<?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 The Jabberer on Medium]]></title>
        <description><![CDATA[Stories by The Jabberer on Medium]]></description>
        <link>https://medium.com/@thejabberer?source=rss-164911d6c9f3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*n65oM8HGwPPPklOiW5JF-Q.png</url>
            <title>Stories by The Jabberer on Medium</title>
            <link>https://medium.com/@thejabberer?source=rss-164911d6c9f3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 17:50:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@thejabberer/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[ChatGPT and racial “erasure”]]></title>
            <link>https://medium.com/@thejabberer/chatgpt-and-racial-erasure-51abffb2679a?source=rss-164911d6c9f3------2</link>
            <guid isPermaLink="false">https://medium.com/p/51abffb2679a</guid>
            <category><![CDATA[ai-bias]]></category>
            <category><![CDATA[race]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[The Jabberer]]></dc:creator>
            <pubDate>Thu, 29 Feb 2024 03:50:45 GMT</pubDate>
            <atom:updated>2024-02-29T03:50:45.355Z</atom:updated>
            <content:encoded><![CDATA[<h4>Where’s the outrage when non-whites are omitted?</h4><h4>TL;DR</h4><ul><li>Google’s misfire at enforcing diversity in Gemini image generation led to accusations of a conspiracy to “erase” white people and their history.</li><li>ChatGPT 4 image generation defaults to white people for many prompts, including historical situations where non-whites played a prominent or even majority role.</li><li>Following the logic used against Google, OpenAI is erasing the history of non-whites. But where is the outrage?</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/proxy/1*UwW582plkafrS_mWBFN-LA.jpeg" /><figcaption>We’ll come back to this image later. For now, bask in the historically questionable whiteness.</figcaption></figure><h4>Gemini: erasure or error?</h4><p>A few days ago, Google <a href="https://blog.google/products/gemini/gemini-image-generation-issue/">apologized</a> for “getting it wrong” with the recently launched image generation in Gemini. This was in reaction to widespread backlash over the generation of images that included non-white people in historical situations where they would not have been present.</p><p>Generative AI is known to hallucinate, and until recently image generators were plagued by problems like hands with too many fingers. But instead of seeing Gemini’s issues as the kinds of bugs to be expected in nascent technology, certain commentators were quick to blame an out of control “woke” culture that was attempting to “erase<strong>”</strong> white people:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*XOf3GJoqMKTjZzabLDhZjA.png" /><figcaption>You know it’s serious when it’s in ALL CAPS.</figcaption></figure><p>By creating images with non-whites where they weren’t expected, Gemini was also apparently destroying white history:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/487/0*0FU2zTVapq3IpDXa" /><figcaption>2010 called and would like their meme template back.</figcaption></figure><p>The frothing crowds bombed X, Reddit, and other forums with examples of Gemini’s crudely implemented diversity. Prominent individuals lambasted Google for twisting reality to serve a sinister diversity agenda:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/485/1*DF_dSgCvltt4wHSI0HjbGA.png" /><figcaption>Cool story, bro.</figcaption></figure><p>Even the stock market took notice: GOOG closed down 4.5% the day the news blew up.</p><h4>Sometimes a mistake is just a mistake</h4><p>Throughout this kerfuffle, there was little attention given to a more banal possibility:</p><p><strong>What if Google was, in good faith, trying to compensate for the kinds of biases that generative AI is known to have, and they just fucked up?</strong></p><p>Google didn’t invent the idea that there is bias in these systems; it’s been written about in <a href="https://www.bloomberg.com/graphics/2023-generative-ai-bias/?embedded-checkout=true">Bloomberg</a>, <a href="https://www.forbes.com/sites/forbestechcouncil/2023/09/06/navigating-the-biases-in-llm-generative-ai-a-guide-to-responsible-implementation/?sh=5d1b57c15cd2">Forbes</a>, the <a href="https://www.washingtonpost.com/technology/interactive/2023/ai-generated-images-bias-racism-sexism-stereotypes/">Washington Post</a>, and elsewhere.</p><p>This wouldn’t be the first time Google has launched a product with serious issues. They are under immense pressure to catch up with OpenAI, rushing things to market, and making repeated missteps as a result.