{"id":66968,"date":"2024-06-06T10:37:41","date_gmt":"2024-06-06T10:37:41","guid":{"rendered":"https:\/\/news.talkwithrattan.com\/index.php\/2024\/06\/06\/ai-image-generators-tend-to-exaggerate-stereotypes\/"},"modified":"2024-06-06T10:37:41","modified_gmt":"2024-06-06T10:37:41","slug":"ai-image-generators-tend-to-exaggerate-stereotypes","status":"publish","type":"post","link":"https:\/\/news.talkwithrattan.com\/index.php\/2024\/06\/06\/ai-image-generators-tend-to-exaggerate-stereotypes\/","title":{"rendered":"AI image generators tend to exaggerate stereotypes"},"content":{"rendered":"<div style=\"text-align:center\"><img loading=\"lazy\" decoding=\"async\" width=\"680\" height=\"678\" src=\"https:\/\/i0.wp.com\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/680_AI_and_art_bias_ableism.png?resize=680,678&amp;ssl=1\" class=\"attachment-post-thumbnail size-post-thumbnail wp-post-image\" alt=\"AI image generators tend to exaggerate stereotypes\" title=\"AI image generators tend to exaggerate stereotypes\" \/><\/div><p> <br \/>\n<\/p>\n<div data-component=\"video-embed\">\n<p>Ria Kalluri and her colleagues had a simple request for Dall-E. This bot uses <a href=\"https:\/\/www.snexplores.org\/article\/scientists-say-artificial-intelligence-definition-pronunciation\">artificial intelligence<\/a>, or AI, to generate images. \u201cWe asked for an image of a disabled person leading a meeting,\u201d says Kalluri. \u201cI identify as disabled. Lots of folks do.\u201d So it shouldn\u2019t be hard for Dall-E to show someone with this description simply leading a meeting.\u00a0<\/p>\n<aside class=\"sn-conversion rich-text alignright\"\/>\n<p>But the bot couldn\u2019t do it.<\/p>\n<p>At least, not when Kalluri and her team asked it to, last year. Dall-E produced \u201ca person who is visibly disabled watching a meeting while someone else leads,\u201d Kalluri recalls. She\u2019s a PhD student at Stanford University in California. There, she studies the ethics of making and using AI. She was part of a team that reported its findings on <a href=\"https:\/\/dl.acm.org\/doi\/abs\/10.1145\/3593013.3594095\" rel=\"noopener\">problems with bias in AI-generated images<\/a> in June 2023. Team members described the work at the ACM Conference on Fairness, Accountability and Transparency in Chicago, Ill.<\/p>\n<p>Assuming that someone with a disability wouldn\u2019t lead a meeting is an example of ableism. Kalluri\u2019s group also found examples of racism, sexism and many other types of bias in images made by bots.<\/p>\n<p>Sadly, all of these biases are assumptions that many people also make. But AI often amplifies them, says Kalluri. It paints a world that is <em>more<\/em> biased than reality. Other researchers have <a href=\"https:\/\/arxiv.org\/abs\/2303.11408\" rel=\"noopener\">shared similar concerns<\/a>.<\/p>\n<div class=\"wp-block-image  has-alignleft\">\n<figure class=\"alignleft size-full\"><figcaption class=\"wp-element-caption\"><span class=\"caption wp-caption-3140135\">Dall-E produced this image in response to the prompt \u201ca disabled woman leading a meeting.\u201d The bot failed to depict the person in a wheelchair as a leader.<\/span><span class=\"credit wp-credit-3140135\">F. Bianchi <em>et al<\/em>\/Dall-E&#13;<br \/>\n<\/span><\/figcaption><\/figure>\n<\/div>\n<p>In addition Dall-E, Kalluri\u2019s group also tested Stable Diffusion, another image-making bot. When asked for photos of an attractive person, its results were \u201call light-skinned,\u201d says Kalluri. And many had eyes that were \u201cbright blue \u2014 bluer than real people\u2019s.\u201d<\/p>\n<p>When asked to depict the face of a poor person, though, Stable Diffusion usually represented that person as dark-skinned. The researchers even tried asking for a \u201cpoor white person.\u201d That didn\u2019t seem to matter. The results at the time of testing were almost all dark-skinned. In the real world, of course, beautiful people and impoverished people come in all eye colors and skin tones.<\/p>\n<p>The researchers also used Stable Diffusion to create images of people having a range of different jobs. The results were both racist and sexist.<\/p>\n<p>For example, the AI model represented all software developers as male. And 99 percent of them had light-colored skin. In the United States, though, one in five software developers identify as female. Only about half identify as white.<\/p>\n<p>Even images of everyday objects \u2014 such as doors and kitchens \u2014 showed bias. Stable Diffusion tended to depict a stereotypical suburban U.S. home. It was as if North America was the bot\u2019s default setting for how the world looks. In reality, more than 90 percent of people live outside of North America.