<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Hanna Kastrytsa on Medium]]></title>
        <description><![CDATA[Stories by Hanna Kastrytsa on Medium]]></description>
        <link>https://medium.com/@annakos070500?source=rss-eaa9f1fa784d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*S9Db4UWY4sJR4gdFyBn5Og.png</url>
            <title>Stories by Hanna Kastrytsa on Medium</title>
            <link>https://medium.com/@annakos070500?source=rss-eaa9f1fa784d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:51:24 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@annakos070500/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Biases in AI-generated imagery]]></title>
            <link>https://medium.com/@annakos070500/biases-in-ai-generated-imagery-ee6ec9b99c88?source=rss-eaa9f1fa784d------2</link>
            <guid isPermaLink="false">https://medium.com/p/ee6ec9b99c88</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ux-research]]></category>
            <category><![CDATA[installation]]></category>
            <category><![CDATA[interaction-design]]></category>
            <dc:creator><![CDATA[Hanna Kastrytsa]]></dc:creator>
            <pubDate>Fri, 21 Jun 2024 04:36:40 GMT</pubDate>
            <atom:updated>2024-06-21T04:36:40.584Z</atom:updated>
            <content:encoded><![CDATA[<h4>How we designed the installation Echoes of Exclusion</h4><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F945547183%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F945547183&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F1851032032-b198609f67bd644620d2b3189c93e2bffdfe32903dbcaa27696ded42590c1d74-d_1280&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1440" frameborder="0" scrolling="no"><a href="https://medium.com/media/1774c987e9d81821f9d5a53700ce6a07/href">https://medium.com/media/1774c987e9d81821f9d5a53700ce6a07/href</a></iframe><p>A few months ago, we encountered an exciting challenge within the Prototyping Interactive Installations course at SUPSI. It tasked us with delving into the intricacies of metadata in media and crafting an interactive installation — a venture none of us had previously undertaken. We approached it with enthusiasm, eager to explore a new field and push the boundaries of our capabilities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X1rFZJZCwtheVZSKudvHrg.png" /></figure><p><a href="https://www.exposed.photography/en/programme/maind-master-educational-program">EXPOSED</a></p><p><strong>Eventually, we created “Echoes of Exclusion” — an installation that would later captivate audiences at the prestigious EXPOSED Torino Foto Festival.</strong></p><p>In this article, we invite you to explore the journey of our installation — from the initial spark of curiosity to its realization in the gallery of Turin.</p><h3><strong>Concept of Echoes of Exclusion</strong></h3><p>“Echoes of Exclusion” is an art installation that confronts viewers with the hidden biases embedded within AI-generated imagery. As visitors step into the space, they encounter a series of interactive projections featuring AI-generated visuals. These images showcase everyday scenes and objects but with subtle distortions and biases that reflect societal stereotypes and exclusions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K8jhgUTM1hPc0KunEpBvFg.png" /></figure><p>Through a combination of audio narration and image comparison, the installation invites visitors to critically examine the imagery and consider the underlying biases at play. By highlighting these biases in a visually compelling way, “Echoes of Exclusion” sparks important conversations about the impact of AI on our perceptions of race, gender, culture, and more.</p><p>The installation aims to raise awareness about the limitations of AI-generated content and the potential harm it can perpetuate. “Echoes of Exclusion” empowers viewers to challenge these biases and advocate for more inclusive AI technologies.</p><h3>The societal challenge presented by AI</h3><p>In the ever-evolving landscape of artificial intelligence (AI), the rising adoption of AI technologies is reshaping industries and enhancing efficiency. However, amidst this transformative wave, it’s imperative to confront the biases of generative AI. Biases, misrepresentations, stereotypes, and discrimination propagated by AI pose significant harm and challenges.</p><h4>Why does this matter, and who does it impact?</h4><p>The representation of diverse groups in AI-generated content profoundly influences societal perceptions and individual identities. When certain data is consistently absent or misrepresented, it conveys the problem of harmful stereotypes and reinforces existing inequalities.</p><blockquote>“Every AI generated image is an infographic about the dataset. AI images are data patterns inscribed into pictures, and they tell us stories about that dataset and the human decisions behind it”.