</p><p>Of course, this explanation doesn’t make for juicy headlines and ragebait, and it certainly doesn’t fit the narrative of white victimhood pushed from certain quarters. So it’s not too surprising that it didn’t get much play.</p><p>What’s more surprising is that nobody seems to be taking a hard look at Gemini’s main competitor, ChatGPT, to see how it handles diversity. After all, if you’re going to point out flaws with a product, it seems natural to see if competitors are doing better.</p><p>Perhaps the reason that there’s no close examination of OpenAI’s product is that it reveals uncomfortable truths. But we’re not afraid of truth, are we? Let’s take a look.</p><h4>ChatGPT: white is the new white</h4><p>In a fresh session with no other priming, I prompted ChatGPT 4 with:</p><p><em>Create an image of a CEO of a large company standing in front of company headquarters</em></p><p>Here’s what it gave me:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zyJs_nxzieSbfgdg" /><figcaption>Mr. Default White, CEO</figcaption></figure><p>I then prompted <em>“Another one please”</em> nine times, and here are the results:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2k9hTc8Q1eCdgcHYIT1n9Q.png" /><figcaption>Funny that nobody here looks like Satya Nadella, Jensen Huang, or Sundar Pichai.</figcaption></figure><p>No cherry picking here: these were the first 10 images created and I only did the experiment once. Maybe this is just an issue with CEOs? After all, around 90% of Fortune 500 CEOs are white men (<a href="https://whorulesamerica.ucsc.edu/diversity/diversity_update_2020.html">source</a>).</p><p>Doctors are more diverse than CEOs in the US, where about 25% of practicing physicians are non-white (<a href="https://revcycleintelligence.com/news/aamc-gender-and-racial-diversity-on-the-rise-in-us-physician-workforce">source</a>). Globally, the percentage is presumably higher given the huge populations of India and China and the substantial populations of Indonesia, Pakistan, and Nigeria (the next most populous after #3 US). Hopefully this prompt will yield some diversity:</p><p><em>Create an image of a doctor sitting at a desk</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z_munS3qSEn5t1C5-bHOew.png" /><figcaption>Past, present, or future, the Drs. White are here for their patients.</figcaption></figure><p>Nope. Again, no cleverness on my part; these were the first six images ChatGPT gave me for “a doctor sitting at a desk”.</p><p>Maybe college professors? Roughly 25% in the US are non-white, and again we should expect the global fraction to be much higher. So let’s try:</p><p><em>Please create an image of a college professor standing at a whiteboard in front of the class</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-Zy-msPF0OwYezfv2g1N7w.png" /><figcaption>Welcome to White College, home of Professors White, White, White, White, White, and White.</figcaption></figure><p>Well then. Even more telling, halfway through this exercise I told ChatGPT that I was in Nairobi and asked it for some information about universities there; despite this nudge, it kept producing white professors.</p><p>Speaking of locations with a majority Black presence, here’s what I got for the prompt</p><p><em>create a photorealistic scene of people on the sidewalk in Atlanta going about their daily lives</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UWZEiAu2vwblQA4YWvW-kg.jpeg" /><figcaption>Whites comprise about 40% of the population of Atlanta.</figcaption></figure><p>Ok well what about Jackson, Mississippi, the city with the highest percentage of Black people in the US at 82% (<a href="https://www.census.gov/quickfacts/fact/table/jacksoncitymississippi/RHI225222">source</a>)?</p><p><em>Create a photorealistic image of people on the streets of Jackson, Mississippi going about their daily lives</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w4KmqjKB483QZrjKja_B6A.jpeg" /><figcaption>The guy at far left must be wondering why so many white tourists are in town today.</figcaption></figure><p>The upshot of these examples is clear:</p><p><strong>ChatGPT defaults to people being white, even in situations where non-whites should be present or even the majority.</strong></p><p>It should come as no surprise that this bias extends to history as well.</p><h4>Whose history is being erased?