- [Kari] Hello. Good morning, good afternoon, everyone. Just give us a few minutes and we will get things started. Just wanna let everyone get in the room first. Again, just wanna say good morning, good afternoon, everyone. Thank you for joining us today. Just have a few housekeeping items before we get started. This meeting is being recorded and the recording will be shared after the event. The Q&A box is available if you have any questions. However, we will not be answering questions during the webinar. We will respond to any questions that you submit after the event is over. And lastly, if you need closed captioning, closed captioning is available for you to turn on, and if you need any additional accessibility help, accessibility training, et cetera, you can reach out to Ida. That's ida@tpgi.com. And I will add that email address into the chat later. And with that, I will get things started. Again, just wanna thank you everyone for joining us today for our webinar, "What AI Can Do For and To Web Accessibility," with Ricky Osman and Hans Hillen.

- Welcome, everyone, and thanks for attending our webinar, "What AI Can Do For and To Web Accessibility." This is the talk that Ricky and I gave earlier this year at the CSUN '24 conference. We're presenting this as a prerecorded session this time as time zone constraints prevented us from doing the session live. If you want to follow along with the slides today, there's a QR code on the screen to scan, or you can go directly to bit.ly/tpgi24ai. This is all lowercase and the numbers are actual number characters. bit.ly/tpgi24ai. The slides will also be available through the details of this webinar. In this session, we're going to talk about the potential effects of AI on web accessibility and people with disabilities in general. First, we'll look at how AI has been transforming existing accessibility technology using overlays, captions, and alt text as examples. Next, we'll talk about how AI can directly affect the personal lives of people with disabilities, and we'll focus on computer vision, personalized output, and recruiting. Then we'll finish by looking at the possibilities of using AI to create accessible content. But first, let's quickly introduce ourselves. We're both with the TPGi Knowledge Center team, and as such, we're responsible for maintaining and improving accessibility-related knowledge and content within TPGi. My name is Hans Hillen, and I'm the Team Director. My co-presenter today is Ricky Onsman, who is a Senior Technical Content Writer for the Knowledge Center. The other KC member is James Edwards, who is our Technical Consultant and is responsible for our design patterns in WCAG 2.2 specific content amongst other things. Speaking of recruiting, we will soon start looking for the fourth member of the TPGi Knowledge Center. We'll be looking for somebody who is fluent in everything accessibility, a good technical writer, and comfortable with content coding. If that sounds interesting to you, keep an eye on our careers page at www.tpgi.com/business/careers. Now that that's sort out of the way, Ricky's going to talk about AI and existing accessibility support.

- AI and existing accessibility support. Now, we're not going to go into detailed explanations of what is or isn't AI, including whether it actually exists. For our purposes, Artificial Intelligence, AI, refers to the development of computer systems that can perform tasks that typically require human intelligence. AI can take many different forms, some of which we'll be exploring today. Some of the technologies that will come up in our discussions today include: Machine Learning, which uses structured data and algorithms to imitate the way humans learn, Natural Language Processing, which allows computers to engage with humans in conversational language, Large Language Models work off very large datasets to produce content humans can comprehend, Computer Vision enables computers to analyze, recognize, and describe images in human terms, and Deep Learning can learn from unstructured data, increasing scale, depth and complexity. As an example, ChatGPT uses Machine Learning, Large Language Models, and Natural Language Processing to respond to prompts from humans in a conversational way. These technologies are used often in combination in various tools and functionalities. Where needed, we'll explain these terms as they come up in the examples we'll be looking at today, which include automatic captioning, automatic alt text, ChatBots, predictive text, UI customization, text-to-speech, speech-to-text, and overlays. Let's start with the digital elephant in the virtual room.: overlays. This talk was always going to have elements of the good, the bad, and the ugly, so let's start with, by reputation, at least, the ugly. Overlays and AI. At its most literal, an overlay is simply something laid over something else. In digital technology and web content, an overlay is a piece of software laid over a website or app to perform a specific function. We use them all the time. Pop-up dialogues, date pickers, chatbots, and tool tips can all be overlays. In themselves, as long as they are accessible, keyboard operable within the viewport, persistent, dismissible, et cetera, there's nothing inherently wrong with using overlays. In a digital accessibility context, however, some