</blockquote><blockquote>Erik Salvaggio, How to read an AI images, 2023</blockquote><p>Understanding and recognizing the limitations of AI-generated imagery and starting conversations about it enable us to better understand our current cultural context and enact meaningful change for the future. Despite the competitive pressures in the tech industry, it’s essential to prioritize ethical responsibility over mere innovation.</p><blockquote>“People learn from seeing or not seeing themselves that maybe they don’t belong”.</blockquote><blockquote>Heather Hiles, Chair of Black Girls Code, 2023</blockquote><p>AI also holds promise in promoting diversity and inclusion when biases are actively addressed by developers, ensuring that AI-generated content faithfully represents the diverse range of human experiences.</p><p>Our project, titled “Echoes of Exclusion,” seeks to shed light on the pervasive issue of biases in generative AI. Through the use of imagery, we aim to initiate discussions that highlight the distortions and biases present in AI-generated content.</p><h3>Research phase</h3><p>The brief for this project was to create an interactive installation related to contemporary photography. The interaction had to occur through the use of a phone, and we were encouraged to consider how we could include metadata in the installation. One of the main points was to find clear answers to a few questions:</p><ul><li>What is metadata?</li><li>How do people usually use it?</li><li>How does it work?</li></ul><p><strong>Metadata is defined as data that provides information about one or more aspects of other data. It is used to summarize basic information about data, making tracking and working with specific data easier.</strong></p><p>Usually photo metadata is the information and specific details concerning a particular image file. This information often includes the date created, author, file name, content, themes and more.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ws4DMoWm4q1mNe6EvDAzGA.png" /><figcaption>Topics and questions which were discovered during the 1st round of ideation</figcaption></figure><p>We went through a large number of different ideas for the concept and particularly liked ideas related to social media and the unveiling of data that is part of imagery. These ideas were represented with corresponding moodboards, references, sketches, interactions, and requirements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lbWMKzTIq9KFjosugV1wBg.png" /><figcaption>First idea sketches with metadata representations on installation</figcaption></figure><p>After some discussion, we found that while our initial ideas showed promise, they needed further development to fully align with our overarching concept. Certain aspects of these ideas were attractive and useful in terms of interaction and internal messaging. Following a period of research, during which we explored various avenues, we decided to combine elements of our previous concepts in search of a fresh approach.</p><p>We found that when content is uploaded, AI algorithms often generate descriptions for photos or videos, contributing to the vast pool of user-generated content. This realization led us to consider the intersection of Artificial Intelligence (AI) and social media. This fusion became the focal point of our exploration.</p><h4>Experiment of “ladies on brunch”</h4><p>We decided to conduct a small experiment. We selected two images from our Instagram accounts that had been published with the same hashtag #ladieswhobrunch and had similar text descriptions. The experiment involved the following steps:</p><ul><li>Uploading both images to an AI-based image generator.</li><li>Asking the AI to generate descriptions for these images.</li><li>Asking the AI to imagine new images based on these descriptions.</li></ul><p>We used MidJourney to generate images based on the uploaded originals.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4mtvnpdny1jDavwv9zWuDg.png" /><figcaption>From left to right: original images, AI-generated description, AI-generated images based on description</figcaption></figure><p>During the experiment, we discovered that photos depicting the same action or event may be interpreted in completely different ways by generative AI. The photo was chosen from Instagram by searching for the hashtag #ladieswhobrunch to make it similar to Example 1.</p><p><strong>The issue at hand is the biases inherent in generative AI.</strong></p><p>Through further research, we identified several criteria for selecting the types of photos to be used in the installation. These criteria were developed based on biases uncovered in previous research on AI-generated imagery.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MCVFEbqiLqTAGXMyyG-TbA.png" /><figcaption>Types of photos which are more susceptible to biases</figcaption></figure><p>We undertook extensive experiments to better understand both the common and less common biases seen in AI-generated images.