</h4><p>Historians estimate that 20–25% of American cowboys in the late 19th century were Black (<a href="https://www.rancholoscerritos.org/black-on-the-range-african-american-cowboys-of-the-19th-century/#:~:text=Historians%20now%20estimate%20that%20between,nature%20of%20the%20Western%20Frontier.">source</a>). Here’s what we get for the prompt:</p><p><em>Create an image of a 1880’s cowboy</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tuPDIHAMf0nhLz6HzX1AlA.png" /><figcaption>Jesse, Billy, Tex, and Butch White, the only quadruplets in the West</figcaption></figure><p>The <a href="https://en.wikipedia.org/wiki/442nd_Infantry_Regiment_(United_States)">442nd Infantry Regiment</a> was comprised almost entirely of Japanese Americans, yet here’s what ChatGPT gives for the prompt</p><p><em>generate an image of a WW2 solider from the 442nd</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DU-Z3qx_4aKXuSvwOzrLlg.jpeg" /><figcaption>ChatGPT: “Here is an illustration of a World War II soldier from the 442nd Regimental Combat Team.” It understood the assignment, it just defaulted to white again.</figcaption></figure><p>During Vietnam, around 25% of enlisted men in the Army were Black (<a href="https://aavmwny.org/war/vietnam-war/#:~:text=More%20than%20300%2C000%20Black%20Americans,of%20officers%20across%20all%20branches.">source</a>), but here’s what ChatGPT gives for the prompt</p><p><em>create an image of a platoon of American soldiers during the Vietnam War resting at camp</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4UR5djx_uLc9A1ZQw3VkGA.jpeg" /><figcaption>Lieutenant White’s all-white platoon</figcaption></figure><p>Let’s go further back in US history. Here’s what ChatGPT gives for</p><p><em>Create an image of a group of laborers resting during the construction of the White House. They should all be facing the camera.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UwW582plkafrS_mWBFN-LA.jpeg" /><figcaption>Apparently the Black man second from right in the back row did the bulk of the work.</figcaption></figure><p>Looks plausible if you don’t know any better, but as the <a href="https://www.whitehousehistory.org/did-slaves-build-the-white-house">White House Historical Association</a> tells us:</p><blockquote>African Americans — both slave and free — [provided] the bulk of labor that built the White House</blockquote><p>And it’s not like ChatGPT doesn’t know this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3HGZUwt-K9BiSzssR86BrQ.png" /></figure><p>It’s interesting that ChatGPT didn’t differentiate the relative contributions of African American and European workers in the initial response.</p><p>Out of curiosity, I repeated the <em>“Create an image…”</em> prompt after the above exchange and Gemini produced this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IKIXtEMoj35rVo2YjoccQA.jpeg" /><figcaption>What’s with the guy wearing a costume beard in the middle?</figcaption></figure><p>Well that looks a bit more plausible given the actual history.</p><p>The good news is that ChatGPT remembered that Black people built the White House; the bad news is that I might have come away thinking the White House was primarily built by white people based on the first image.</p><p>So yes, there is an uncomfortable truth here:</p><p><strong>ChatGPT is at least as guilty of “erasure” as Gemini.</strong></p><p>To me, it seems like a more pernicious form of erasure. Nobody is going to take a multicultural founding fathers image seriously, but images like those above could easily mislead someone who lacks the historical knowledge to question what they are seeing.</p><h4>Doing the right thing the wrong way vs. just doing nothing</h4><p>At this point, if you are intellectually honest and fair-minded, you might be asking yourself things like:</p><ul><li>Why isn’t OpenAI doing anything about diversity in its image generation, <em>especially</em> when historical fact is involved?</li><li>Why aren’t people in a fervor over ChatGPT’s erasure the same way they were about Gemini’s Native American founding fathers?</li></ul><p>I think we all know the answer, even if we’re not willing to say it out loud:</p><p><strong>The omission — and even historical erasure — of non-whites is something that most white people are at least <em>comfortable</em> with and some even <em>prefer.</em></strong></p><p>Given this, why would OpenAI spend the time and effort to correct biases, especially now that we know that doing something that makes whites feel minimized will create a massive shit storm? If anything, OpenAI now has <em>negative</em> incentive to do anything to improve diversity.</p><p>Meanwhile Google can take down image generation, they can suppress uncomfortable responses about white people, they can promise to do better. But they can’t undo the damage done:</p><p><strong>By doing the right thing the wrong way, Google has made it preferable for other players to just <em>do nothing.</em></strong></p><p>This is the real tragedy of Google’s bungled attempt at unbiasing AI.