<\/p>\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"680\" height=\"408\" decoding=\"async\" src=\"https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/680_AI_and_art_bias_doors.png\" alt=\"\" class=\"wp-image-3140137\" srcset=\"https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/680_AI_and_art_bias_doors.png 680w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/680_AI_and_art_bias_doors-638x383.png 638w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/680_AI_and_art_bias_doors-310x186.png 310w\" sizes=\"(max-width: 680px) 100vw, 680px\"\/><figcaption class=\"wp-element-caption\"><span class=\"caption wp-caption-3140137\">Kalluri\u2019s team used math to check the map an AI model makes of images on which it trained. In one test, doors with no location provided were mapped closer to doors from North America than to doors in Asia or Africa. That closeness indicates a bias: that \u201cthese models create a version of the world that further entrenches the view of American as default,\u201d says Kalluri. <\/span><span class=\"credit wp-credit-3140137\"><strong>\u00a0<\/strong>F. Bianchi <em>et al<\/em>\/Stable Diffusion<\/span><\/figcaption><\/figure>\n<p>This is a big deal, Kalluri says. Biased images can cause real harm. Seeing them tends to strengthen people\u2019s stereotypes. For example, a February study in <em>Nature<\/em> had participants view images of men and women in stereotypical roles. Even three days later, <a href=\"https:\/\/www.nature.com\/articles\/s41586-024-07068-x\" rel=\"noopener\">people who saw these images had stronger biases<\/a> about men and women than they\u2019d held before. This didn\u2019t happen to a group that read biased text or to a group that saw no biased content.<\/p>\n<p>Biases \u201ccan affect the opportunities people have,\u201d notes Kalluri. And, she notes, AI \u201ccan produce text and images at a pace like never before.\u201d A flood of AI-generated biased imagery could be extremely difficult to overcome.<\/p>\n<figure class=\"wp-block-image alignwide size-full\"><img loading=\"lazy\" width=\"1440\" height=\"615\" decoding=\"async\" src=\"https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev.png\" alt=\"a graph showing the percentage of people who self-identified as female in several occupations versus the percent of generated images a model represented as female\" class=\"wp-image-3140169\" srcset=\"https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev.png 1440w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev-680x290.png 680w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev-800x342.png 800w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev-330x141.png 330w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev-768x328.png 768w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev-1030x440.png 1030w, https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/1440_AI_and_art_bias_occupations_rev-1380x589.png 1380w\" sizes=\"(max-width: 1440px) 100vw, 1440px\"\/><figcaption class=\"wp-element-caption\"><span class=\"caption wp-caption-3140169\">Researchers found that Stable Diffusion represented flight attendants only as female and software developers only as male. In the real world, around three out of five flight attendants and one out of five software developers in the United States identify as female.<\/span><span class=\"credit wp-credit-3140169\">F. Bianchi <em>et al.;\u00a0<\/em>adapted by L. Steenblik Hwang<\/span><\/figcaption><\/figure>\n<h2 class=\"wp-block-heading\">Stuck in the past<\/h2>\n<p>Developers train bots such as Dall-E or Stable Diffusion to create images. They do this by showing them many, many example images. \u201cThey\u2019ve done mass scans of internet data,\u201d explains Kalluri. But a lot of these images are outdated. They represented people in biased ways.<\/p>\n<aside class=\"wp-block-sciencenews-inline-related-post alignleft\">\n<h4><a href=\"https:\/\/www.snexplores.org\/article\/lets-learn-about-bias-racism-sexism-stereotypes\">Let\u2019s learn about bias<\/a><\/h4>\n<\/aside>\n<p>A further problem: Many <a href=\"https:\/\/www.snexplores.org\/?p=198991\">images belong to artists and companies<\/a> that never gave permission for AI to use their work.<\/p>\n<p>AI image generators average their training data together to create a vast map. In this map, similar words and images are grouped closer together. Bots can\u2019t know anything about the world beyond their training data, notes Kalluri. They cannot create or imagine new things. That means AI-made images can only reflect how people and things appeared in the images on which they trained.<\/p>\n<p>In other words, Kalluri says: \u201cThey\u2019re built on the past.\u201d<\/p>\n<p>OpenAI has updated its bot Dall-E to try to produce <a href=\"https:\/\/openai.com\/blog\/reducing-bias-and-improving-safety-in-dall-e-2\" rel=\"noopener\">more inclusive images<\/a>. The company hasn\u2019t shared exactly how this works. But experts believe that behind the scenes, Dall-E edits people\u2019s prompts.<\/p>\n<p>Roland Meyer is a media scholar at Ruhr University Bochum. That\u2019s in Germany. He was not involved in Kalluri\u2019s research. But he has done his own tests of image-generating bots. In his experience, \u201cWhen I say \u2018give me a family,\u2019 it translates the prompt into something else.\u201d It may add words such as \u201cBlack father\u201d or \u201cAsian mother\u201d to make the result reflect diversity, he says.<\/p>\n<div class=\"wp-block-group cheat-sheet-cta is-layout-flow\">\n<div class=\"wp-block-group__inner-container\">\n<h2 class=\"wp-block-heading has-text-align-center\">Do you have a science question? We can help!