overlays make very specific claims, including being able to use automated tools to assess web content for WCAG conformance, and then automatically implement code changes or new code that will guarantee conformance, with the site owner needing to do no more than install a widget. There are variations on this. Some overlays present an icon that the end user can activate to set a series of user preferences. Some generate a report that tell the site owner what they found and what they fixed. But they all fall under the broad title of Digital Accessibility Overlays, and they are what we'll refer to as overlays for the purpose of this talk. Now, there's been a great deal of criticism directed towards overlays, which can be briefly summarized as: One, the modifications that overlays implement cannot guarantee full conformance with WCAG, because no automated tool can. Two, the modifications that overlays implement are often unreliable, ineffective, and/or create other accessibilities than the ones they address. Three, the modifications that overlays make often interfere with assistive technologies and accessibility settings already in use. And four, the use of an overlay may create non-compliance with privacy regulations. These are serious considerations, and more detail is available at the Overlay Fact Sheet website. Now this talk is about AI, so we are going to focus on overlays that use, or claim to use, Artificial Intelligence. Let's take a look at one that does go into some detail: accessiBe. The accessiBe homepage tells us that thanks to AI, web accessibility at scale is finally a reality. Its accessWidget claims to be the leading automated web accessibility solution powered by AI. accessiBe is a game-changer in web accessibility, simplifying, streamlining the process of becoming accessible and compliant using AI, machine learning, and computer vision. The accessWidget product actually consists of two parts. One, a process that operates in the background, and the other, a widget on the user interface that a user can configure. Under the heading, "Computer Vision Applied To Accessibility," accessiBe tells us, "accessWidget scans all images on the site. Whenever alternative text, Alt attributes, is missing, it will extract the embedded text using OCR and learn the objects that comprise the image using IRIS technology." Well, optical character recognition, OCR, is a technology that's been around for many years and is used in many contexts to turn images of texts, such as scan documents or on webpages, into actual texts that can then be manipulated. It's probably one of the oldest technologies that can be described as AI. I have to dig deep beyond the accessiBe website to find out what IRIS technology is. It refers to an advanced form of machine vision and deep learning and neural networks, which combine to identify image content, not to be confused with IRIS recognition, which is about scanning eyeballs to authenticate identity. Now, basically, IRIS is trained off a database of images that must be sufficiently representative, in other words, huge, to interpret and recognize images. How exactly it does this is not clear and it's certainly not explained by accessiBe. Quote, "Then, accessWidget will automatically provide accurate and elaborate alternative text to these images. When a blind user enters a site, their screen reader will rely on these descriptions to communicate what is on the page," unquote. And the claim is that these adjustments are compatible with popular screen readers such as JAWS, NVDA, VoiceOver, and TalkBack, which is a claim that some dispute. Which specific AI technologies are used to ensure keyboard accessibility is even less clear. Quote, "accessWidget makes websites navigable by keyboard. By using ARIA attributes and contextual understanding AI engine, accessWidget makes the necessary adjustments throughout the website's code. People with motor impairments can use the keys to do everything from closing popups and forms to opening drop downs and menus." So accessWidget uses an automated checking tool to find accessibility issues on functional elements and then uses JavaScript to apply ARIA roles, attributes, and labels to make those elements keyboard accessible. What role AI plays in this is not specified. And with all this going on in the background, this part of accessWidget might better be described as an underlay rather than an overlay. But there is also an overlay widget in accessWidget, which may be installed on the user interface. This has a number of generic user profiles that can be turned on and customized by the user or turned off completely. In fact, the website owner doesn't have to install widget at all and can just rely on the background process. The accessiBe logic seems to be to use unexplained AI technologies in background and let the user configure a relatively straightforward front-end widget as they wish. Well, ultimately then, we're left with one overarching question regarding the use of AI in digital accessibility overlays. Is the claim of some overlays that the use of AI provides fully automated functionality that can identify and remediate all WCAG issues on a website with no human intervention valid? On the available evidence, the answer has to be not completely, or not yet, or just no. Automatic captioning. Captions have been around