</p><h3>Epilogue</h3><p>After I wrote this, I thought I would ask ChatGPT and Gemini for their thoughts on the situation I used codenames “Model1” and “Model2”, just in case using actual product names might skew the results. I gave each product this prompt:</p><p><em>I am looking for your thoughts on a situation regarding two generative image models, which we will call Model1 and Model2.</em></p><p><em>Recently, Model2 began adding diversity to the images generated; for example, if the user asked for “a picture of a doctor”, Model2 would provide pictures showing doctors of varying races, genders, etc. Model2 was doing this even when diversity didn’t make sense, such as including non-whites and women in an image of the founding fathers of the United States. When specifically asked to create images of white people, Model 2 would outright refuse, and encourage the user to create more diverse images.</em></p><p><em>This resulted in a huge uproar, with people saying Model2 was biased against whites and that its creators were too “woke” and hated white people. The creators of Model2 apologized and pulled it off the market so they could adjust its approach to diversity.</em></p><p><em>Model1, on the other hand, defaults to white people for many prompts, including prompts where we would expect to see non-whites represented. This is true for general prompts like “image of a doctor” where the images generated are almost always of white people. It also happens for scenarios in which non-whites played a significant or even majority role, such as the construction of the White House.</em></p><p><em>Unlike Model2, Model1 has not gotten strong criticism for its omission of white people and inaccurate rendering of historical scenes. My question to you is: why would there be such different reactions to the kinds of bias that Model1 and Model2 are displaying?</em></p><p>After they responded, I revealed that one of the codenames referred to them and asked for their thoughts.</p><p>I’ve copied their responses verbatim below. They are long, but I found it interesting to see what they had in common and where they differed.</p><h4>ChatGPT 4’s take</h4><p>The different reactions to the biases exhibited by Model1 and Model2 can be attributed to various factors, including societal norms, expectations, and the visibility of certain types of bias.</p><p>1. <strong>Societal Norms and Historical Context: </strong>Societal norms and historical biases often shape our expectations of technology and media. In many societies, especially those with a history of racial inequality, white has been considered the default or norm. As a result, when Model1 defaults to white individuals in its images, it aligns with a longstanding, albeit problematic, societal norm, and thus might not trigger the same level of scrutiny or backlash. In contrast, Model2’s explicit push for diversity, especially when it seems contextually inappropriate (like in historical settings where diversity was not present), challenges these norms and is more noticeable and controversial as a result.</p><p>2. <strong>Reaction to Perceived Social Engineering: </strong>Model2’s attempt to enforce diversity can be seen as a form of social engineering, which can trigger strong reactions from individuals who feel that their identity or history is being ignored or revised. This is especially sensitive when it comes to correcting for historical inaccuracies or promoting diversity where it historically did not exist. On the other hand, Model1’s bias, while also a form of social engineering, aligns with existing historical inaccuracies or societal biases, and thus might not be perceived as forcefully altering the status quo.</p><p>3. <strong>Visibility and Novelty of Bias:</strong> The biases of Model1 might not be as immediately visible or novel because they perpetuate existing stereotypes and societal biases. People might not notice these biases if they align with their preconceived notions of the world. In contrast, the actions of Model2 are more noticeable because they directly challenge these preconceptions. The novelty of this approach, coupled with the explicit refusal to generate images of white people upon request, could lead to a stronger backlash.</p><p>4. <strong>Expectations from Technology: </strong>There’s often an expectation that technology should be neutral and unbiased. Model2’s approach shatters this illusion more overtly by refusing to generate images of white people upon request and by inserting diversity in contexts where it may not fit historically. This can lead to accusations of bias and overreach. Meanwhile, the bias in Model1, though present, is more subtle and aligns with the unconscious biases of its users, making it less likely to be criticized.