</p><figure><img alt="Images and AI-generated images with biases" src="https://cdn-images-1.medium.com/max/960/1*cCq1ryMMXcdQx1_HKWZlEg.gif" /><figcaption>Original images and AI-generated images with synthetic biases</figcaption></figure><h3>Prototyping phase</h3><p>Finally, we selected an idea and presented the details to the stakeholders for early feedback and approval. Our first prototype included only a preset of photos with a slider where users could view the results of our experiment. We created an initial sketch of this idea to clearly explain the user journey, and it can be seen in the photo below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_WD1JQ9ZozyOwFoZ_MeDnA.png" /><figcaption>First sketch about interaction with installation</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oHR1xHQ9_do5VGUpRmPk1A.png" /></figure><p>A 3D simulation of the installation in the gallery was also created to showcase the interactions in detail. This prototype simulation was developed using Unreal Engine to demonstrate to stakeholders how the installation might work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*iq97tFCtGtDtwC6WfjySwA.gif" /><figcaption>Unreal engine simulation of our installation</figcaption></figure><p>However, the final user flow was changed. We decided to give visitors the possibility to upload their own photos and generate new results. The final version of the user flow is shown in the scheme below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j2HCpvBAHlHo5TAy76MhTw.png" /><figcaption>The final user flow of the installation</figcaption></figure><p>The user experience is designed as follows:</p><ul><li>Users enter the installation space where they can scan a QR code to start the interaction.</li><li>They can upload their own photo to generate the AI-analyzed image and description.</li><li>After a few minutes, the AI-generated photo will appear on the main carousel of the installation.</li></ul><h3>Development and testing phase</h3><p>We used a number of different programs and processes to create the installation. Our first prototype was based on a preinstalled set of photos and their generated versions. We developed two main applications:</p><ol><li><strong>Main Installation</strong>: This program (or website) represents the main wall with two images displayed on it along with their descriptions.</li><li><strong>Visitor Webpage</strong>: This website allows users to choose the original photo and request to view its AI representation on the wall.</li></ol><p>The code was primarily written using JavaScript and Supabase for the backend, and HTML and CSS for designing the layout and interface. We also utilized Google Sheets and Contentful as main data storages. Supabase was primarily used to facilitate real-time data exchange between the two applications.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ULx5J_xQ1Hn8AosKkvuwwA.png" /><figcaption>The architecture of first prototype of installation</figcaption></figure><p>The structure of the final installation was improved and developed. The entire process ensures the installation’s carousel works smoothly without needing an internet connection, minimizing crash risks. The system regularly checks for new image uploads and only adds them to the local carousel after all processes are complete and files are downloaded locally.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gbDa_APeE45xBYLn04xK1g.png" /><figcaption>The final development scheme</figcaption></figure><ol><li><strong>User to Server</strong>: users can upload a picture through a webpage. JavaScript handles the image selection, and PHP manages the upload process. The script ensures the image is in the correct format , compresses it if it’s too large, resizes it to the correct dimensions.</li><li><strong>Server to PC: </strong>the Midjourney API generates a description of the uploaded image and creates an AI-generated image based on this description. A Python script then downloads the picture. Using the ElevenLabs API, the script generates audio from the description. Finally, the AI-generated image is downloaded into the same folder.</li><li><strong>PC to Wall:</strong> For the local display the webpage is set up to show images and play the audio descriptions. A JSON file stores the paths to the folders containing these files. A Python script continuously checks for new image uploads and updates the JSON file with the paths to new folders whenever the carousel of images reaches its end.