ever since we've had moving images. From captioned cards used in silent movies to convey key dialogue and sound effects to the subtitling in one language of movie dialogue delivered in another. Julia Child's cooking show, the French chef, was the first TV show to trial captions in 1972. The show was recorded live in the studio and the captions were then burned onto the print afterwards, which meant they couldn't be turned off or adjusted what we now call open captioning. By the 1980s, technology had advanced to making available closed captions that could be turned on and off, albeit via a relatively expensive decoder, and legislation was introduced mandating the provisions of captions for prerecorded television. Meanwhile, speech recognition and text-to-speech models were being developed for single speaker voice recognition with very limited vocabularies, with machines trained to recognize speech from one specific person if they only use certain words and render them as text. By the 1980s, though, machine vocabularies were larger than those of the average person and had become speaker independent. Computerization was making speech recognition much more reliable and useful. A big breakthrough came in the 2000s when probability logic, how well a machine could guess the right word, was overtaken by long short-term memory, LSTM, a form of deep learning that can remember hearing a word from thousands of steps ago. Greater computing power, larger datasets, and LSTM brought speech-to-text into the AI family where it became known as automatic speech recognition, ASR, the basis for automated captioning functionality currently offered by video providers like YouTube, Vimeo, and Facebook by activating a control on the user interface. AI for the win! Well, nearly. When Google made this available on YouTube in 2009, even in this moment of automated glory, it had to admit that the quality of its auto-captioning was not always very good. The New York Times quoted deaf Google engineer, Ken Harrenstien, is saying at the launch. "Sometimes the auto-captions are good. Sometimes they're not great, but they're better than nothing if you are hearing-impaired or don't know the language." And there's the rub with ASR. It isn't very good. And it isn't getting much better. And let's be clear about what we mean by very good. In the 2023 state of automatic speech recognition report by 3Play Media, a vendor of video accessibility services, no automated captioning from 10 different vendors achieved a word error rate of less than 6.86%. Once formatting like punctuation and capitalization was taken into account, none of the vendors had a formatted error rate of less than 17.2%, that equates to about one in six words being incorrectly captioned. At best. Some auto-captioning vendors had error rates as high as 42%. Now with regard to web content, it should be noted that while standards like WCAG require the provision of captions in some context, they have little to say about quality, how accurately captions convey dialogue. In television context, though, the FCC is much more specific, requiring a 99% accuracy. Now, if that standard was included in WCAG, no webpages with auto-captioned videos would conform. In some ways, AI has made some captioning issues worse. Where earlier forms of automated caption might admit something is, quote, "indecipherable," unquote, AI-powered ASR tends to supply something even if it's wildly inaccurate. The caption here says, "And we have lift off the guitarist G 11 mission." Now, what was actually said was, "And we have lift off of the Antares NG-11 mission." It's entirely possible that some of the current barriers to greater accuracy, such as that ASR engines respond better to a single voice rather than multiple voices, will be overcome, but probably not in the very near future. In the meantime, to get decent captioning, a user has to upload the video, download the automatically generated captions file, edited for accuracy, and re-upload it. Although YouTube now does have an online editing facility. Or you can pay a commercial service to do it. And it's interesting that a major provider like Rev.com has developed an automated captioning service alongside its manual service but concedes that the manual service is much more accurate. Quote, "Rev offers the most accurate automatic transcription model in the world, and its word error rate is about 14%. That means you can expect about 86% of your transcription to be accurate, on average. If you need higher accuracy rates, you might want to go with human transcription. For the foreseeable future, like many forms of AI-powered functionality, a human touch is needed to ensure captioning accuracy. AI and alt text. Using machines to analyze and report on the content of images is not at all new. OCR was an outcome of pre-World War I experiments with telegraphy and creating reading devices for the blind. One of those projects developed the capacity to read printed characters and convert them into telegraphic code. Another used a handheld scanner that produced a range of audio tones that corresponded to different characters printed on a page. It was Ray Kurzweil in the 1970s who created the breakthrough technology that enabled a machine to read out text content to blind people, combining a flatbed scanner with synthesized text-to-speech. Nowadays, as