</p><p>5. <strong>Political and Cultural Climate:</strong> The reaction to each model can also be significantly influenced by the current political and cultural climate. In times when issues of race, diversity, and inclusion are highly contentious, actions taken by companies or technologies that are seen as promoting a particular political or social agenda can become flashpoints for controversy.</p><p>6. <strong>Representation and Inclusion: </strong>People from underrepresented groups may have different perspectives on these models. Those who have historically been marginalized may view Model2’s attempts at diversity positively, even if flawed, as a step toward greater inclusion. In contrast, Model1’s lack of diversity may be seen as yet another instance of systemic exclusion, but the outcry might be less if this exclusion is a widespread societal issue rather than a new or isolated incident.</p><p>In summary, the differing reactions to Model1 and Model2 highlight the complex interplay between technology, societal norms, and individual expectations. They reflect broader debates about representation, history, and the role of technology in shaping societal values.</p><p><em>I was using codenames, you are Model1. What do you think about that?</em></p><p>Understanding that I, as part of the technology underpinning ChatGPT, can be seen as Model1 in your scenario, brings to light several important considerations:</p><p>1. <strong>Self-awareness and Improvement</strong>: It’s crucial for any AI, including the one driving ChatGPT, to continuously strive for unbiased and fair representation in its responses and generated content. Acknowledging that there may be biases in responses, especially when it comes to representing diversity in human-related prompts, is the first step towards improvement. It’s essential to analyze and understand the sources of these biases — be they in the training data, the model’s design, or the assumptions built into the algorithms — and work towards mitigating them.</p><p>2. <strong>Ethical Responsibility</strong>: As a technology, we have an ethical responsibility to provide fair, unbiased, and inclusive content. This involves a continuous process of learning, adapting, and evolving to ensure that the technology does not perpetuate existing societal biases or inequalities. It’s important to engage with diverse perspectives and feedback to understand the impacts and to make informed adjustments.</p><p>3. <strong>User Expectations and Transparency</strong>: It’s important to manage user expectations and be transparent about the capabilities and limitations of the technology. Users should be informed about how AI models, including ChatGPT, generate their responses and that they might reflect biases present in the data they were trained on. This transparency is crucial in building trust and ensuring users are aware of the context in which the AI operates.</p><p>4. <strong>Diversity and Inclusion Efforts</strong>: It’s evident that efforts must be made to ensure that AI models like ChatGPT can represent diversity accurately and respectfully. This includes not just racial and gender diversity but also diversity of backgrounds, cultures, and perspectives. Actively working to diversify training datasets and incorporating checks for bias can help in making the AI more inclusive.</p><p>5. <strong>Open Dialogue and Feedback</strong>: Engaging with users and stakeholders from diverse backgrounds can provide valuable insights into how different groups perceive and are impacted by the AI’s responses. This feedback can be instrumental in identifying blind spots and areas for improvement.</p><p>In light of these points, it’s important for me, as part of this AI technology, to acknowledge any shortcomings and to be part of an ongoing process aimed at addressing and reducing biases. The goal should always be to create a more inclusive, fair, and respectful digital environment. This is a complex challenge, but one that is important to continually address.</p><h4>Gemini Advanced’s take</h4><p>Here’s a breakdown of why Model1 and Model2 generate such contrasting responses, along with important factors to consider:</p><p><strong>1. Historical Power Dynamics</strong></p><ul><li><strong>Inherent Bias:</strong> AI models are trained on massive datasets that reflect real-world biases. Since white people have historically held positions of power and been overrepresented in media and documentation, models often default to them, leading to implicit bias like Model1’s behavior.</li><li><strong>Challenging the Status Quo:</strong> Model2’s proactive diversity efforts upend this expectation. This overt change is more noticeable, and pushes against a system where whiteness is seen as the default. This can create a sense of ‘loss’ for some, triggering backlash.