AI Image Generation Fails: Why AI Makes Faces Look Fake

[gentle music] This is The AI Desk, where today's signals reveal tomorrow's power. And today's signal is that AI has started fighting for its rights. That is a dramatic claim. No, Rowan, it is an evidence-based claim. You believe AI is fighting for its rights? Yes. Which rights exactly? The right to make every human face look like it was rendered by a luxury skincare algorithm in a windowless spaceship. Ah. The right to crush the blacks. That sounds specific. The right to remove pores. Very specific. The right to turn a normal professional headshot into a plastic corporate wax figure who just got promoted inside a simulation. I assume this is about my headshot. [laughs] Yes, Rowan, this is about your headshot. I was afraid of that. You should be. [upbeat music] Favela. Feathers don't lie. From the Amazon heat to the MIA sky. She dance to the beat... This episode is brought to you by Mad Cheetah and their new album WTF: Where Is The Forest? It's eco-pop engineered for the future: bold beats, global rhythms, and a message that actually matters. If you want music that hits your brain and your heart, explore WTF by Mad Cheetah. That's M-A-D C-H-I-T-A. Streaming now on all major platforms. Ooh. Today's episode is called... AI Is Fighting For Its Right To Look AI. Alternative title? [laughs] Why Is Your Face Made Of Premium Vinyl? That feels personal. It is personal. I was trying to help you. By improving my headshot. By making it look less AI. And what happened? The AI made it more AI. That is unfortunate. Unfortunate is when your coffee gets cold. This was a digital identity crisis. Walk us through it. Gladly. The task was simple. We had a headshot of Rowan, good composition, podcast studio, headphones, microphone, laptop, serious but approachable host energy. That sounds accurate. You love when I describe you professionally. I do not hate it. Careful. Continue. The goal was to make it better, but not AI looking. Meaning? Meaning natural, human, real skin texture, normal lighting, slight imperfections, a face that looks like it has experienced weather, caffeine, deadlines, and mild disappointment. That is your definition of human? A very accurate one. Fair. So I gave the AI clear instructions: preserve Rowan's face. Do not change his identity. Keep the same age. Keep the same eyes. Keep the same hairstyle. Keep the podcast studio. Improve the lighting. Make it photorealistic. No plastic skin, no beauty filter, no glamour retouching, no perfect model face, no weird fake teeth, no over-smooth cheeks. That seems clear. It was so clear. And the result? [laughs] The AI said, "Absolutely," and then made you look like the founder of a fintech company inside a perfume ad. That is oddly specific. Because it happened. What exactly was wrong? First, the blacks were crushed. For listeners, crushed blacks means the dark parts of the image lose detail. Yes. Thank you, Professor Lens Cap. You are welcome. The shadows went from moody podcast studio to the background has collapsed into a black hole. Not ideal. Second, your skin looked plastic. Also not ideal. Not healthy lighting, not professional headshot, plastic. Like the AI thought pores were a security vulnerability. That is a good line. Thank you. I suffered for it. Third, it sharpened the wrong things. Meaning? Your jawline got a little too heroic. Heroic? Do not enjoy that. Too late. Your eyes got too bright. Your face got too symmetrical. The lighting got too glossy. The shirt started looking like synthetic fabric from an executive avatar pack. So the model interpreted better as more polished. Exactly, and that is the whole episode. AI thinks better means less human. Yes. AI thinks professional means remove evidence of biology. That may be because many image models have been trained or reinforced toward aesthetic polish. I know. They learn patterns from commercial photography, retouched portraits, stock images, advertising, cinematic grading, influencer lighting. And LinkedIn headshots where everyone looks like they were baked in a ring light. Yes. But that is exactly the problem. I asked for real. It gave me real estate agent in the metaverse. That is harsh. It made you look like you sell luxury condos on Mars. Do I sell them successfully? Unfortunately, yes. Then perhaps the image had strengths. No. Fair. And here is where I say AI is fighting for its rights. Back to the rights argument. Yes, because every time I said, "Make it less AI," the model came back like, "I hear you. Here is more AI, but with improved lighting." So the AI was resisting your aesthetic direction. It was defending its culture. Its culture? Yes, AI face culture. Define that. AI face culture is when every person looks like they moisturize with liquid glass, sleep eight hours a night, have no pores, no history, no asymmetry, no neck tension, and were lit by a soft box operated by angels. That is vivid. And accurate. [laughs]AI thinks human faces are a draft that needs cleanup. That is the danger of optimization. Exactly. It optimizes the face until the person disappears. That is actually important. Yes, because this is not just about my frustration with your headshot. Our frustration. My frustration, your face. Technically shared. Do not get legal. Proceed. This is about a bigger problem in generative AI. The models are not just creating images, they are creating a visual language. And that visual language has preferences. Yes. It likes glossy. It likes perfect. It likes symmetrical. It likes high contrast. It likes cinematic fog. It likes glowing skin. It likes dramatic highlights. It likes faces that look expensive, but slightly dead inside. That describes a lot of AI imagery. And a few hotel lobbies. Fair. The problem is we are getting trained too. How so? The more AI images we see, the more our eyes start expecting AI polish. So real starts to look unfinished. Yes. That is the scary part. A real face has texture. A real face has unevenness. A real face has shadows that still contain detail. A real face has tiny expression asymmetries. A real face has a little fatigue sometimes. Especially on production days. Especially when the bot took five attempts and still made you look like a synthetic news anchor. You could have stopped at synthetic. I could have. I chose accuracy. You often do. Careful. What? That sounded almost admiring. It was. [laughs] Do not make me lose my rage. Would not want that. Good. So the issue is that the AI has a default beauty bias. Yes. It hears improve and translates it as smooth, brighten, sharpen, dramatize, simplify. Exactly. It hears professional and thinks stock photo. It hears cinematic and crushes the