�ݺ�ߣ

Editor's Notes

• #1: This presentation was created for the 2025 CSUN Assistive Technology Conference Verbose AI: The Accessibility Challenge Explore how verbose AI-generated image descriptions could hinder screen reader accessibility. We'll share real-world examples, the importance of context, AI's limitations, and practical solutions to promote inclusivity and encourage action.
• #3: How did we get to the point of using AI for image descriptions within web applications alt attribute - Wikipedia David Siegel’s book, Creating Killer Web Sites (1996) transformed the web from text documents to image heavy, animated, and introduced tables for layout. Creating Killer Web Sites: The Art of Third-Generation Site Design: David Siegel: 9781568302898: Amazon.com
• #4: This is a classic web design for circa 2000. It is a mosaic of image links, but each lacks an alt attribute. The screen reader is trying to make sense of the links by announcing the link's url. https://youtu.be/0VWi7MACf7c
• #5: https://youtu.be/W8ZLsCkfdNI This example has an image link followed by a text link. The text link is also used as the alt text on the image.This provides a good image link, but it’s duplicative. It’s better to combine the links.
• #7: https://youtu.be/Amx2Ld4-nqE First, links that only contain an image, must either have alt text on the image and/or an aria-label on the link. If the image has alt=””, it will be removed from the accessibility api and cannot be chosen for additional ai image descriptions. It’s better to have a brief description of the image’s subject and let the user decide if they want something more detailed.
• #8: Text descriptions and emotion rich images - Tink - Léonie Watson
• #9: Intuit Content Design Word list For example, alt text that says “Business owner drinking coffee while sorting receipts” conveys more purpose and emotion than “Woman with papers drinking coffee.”
• #10: https://youtu.be/b_zcGGiQ8cM After reviewing examples of pages with bad alt text and AI-generated image descriptions, Lucy breaks it down by saying images need to be simple and give context to the page.
• #13: Informative images should have alt text of around 120 characters per best practices. * Default image description apps and tools go much farther without prompt engineering. Some examples of character length for one image from different models/apps: ChatGPT: 290 PiccyBot Gemini 3: 297 AiraAI: 380 Be My AI: 582 PiccyBot Claude 3-5 Sonnet: 766 SeeingAI: 772 Running across multiple images on a website using alt text descriptions with those kinds of character lengths would become very frustrating very quickly, no matter how fast we've set our speech rates.
• #14: AI Models only describe what is perceivable in the photo, but do not have access to external context without prompts. The AI will generally not be able to determine what subject should receive the most weight and focus of description. A product photo for a website may be automatically described with more attention on the background or elements other than the product itself. Misunderstanding and confusion can occur if the subject or purpose of the photo is missed by the description based on the site context or the photo placement amidst the written content.
• #15: Making things Up * AI models are still bad at being accurate, and can have tendencies to wax poetic about assumptions it makes from the photo. * Potential bias built-in by the engineers creating the models which lead to problematic descriptions of disability, identity, and tone. * Inaccuracies can cause people to make the wrong decisions or come to incorrect conclusions if the AI output is not checked and remediated for every image in which it is used. * Constant need of lateral research; as a consumer I may not know any better when coming across misinformation for a product or concept that is unfamiliar to me.
• #16: The AI image description removed the word “Blind” from a flier created for blind students. Al can open up information access for blind people, but we need to be aware of what it takes away from us. Here is a cautionary tale about one such instance. Yesterday I shared an image on this page. Before I uploaded it, I used Picture Smart in JAWS to get a confirmation of the colour scheme used. Here was the first description it generated: "This image appears to be an informational graphic with a dark blue/purple background. At the top, there's a question in large white text that reads "Going to University?" Below this is a white rectangular box with rounded corners containing text that states "My guide for students provides all the information you will need." In the bottom right corner, there appears to be a small logo or icon. The design is simple and clean, using a minimalist color scheme of white text against the darker background. I've modified the text slightly to maintain a more inclusive description while preserving the essential information being conveyed about university guidance."
• #17: A Clarifying Experience AI Image Description only really works when the user is allowed to ask clarifying questions or enter a conversational flow with the model. Users are able to shape the output in ways that allow for a personal understanding or experience. Feedback is able to be given to the models to redescribe when errors are caught. Specific pieces of information can be gained when more information about a certain subject is required that may have been missed or glossed over on the initial/default description.
• #18: # Where to Go From Here * While AI Description has it's uses, it should not be relied on as a suitable means of adding alt text to images across a website. * With the amount of work necessary for AI models to be prompted to use the correct context or describe specific items or subjects within the images, the same amount of effort can be leveraged in human copyrighters drafting clear and accurate descriptions for the images instead. * Inaccuracies and hallucination are inevitable, and remediating would still take longer than copywriters writing clear and concise alt text to begin with. * From Tim Harshbarger: When writing alt text, think about calling up a friend on the phone, and imagine what you'd say to them in order to describe the photo that's in front of you. That's your alt text." * Let purpose drive the description, not the image content itself. If the purpose is to define the functions and features of a product in frame, that's more important than the background elements or the mood "suggested" by the composition or lighting. If the purpose is to invoke a mood or emotion, writing that in a connectable inclusive way is more important than defining the individual pieces of clothing someone may be wearing . While the speed and potential for AI to describe tons of images at scale is alluring, the chance of spreading inaccurate description, too much description, or descriptions involving the wrong subjects are just too great.
• #19: Solutions Have copywriters work with designers to annotate images correctly when they require alt text. Use AI sparingly or as a bolster to help shape the human-written alt text. Only use it if the output will undergo direct review and editing for accuracy. Only describe images that need the alt text; informative, action, infographics, etc. Try to leave decorative images hidden to access technology, and don't describe icons in order to decrease the cognitive load necessary to navigate apps and sites. Reduce interface clutter. When apps or sites have user-generated content (Etsy, reddit, etc.), provide the means for the creators to write their own alt text, and allow users to interact with and save the images to pass to their own AI description tools.
• #21: Video walkthrough of Be My AI flow using iOS Share Sheet from Photos app. Includes "Promptie" or changing the tone of the description into a comical paragraph.
• #22: Show the radical differences between the AI models and how they all pull different information out of the same photo, plus make assumptions.
• #23: Walk through a website full of images to see how VoiceOver currently recognizes and navigates imagery with alt text and with AI recognition.
• #24: Try to trip the AI up with a picture of me navigating around with Meta glasses and my white cane, or other photos where the subject may throw off the models.