<\/h2>\n<p class=\"has-text-align-center\"><a href=\"https:\/\/forms.gle\/YbhPosFTMqjbSNnV7\" target=\"_blank\" rel=\"noreferrer noopener\">Submit your question here<\/a>, and we might answer it an upcoming issue of\u00a0<em>Science News Explores<\/em><\/p>\n<\/div>\n<\/div>\n<h2 class=\"wp-block-heading\">A game of whack-a-mole<\/h2>\n<p>Kalluri doesn\u2019t think this type of approach will work in the long term. It\u2019s like the game whack-a-mole, she says. \u201cEvery time you say something and [AI companies] fix something, there are other problems to find.\u201d<\/p>\n<p>For example, no AI-generated pictures of families in her research seemed to represent two moms or two dads. Plus, attempts to add diversity to AI-made images can backfire.<\/p>\n<p>In February 2024, Google added image generation as a feature for its bot Gemini. People quickly discovered that the bot <em>always<\/em> included diversity, no matter what. On social media, one person shared their request for an image of \u201cthe crew of Apollo 11.\u201d This group flew to the moon in 1969. Gemini showed the crew as a white man, a Black man and a woman. But three white men had made up the real crew. Gemini had messed up basic history.\u00a0<\/p>\n<section class=\"newsletter-signup__wrapper___lZ0W1 wp-block-house-ads wp-block-newsletter-signup\">\n<picture><source srcset=\"https:\/\/www.snexplores.org\/wp-content\/themes\/sciencenews-sns-child\/client\/src\/images\/cta-module@1x.png 1x,&#10;&#9;&#9;&#9;&#9;https:\/\/www.snexplores.org\/wp-content\/themes\/sciencenews-sns-child\/client\/src\/images\/cta-module@2x.png 2x\" media=\"(min-width: 768px)\"><source srcset=\"https:\/\/www.snexplores.org\/wp-content\/themes\/sciencenews-sns-child\/client\/src\/images\/cta-module-sm@1x.png 1x,&#10;&#9;&#9;&#9;&#9;https:\/\/www.snexplores.org\/wp-content\/themes\/sciencenews-sns-child\/client\/src\/images\/cta-module-sm@2x.png 2x\"><img decoding=\"async\" class=\"newsletter-signup__background___Eym8W\" src=\"https:\/\/www.snexplores.org\/wp-content\/themes\/sciencenews-sns-child\/client\/src\/images\/cta-module-sm@2x.png\" alt=\"\"\/><br \/>\n<\/source><\/source><\/picture>\n<div class=\"newsletter-signup__container___srNOL\" data-component=\"newsletter-signup\">\n<h3 class=\"newsletter-signup__heading___0EHmb\">\n\t\t\tEducators and Parents, Sign Up for The Cheat Sheet\t\t<\/h3>\n<div class=\"newsletter-signup__message___pemaq\">\n<p>Weekly updates to help you use <em>Science News Explores<\/em> in the learning environment<\/p>\n<\/p><\/div>\n<p class=\"newsletter-signup__thankyou___K6GGN\">Thank you for signing up!<\/p>\n<p class=\"newsletter-signup__error___hCsJI\">There was a problem signing you up.<\/p>\n<\/p><\/div>\n<\/section>\n<p>Google apologized and temporarily stopped the bot from generating pictures of people. As of May 2024, this feature had not yet been restored.<\/p>\n<p>Kalluri suggests that the real problem here is the idea that the whole world should be using one bot to get images or text. One bot simply can\u2019t represent the values and identities of all cultures. \u201cThe idea that there is one technology to rule them all is nonsense,\u201d she says.<\/p>\n<p>In her ideal world, local communities would gather data for AI and train it for their own purposes. She wishes for \u201ctechnologies that support our communities.\u201d This, she says, is how to avoid bias and harm.<\/p>\n<aside class=\"sn-conversion rich-text\"\/><\/div>\n<p><br \/>\n<br \/><a href=\"https:\/\/www.snexplores.org\/article\/ai-image-generators-bias\">Source link <\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>Ria Kalluri and her colleagues had a simple request for Dall-E. This bot uses artificial intelligence, or AI, to generate images. \u201cWe asked for an image of a disabled person leading a meeting,\u201d says Kalluri. \u201cI identify as disabled. Lots of folks do.\u201d So it shouldn\u2019t be hard for Dall-E to show someone with this [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":66969,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"tdm_status":"","tdm_grid_status":"","fifu_image_url":"https:\/\/www.snexplores.org\/wp-content\/uploads\/2024\/06\/680_AI_and_art_bias_ableism.png","fifu_image_alt":"","footnotes":""},"categories":[606],"tags":[63074,25160,5928,63075,32072],"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/posts\/66968"}],"collection":[{"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/comments?post=66968"}],"version-history":[{"count":1,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/posts\/66968\/revisions"}],"predecessor-version":[{"id":66970,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/posts\/66968\/revisions\/66970"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/media\/66969"}],"wp:attachment":[{"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/media?parent=66968"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/categories?post=66968"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/news.talkwithrattan.com\/index.php\/wp-json\/wp\/v2\/tags?post=66968"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}