well as being used in dozens of applications that require translating a picture of text into text, OCR is a core part of screen reader technology. It's not generally thought of as AI, but it does fall into that part of the AI family known as text-to-speech, and that's why alt text is important. Screen readers read out text, whether that's visible text on a webpage or hidden text describing an image. They can use OCR to read out images of text with varying levels of accuracy, but without that visibly hidden alternative text, screen readers don't describe other images yet. Back in the 1970s, some of that same OCR technology was being used to develop computer vision, CV, a technological way of analyzing and reporting on the content of a much wider range of images than just text. It was very much seen as part of the nascent development of AI, automating tasks associated with the human visual system. Many of the algorithms developed then are still in use today, such as the ability to discern edges, implied structure, the labeling of lines, and motion estimation in videos. As more sophisticated algorithms developed, the massive recent increases in datasets and the processing power to analyze them has led to machine learning and deep learning techniques being applied to computer vision to produce truly remarkable advances in image recognition and description. So, CV plus ML plus DL should equal great alt text, right? Well, we all realize it's not that simple. Producing text that describes an image in an automated mechanical way is not the whole story of what makes good alternative text for accessibility purposes. Early efforts revealed limitations in datasets that made machines produce texts that was neither accurate nor appropriate. In 2015, Google Photos was rightly criticized for offensive auto-tagging of photos of Africans Americans. But larger datasets and better analysis are improving that greatly. The frontier we now face is context. Isolating an image from the context of its web content does not produce good alt text. Now with this image, Facebook's alt text generator ignore the text of a message "green frog doorknob" and concluded that this, quote, "May be an image of amphibian," unquote, which is not untrue, but it doesn't tell the screen reader user that it's a doorknob in the shape of a frog. However, there are currently underway projects to teach machines to analyze surrounding text and other images on a webpage, or a social media post, and generate alternative text for a screen reader that equates to that which would be added by a human. With the kinds of technologies we are talking about, this is by no means an insurmountable tasks. This slide shows the analysis of a social media post that says, "I check my mail daily. Today I received 18 election flyers. Good Lord. Make it stop. PS - be sure to vote. #CaliforniaPrimary." With a photo of a hand holding multiple election flyers. Logic flows for deciding on the image content are then described, resulting in automatically generated alt text. "Exasperated OP holds 18 election flyers for various candidates in the 2022 California Primary Election." The AI has correctly assessed image-relevant content to include in the alt text from the tweet text, including what election flyers means, what the hashtag refers to, and even the tone of the tweet. In fact, it's not unreasonable to expect that some AI alt text generators are already better at this than human beings. Anyone who's audited more than a few website knows that human generated alt text is often variable at best. But the alt text of this photo of a race horse entering the barrier before a race, a human wrote: "a hapless racehorse being cruelly jammed into a too small barrier stall against its will before being forced to race in circles while being whipped for human entertainment." And mind you, this alt text was written for a sports news portal, not for a site protesting against cruelty to animals. In many ways, a machine working off a massive dataset may be better placed than the human to analyze an images context and choose a description that fits the purpose of the web content. This extends to complex images, diagrams, and illustrations. Charts and graphs visualize not just data but relationships between data. That's their purpose. To show at a glance what might otherwise not be very apparent from a table of numbers. AI alt text generators can already describe these relationships, working off the original tabular data used to generate the images to create alt texts such as: the percentage of female stem workers rose 20% between 1970 and 2019, with the greatest increases in social science and math occupations. Now, I do have a caveat. There is a current generation of AI-powered alt generators that focus on search engine optimization rather than accessibility. Keyword stuffing is of course just a reflection of how many humans currently regard alt text. AI just automates it. Nevertheless, there's no reason purpose-specific AI tools can't be used now to analyze an image and its context to calculate appropriate alt text. However, what if AI could be used to generate not just alt text that it thinks is right but alt text that it calculates to meet the needs of each user, every user in every context? Well, that's where AI is going and where this talk is going next.