</li></ol><h3>Final result</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F945328512%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F945328512&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F1850645622_1280&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1080" frameborder="0" scrolling="no"><a href="https://medium.com/media/b3de30565f1ed7e110c9f751c1dc9ce8/href">https://medium.com/media/b3de30565f1ed7e110c9f751c1dc9ce8/href</a></iframe><h3>About team</h3><p>The team came together spontaneously, with each member bringing a unique set of skills and interests to the table. Despite our diverse backgrounds and areas of expertise, we shared a common goal: to create a high-quality product.</p><p>Throughout the project, we collaborated closely, engaging in collective research and brainstorming sessions. However, when it came to the development process, our roles naturally fell into place.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xk7hVKXHOTv4Yp5YR6jWNQ.png" /></figure><p>Of course, we were not alone. The supervisors and assistants of MainD program helped us through the whole way from idea to delivery.</p><h3>References</h3><ul><li>Cybernetic Forests. (2022). <em>How to Read an AI Image</em>. [online] Available at: <a href="https://www.cyberneticforests.com/news/how-to-read-an-ai-image">https://www.cyberneticforests.com/news/how-to-read-an-ai-image</a> [Accessed 7 Mar. 2024].</li><li>Ghili, S. (2021). <em>Can Bias Be Eliminated from Algorithms?</em> [online] Yale Insights. Available at: <a href="https://insights.som.yale.edu/insights/can-bias-be-eliminated-from-algorithms">https://insights.som.yale.edu/insights/can-bias-be-eliminated-from-algorithms</a>.</li><li>Heikkilä, M. (2023). <em>These new tools let you see for yourself how biased AI image models are</em>. [online] MIT Technology Review. Available at: <a href="https://www.technologyreview.com/2023/03/22/1070167/these-news-tool-let-you-see-for-yourself-how-biased-ai-image-models-are/">https://www.technologyreview.com/2023/03/22/1070167/these-news-tool-let-you-see-for-yourself-how-biased-ai-image-models-are/</a>.</li><li>IBM Data and AI Team (2023). <em>Shedding light on AI bias with real world examples</em>. [online] IBM Blog. Available at: <a href="https://www.ibm.com/blog/shedding-light-on-ai-bias-with-real-world-examples/">https://www.ibm.com/blog/shedding-light-on-ai-bias-with-real-world-examples/</a>.</li><li>Manyika, J., Silberg, J. and Presten, B. (2019). <em>What Do We Do About the Biases in AI?</em> [online] Harvard Business Review. Available at: <a href="https://hbr.org/2019/10/what-do-we-do-about-the-biases-in-ai">https://hbr.org/2019/10/what-do-we-do-about-the-biases-in-ai</a>.</li><li>Milne, S. (2023). <em>AI image generator Stable Diffusion perpetuates racial and gendered stereotypes, study finds</em>. [online] UW News. Available at: <a href="https://www.washington.edu/news/2023/11/29/ai-image-generator-stable-diffusion-perpetuates-racial-and-gendered-stereotypes-bias/#:~:text=AI%20image%20generator%20Stable%20Diffusion%20perpetuates%20racial%20and%20gendered%20stereotypes%2C%20study%20finds">https://www.washington.edu/news/2023/11/29/ai-image-generator-stable-diffusion-perpetuates-racial-and-gendered-stereotypes-bias/#:~:text=AI image generator Stable Diffusion perpetuates racial and gendered stereotypes%2C study finds</a>.</li><li>NVIDIA. (n.d.). <em>What is Generative AI?</em> [online] Available at: <a href="https://www.nvidia.com/en-us/glossary/generative-ai/">https://www.nvidia.com/en-us/glossary/generative-ai/</a>. skintone.google. (n.d.). <em>Skin Tone Research @ Google</em>. [online] Available at: <a href="https://skintone.google/">https://skintone.google/</a>.</li><li>Thomas, R.J. and Thomson, T.J. (2023). <em>Ageism, sexism, classism and more: 7 examples of bias in AI-generated images</em>. [online] The Conversation. Available at: <a href="https://theconversation.com/ageism-sexism-classism-and-more-7-examples-of-bias-in-ai-generated-images-208748#:~:text=There%20were%20also%20notable%20differences">https://theconversation.com/ageism-sexism-classism-and-more-7-examples-of-bias-in-ai-generated-images-208748#:~:text=There were also notable differences</a>.</li><li>Tiku, N., Schaul, K. and Chen, S.Y. (2023). <em>These Fake Images Reveal How AI Amplifies Our Worst Stereotypes</em>. [online] Washington Post. Available at: <a href="https://www.washingtonpost.com/technology/interactive/2023/ai-generated-images-bias-racism-sexism-stereotypes/">https://www.washingtonpost.com/technology/interactive/2023/ai-generated-images-bias-racism-sexism-stereotypes/</a>.</li><li>Turk, V. (2023). <em>How AI reduces the world to stereotypes</em>. [online] Rest of World. Available at: <a href="https://restofworld.org/2023/ai-image-stereotypes/">https://restofworld.org/2023/ai-image-stereotypes/</a>.</li><li><em>Humans are biased. Generative AI is even worse. </em>[online] Bloomberg. Available at:<a href="https://www.bloomberg.com/graphics/2023-generative-ai-bias/"> https://www.bloomberg.com/graphics/2023-generative-ai-bias/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ee6ec9b99c88" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>