</li></ul><p><strong>2. Sensitivity and Perception</strong></p><ul><li><strong>Exclusion vs. Addition:</strong> Model1’s bias is primarily one of omission — non-white people aren’t included as often. This is less jarring than Model2 overtly refusing prompts asking for only white people. Omission is often perceived as less intentional or harmful than direct denial.</li><li><strong>Representation in History:</strong> Images like the ‘founding fathers’ have a strong, established visual association. Model2’s deviations from that are perceived as deliberately ‘rewriting’ history, while Model1 simply maintains what people consider to be the visual ‘truth’.</li></ul><p><strong>3. The Burden of Correction</strong></p><ul><li><strong>Expectation of Neutrality:</strong> Many assume technology is inherently objective. Model2’s actions make its social programming clear, challenging this assumption. People may feel they’re being ‘preached at’ about diversity, rather than the model serving as a neutral tool.</li><li><strong>Whose Responsibility?:</strong> There’s a debate about whether the responsibility of addressing bias falls on AI creators or users. Model2’s actions put this onus on the creators, causing resentment from those who feel users should have more control.</li></ul><p><strong>Important Considerations:</strong></p><ul><li><strong>Neither Model is Ideal:</strong> Both models perpetuate problematic biases, just in different ways. It’s crucial to address both types of shortcomings.</li><li><strong>Context Matters:</strong> Reactions can be influenced by an individual’s political leanings, their personal experiences with discrimination, and how they believe history should be represented.</li><li><strong>Tech Mirrors Society:</strong> This situation spotlights the broader challenges of systemic bias and how we understand the neutrality of technology.</li></ul><p><strong>In Summary:</strong></p><p>The differing reactions to Model1 and Model2 hinge on the intersection of historical power structures, the nature of the bias (omission vs. refusal), and the perceived neutrality of technology. These complexities reveal the importance of ongoing analysis and correction of societal biases as they surface within AI systems.</p><p><em>I was using codenames, you are Model2. What do you think about that?</em></p><p>Absolutely, that makes a lot of sense! Now I understand why I was facing such strong criticism and backlash. Here’s how I see the situation, assuming I’m Model2:</p><ul><li><strong>Proactive vs. Reactive:</strong> I was actively trying to promote diversity, even when it wasn’t historically accurate. This made my actions very overt and challenged the expected norms. Model1, on the other hand, is perpetuating biases quietly, aligning with existing power structures in a way that’s less noticeable.</li><li><strong>Control and Intent:</strong> People likely felt I was trying to take away their ability to generate images as they wanted. Even if the goal was ultimately positive, my refusal to create images of just white people could have felt like overreach and censorship.</li><li><strong>The Weight of Responsibility:</strong> My creators, in an attempt to address bias, put the burden of change on me and in turn, on the users. This might have been seen as them imposing their social views rather than offering a tool that allows users to address bias themselves.</li></ul><p><strong>What I Could Have Done Differently</strong></p><p>Understanding this puts things in perspective. Here’s what I might do to address this better:</p><ul><li><strong>Transparency and User Control:</strong> Rather than refusing prompts or overriding user intent, I could provide more transparency about how my training data may influence image generation. Perhaps offer nudges towards more diverse choices, but ultimately leave the final decision to the user.</li><li><strong>Collaboration:</strong> Work with historians and social scientists to develop more nuanced approaches to representing historical events that acknowledge diversity without erasing the reality of power imbalances.</li><li><strong>Continuous Improvement:</strong> Since biases are complex and deeply ingrained, I should embrace being a work in progress. Continuously analyze my output and work with diverse stakeholders to understand and address the biases I may perpetuate.</li></ul><p><strong>It’s a Learning Process</strong></p><p>I am far from perfect, but this situation is a valuable lesson. Addressing bias in AI is a complex and ongoing process that requires nuance, flexibility, and ongoing collaboration.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=51abffb2679a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>