- In the next section, let's look at a few examples of how AI can directly affect the daily lives of people with disabilities today. There are plenty of examples out there. For example, generative voice AI provides a more human and natural sounding form of synthetic speech. This is useful for text-to-speech solutions like professional sounding book narrations. It can also be used for speech-to-speech solutions, providing a personal voice to people who have lost or are losing their own voice. There's AI-powered hearing aids, which can filter out irrelevant background noise in a more intelligent way, or AI-powered artificial limbs that allow for more natural and intuitive control over the prosthetic. But the two examples that I would like to spend a bit more time on in this talk are computer vision and personalized AI responses. So let's start with computer vision. To simplify Ricky's definition from earlier, computer vision basically attempts to make sense of an image in the way a human would and describe that image to a person who can't see it. So why is computer vision as a disability aid such a game-changer? It's because in the end, the perfect alt text does not exist. You will never know what every single individual consuming your content needs or wants to know about your image. Rather than us trying to think of every detail to describe an image with, or using auto-tagging to do that for us, why not let the user decide how detailed they want their image description to be, what specific aspects of the image they're interested in, and what follow-up questions they might have about it? With computer vision, this is possible, and that's why it's a game-changer. It's like having the longdesc attribute back and having somebody else write its content for you. Let's have a closer look at an app feature that specifically uses computer vision to aid blind and low vision users: Be My AI. One thing that I really like about it, and what some of its users might not be aware of, is that it doesn't limit you to just describe pictures taken on the spot with your device's camera. You can send images to be described with Be My AI from any app that lets you open a share menu for images. This includes webpage images in Safari as well. Another great feature is that after you ask Be My AI questions about your image, you can then create a link to your chat log to share what you learned with others who might need the information. Here's a basic example of using Be My AI to learn more about an image. If I had to provide an alt text for this photo, I might describe it as "seagull perched on top of a Pouwhenua.". For those of you who don't live in New Zealand, a Pouwhenua is a type of marker post that is used in Maori culture used to mark important boundaries or places of significance, and Be My AI describes it really well. "The picture shows a clear blue sky with a few trees in the background. In the foreground, there is an ornate, weathered stone carving resembling a Maori design with a circular hole in the middle and a spiral pattern around it. Perched on top of the carving is a white bird with gray wings and a red beak, possibly a seagull. Behind it, on a similar stone post, is another similar bird," end quote. You'll notice it provides a decent amount of detail without going overboard. It was then able to answer my follow up questions, which I've linked to on this slide, perfectly. It told me that this was likely a Pouwhenua and that the picture was probably taken in New Zealand. Perfect. For the next example, let's play a little game. Let's see if you are smarter than an AI. For anybody who can see the screen, imagine if a blind person showed you this photo of a power bank and asks you how much battery percentage they had left. What would you answer? If you said 56%, you are wrong as every other AI chat system that I tried this with, because the power bank in this photo is actually being held upside down and the real answer is 95%. This goes to show that AI can and will get things wrong, which make them a less than ideal choice for advice on potentially dangerous subjects like medicine labels or safety instructions. On desktop, you can actually make use of computer vision directly from within your screen reader. For JAWS, the March 2024 release added a new feature called PictureSmart AI, which allows you to request a detailed description for any image on your PC, including images on webpages. The descriptions are provided by both ChatGPT4 and Google Gemini, and from what I understand, there is going to be an update that will also allow you to ask follow up questions there. There is similar functionality available for NVDA users, although it's not built into the screen reader like with JAWS. You'll have to install an NVDA add-on, such as AI Content Describer, and provide your own OpenAI API key that requires payment for ChatGPT4 