't it is not just about like it is part of the systemic inequities that we experience is that piece. My second question is how do we manage infantilization with individuals? A lot of times with Be My Eyes we had the whole discussion of blurring out faces because of privacy. That's not equitable inclusion. How do we manage that bias? What are we doing about that?
JENNISON ASUNCION: Two questions.
This is the third one. AI recognizing AI created images. I know that visually there are things that you can look for, like hands and that kind of thing. But how do we leverage these technologies to recognize pictures that have been made by AI? Thank you.
JENNISON ASUNCION: All right. Christopher, since Mary answered, do you want to elucidate a little more?
CHRISTOPHER PATNOE: This is why the partnership is so critical. We are humans with things that we love and don't love. We want to play games. We want to have selfies and post them. By partnering with communities we understand what's important instead of the stuff of someone who is not part of the community would think is important. Knowing how your shoes are is really important. Because is an aspect of who you are. And how you express yourself. If we can't show you that, we are not doing you justice. We don't know to show you that if we don't know to work with you to know the things that are important.
JENNISON ASUNCION: Anyone else?
I wanted to speak to that as well. Because something we been talking about is interrogate an image. What we can do with generative AI and with Firefly is you can replace objects in there. So you don't want books on the shelf. You want to have, you know, a bump of trophies, no problem. You can do that. If before this talk today I took a picture of one of my friends and, you know, highlighted a section and indicated that I wanted to have a monkey on her shoulder. And just like that, it was done.
Now if she is going to take that picture and put it into her dating app because she is trying to impress some local monkey trainer he is going to want to have content authenticity information on there so he knows it is a real image. Allow people to trust what's real and what's not and to have fun. Because we don't want everything to be real. Sometimes we want to have fun.
JENNISON ASUNCION: I'm not sure, Andrew, if you want to within the context of the third question that Mary asked about bias and how to manage that.
ANDREW KIRKPATRICK: The third question was about detecting whether have been modified. And I think that our belief right now is that we're not going to be able to. I could be wrong. I'm not they didn't ask me to come up here because I'm a futurist. But this is part of why with the content authenticity initiative we are looking for a mechanism for people to make a positive affirmation of what's true and how they can prove it. As opposed to detecting whether something has been modified. Because that feels like that's, you know, a neverending arm's race that's going to be something that, you know, is going to be unsuccessful in order to say that's a faked image.
JENNISON ASUNCION: We have time for one more quick question.
Hi Jennison. It is Sally from Apple. It might be a quick question but it might be a quick answer. I'd like to ask Saqib and Andrew, do you think that gen AI can solve for context. Something that I don't think has been covered a lot in discussions. Things can be very different based on the context within the page they are delivered in or an app or a different environment.
JENNISON ASUNCION: Sure. Maybe I will ask Saqib to take this first.
SAQIB SHAIKH: Yeah, I'm very optimistic for this. You don't want one size fits all descriptions. And here we have as you gave the example of text on a web page but also of an individual where are they located, what have they done in the past and what are their personal preferences. And so I'm very optimistic about this idea. The whole purpose of the description is to serve the human who's consuming it. So yeah. Highly personalized descriptions is very exciting, too.
Can we squeeze in one?
ANDREW KIRKPATRICK: Yeah. I think it can. I'm not going to give you a timeline for that.
Oh, come on.
ANDREW KIRKPATRICK: But depending on what the context nor when that description is generated, if it is generated, you know, at an offering time, it is quite possible that, you know, web content that you experience in the future may be assembled by AI more than it is assembled by an individual author. So there is pieces being brought from different places. So the context has to be established in that way. And I agree, also to Saqib's point, that personalization, user's ability to ask further questions because even if I'm 100% certain that I'm putting the right alternative text on an image I can be 100% certain it is not going to be just right for everyone. That's another great capability that we'll get is to be able to have users ask additional questions.
JENNISON ASUNCION: Please join me in thanking my panelists.
(Applause.)
JENNISON ASUNCION: Thank you.
FRANCESCA CESA BIANCHI: Thank you for Jennison and to the panelists for this great session. And we are going to have a short break. At 2:45 in this room we will have the panel discussion on aging and technology policies.
(Event concluded at 2:30 p.m. eT)
***
This text, document, or file is based on live transcription. Communication Access Realtime Translation (CART), captioning, and/or live transcription are provided in order to facilitate communication accessibility and may not be a totally verbatim record of the proceedings. This text, document, or file is not to be distributed or used in any way that may violate copyright law.
***