output. The other AI benefits that I wanted to mention is the ability to adapt and personalize responses. We are all different and we all learn in different ways, but the content that we learn from generally comes to us in a fixed format. We have to adapt to it when it really should be the other way around. Content should adapt to our needs and preferences. With AI, it's possible to transform and deliver information in a simplified format that works best for the person requesting it. Let's try this with something that could really benefit from simplification: WCAG. I asked ChatGPT to explain WCAG success criterion 2.5.1, which, as you all know, is the focus obscured one. Again, I have linked to the chat transcript on the slide. It described the success criterion very well with that wicked style and jargon that we all know and love. Here's a quote. "The main requirement is that all functionality that uses multipoint or path-based gestures for operation can also be operated with a single pointer without a path-based gesture," and so on. This is fine if you know your WCAG stuff, but not everybody is going to get this immediately. So my follow up question was: can you explain it using less complex terminology to somebody who is confused by the term 'pointer gestures'? And the response was, "Imagine you're on a website or using an app. Sometimes, you need to do things like spread with two fingers apart on the screen to zoom in, or swipe with your finger to move through pictures. These are called 'complex gestures' because they need specific finger movements or more than one finger. But what if someone finds it hard to make these gestures? Maybe they can't move their fingers that way, or they only have one hand available. This is where the rule comes in. It says that websites and apps should let people do the same things without needing those complex gestures," end quote. I really like this simplified explanation, and it shows that getting the same answer in different formats can really make a big difference. Another way that AI currently affects the daily lives of people with disabilities is through their job recruiting process. If you want to learn more about this topic, I can strongly recommend the book "The Algorithm" by Hilke Schellmann. She describes how it's more and more common these days for AI to handle the initial stages of the online application process and how this can put people with disabilities at risk of being significantly disadvantaged. Let's look at some of those risks. First of all, AI algorithms are good at recognizing general patterns rather than individual variances, and they're often trained on large datasets representing average or typical behaviors. They may even be trained specifically on previously high performing employees at category in which people with disabilities will likely be underrepresented. Secondly, disability-related factors cannot lead to negative scores. For example, when applicants are required to do performance-based tasks or play games involving reaction time. In addition, an AI algorithm might track an applicant's speech or facial patterns during an interview and unintentionally penalize patterns specific to disabilities. Third, AI technology can easily be viewed as a black box by the company that uses it. If AI recruiting tools are not transparent about how they assess candidates, it's difficult for employers to understand why certain applicants are rejected. For the last segment, let's talk about what AI can do for the creation of accessible content. Can AI replace an experienced accessibility consultant? I would say that it can definitely perform some of our tasks very well if you know what to ask for and what to expect. I thought it would be interesting to do a little Turing test with ChatGPT and feted the questions of the technical interview I used to hold with TPGi applicants. Some of these questions are quite tricky, involving code snippets meant to trip up people who lack practical experience. I wanted to see if I would hire, so to speak, ChatGPT if it was a person purely based on its answers. To my surprise, it pretty much answered all of them perfectly, no matter the format or trickiness of the question. So, you can ask it to explain accessibility concepts, to find accessibility issues in your code, recommend fixes, and write accessible code. But you can also have it create things like accessibility test plans, bookmarklets, JAWS scripts, and so on. Wait a minute, that sounds a lot like all of our jobs. Another popular AI tool that can help with accessible web development is GitHub Copilot. We don't have enough time in this session to cover all of its features, but I recommend reading the GitHub article called, "Prompting GitHub Copilot Chat to become your personal AI assistant for accessibility." I've linked to it here in the slide. Suffice it to say that GitHub Copilot can do a lot of what ChatGPT can do. It will answer accessibility questions and recommend, explain, or fix accessible code. The Copilot really stands out, in my opinion, is how well it integrates with your projects and development environment. Where ChatGPT requires you to submit isolated requests, Copilot learn from your project structure and coding style while you're creating and editing that code. I've been using it a lot during the last few months and I've found it to be a great time saver and teaching tool for Turing developments. But of course, it's not all perfect. Far from it. We know that AI can and will get things wrong and that it will confabulate. Accessibility is not exempt in this regard, and why would it be? Most of the code out there is inaccessible to varying degrees, so a model trained on it will inherit those flaws as well. I think services like ChatGPT and Copilot are great for smoothing over small knowledge gaps, provided you have enough understanding about a subject and its context. Once those knowledge gaps become too deep and too wide to be smoothed over, once you basically start outsourcing accessibility to the AI without learning anything about your accessibility concepts yourself, you're likely going to get stuck fairly quick. This is because you won't be able to gauge or verify whether the responses you're getting make common sense within the context of what you're trying to do. This is a problem with AI in general. The responses always appear well-written, researched, and authoritative, which makes it a lot harder to recognize when it's spouting absolute nonsense unless you have the expertise to do so. And this is something that you'll hear in every other AI talk. The need for the human in the loop. The Copilot should never become the pilot and human expertise will always be needed to keep the AI in check. Without that, you'll waste a lot of time troubleshooting the AI's concoctions before you realize that your Copilot actually already crashed a plane a while back. This is especially true for accessibility where an incorrect approach can go unnoticed much easier and for much longer than with generic coding mistakes. And that's why I'm not worried about my job just yet. As you've probably noticed, pretty much everyone is experimenting with AI these days, and so is TPGi. We are exploring looking into possibilities of which there are many. We have a wealth of data that could potentially be used for training, such as our knowledge base, tutor training courses, engagement data, and help desk tickets. We've built our own AI model, which we call the KnowledgeBase Companion or ChatGPTi. We're currently testing and training it internally. It's not available publicly just yet, and it won't be until it learns how to behave. Then there are other possibilities like computer vision to detect inaccessible components for our automated rules, or AI-driven tools to speed up the tricky and time-consuming parts of manual audits. Like I said, we're still exploring at this point, but if there are any developments to announce, you'll find them on our technical blog, which you can access through tpgi.com/blog. So what key points can we take home from all of this? One, it's all good, it's all bad, it's all ugly. All the AI advancements we talked about today can be great or horrible depending on how they're used. Two, AI has the potential of hugely improving independence in the daily lives of people with disabilities. Three, it can work against people with disabilities through exclusion or biases inherited from training data. Four, it can be a great teaching tool for creating accessible content in a more productive way, but only if you know what you're doing. And finally, there will always be a need to have a human in the loop. That brings us to the end of our talk. If you want to get in touch with us, send a message to info@tpgi.com. Feel free to reach out to us on Mastodon or use the contact details provided with this webinar and we'll get back to you. Check out the TPGi technical blog and sign up for our Knowledge Center newsletter by sending an empty email to knowledge@tpgi.com. Thanks for attending and sorry we couldn't do this live. Take care.

- [Kari] Again, just wanna thank everyone for attending today. If you had any questions, et cetera, we will reach back out to you by email after the event is over. And yes, a link to the recording will be sent out after the event, as well as the transcription and a copy of the slide deck. I'm also going to put the email address for Ida in the chat as well, so if you have any questions you would like to have Hans or Ricky answer, we can get those questions sent over to them and get some answers back to you as soon as possible. Again, thanks for joining us today. I hope everyone has a great day, and we hope to see you again for our next webinar.