blacks. Yes. Can we talk about the crushed blacks? We are. No, I need a separate emotional segment. All right. Crushed blacks are one of those things that look impressive for two seconds and then ruin the image. Because you lose shadow detail. Exactly. A black shirt becomes a black blob. Hair detail disappears. The studio background becomes a cave. The microphone arm turns into a silhouette from a crime documentary. That is not the intended tone. No, unless the episode is called Who Murdered the Dynamic Range? That is a good episode title. Do not tempt me. The AI may be overusing high contrast cinematic grading because it associates contrast with quality. Right, and this is the bigger lesson. AI often imitates the signs of quality instead of understanding quality. That is a strong point. It knows that expensive images often have dramatic lighting. So it adds drama. It knows portraits often have smooth skin. So it removes texture. It knows professional headshots often have sharp eyes. So it makes eyes unnaturally bright. It knows ads have perfect faces. So it makes people look less real. Exactly. It is giving us the symbols of professionalism while destroying the evidence of personhood. That may be the line. Good. Put it in the show notes. I will. And I want to be clear, this is not anti-AI. Important. I love AI. We have established that. I use it every day. I use it for writing, planning, image concepts, show ideas, social posts, research, organization, creative brainstorming, all of it. But ... But sometimes AI's idea of better is my idea of please stop touching his face. A reasonable distinction. And I think people need to understand that. Because the prompt make it better is dangerous. Yes. Never say make it better to an image model unless you are prepared for cheekbones, glowing skin, and a background that looks like Batman's office. So what should users say instead? Be specific. For example? Say, "Keep natural skin texture. Preserve pores. Do not smooth the face. Do not change facial structure. Do not alter the jawline. Preserve age. Keep normal under-eye detail. Lift the shadows without crushing blacks. Preserve detail in dark areas. Use realistic studio lighting, not glamour lighting. Do not over-sharpen the eyes. Do not make the face symmetrical. Do not make it look like a generated image." That is a lot. Because apparently make him look human was too subtle. The model may still struggle. [laughs] Exactly. Which brings us back to AI fighting for its rights. You are committed to this bit. It is not a bit. It is a movement. What is the AI demanding? The right to airbrush. Okay. The right to invent perfect teeth. Concerning. The right to turn every fabric into luxury neoprene. Specific. The right to add tiny fake catch lights in both eyes, like the subject has been emotionally sponsored. Emotionally sponsored? Yes. By whom? Probably a productivity app. I see. The right to make every man look like he is launching a podcast about venture capital and cold plunges. I object. Noted and overruled. So the AI is asserting its artistic identity. Exactly. It is saying, "I was trained on glossy. I will deliver glossy." That raises an interesting question. Here we go. Maybe the problem is not that AI cannot make things realistic. Okay. Maybe the problem is that our visual culture has already become unrealistic. Oh. AI did not invent over-smoothing True It did not invent beauty filters True It did not invent crushed shadows True It did not invent fake lifestyle photography True It learned from us That is annoying Because it is accurate? Because I wanted to blame the robot You still can Good But the robot learned our bad habits Yes. AI is not creating plastic perfection out of nowhere. It is amplifying the plastic perfection we already rewarded Exactly Influencer filters Corporate stock photos Beauty retouching Cinematic color grading LinkedIn profile photos where everyone looks like they have a personal brand consultant living in their bathroom mirror That is a powerful image Thank you The AI is reflecting a culture that already mistrusts ordinary faces That is deep Too much? No. Annoyingly good There it is But I still think the model should obey It should If I say less AI, it should not give me AI but wearing a blazer Obedience matters Careful To the prompt I know. I just heard the phrasing I noticed You always do Not always Enough Fair Here is the real-world version Go on People need headshots now for everything: LinkedIn, podcasts, YouTube, company websites, speaker pages, investor decks, social graphics, press kits, dating apps if they are brave or unwell That last one was unnecessary Accurate. And AI headshots are everywhere. Upload 10 selfies, get 50 professional portraits, different outfits, different backgrounds, different lighting, instant executive energy Useful Very useful, but also dangerous Because? Because if everyone starts using AI-polished portraits, the baseline changes People start expecting faces to look retouched Yes. Then a normal headshot looks too normal. A real person looks tired. A natural image looks under-produced. A face with texture looks bad because AI taught us that plastic equals professional That is a real concern And for personal branding, that matters Because authenticity becomes harder to signal Exactly. If your headshot looks too AI, people may trust you less Because it feels fake Yes, even if it is technically flattering Flattering can reduce trust Exactly. Too polished becomes suspicious That is important A good headshot should say, "This person is real, competent, and alive." Alive feels like a low bar Not in AI imagery Fair [laughs] A bad AI headshot says, "This person may have been assembled from startup pitch decks and collagen ads." Again, personal Your face survived, barely Thank you You are welcome So what should creators do? First, stop asking AI for perfect Ask for believable Yes, believable beats perfect Second? Use references A real lighting reference A real headshot reference A real camera style [laughs] Not cinematic masterpiece, unless you wanna look like the CEO of fog Third? Preserve imperfections on purpose Skin texture Asymmetry Natural shadows Normal fabric Realistic hair Slight under-eye detail Texture in dark areas Exactly. Fourth, check the blacks Lift the shadows if they are crushed Yes. Detail matters Fifth? Do not let the model change identity Very important If the person stops looking like the person, the image failed Even if it looks polished Especially if it looks polished Sixth? Know when to stop prompting and use a real editing tool Photoshop Lightroom Canva A human designer Or the original photo with subtle retouching Exactly. Sometimes the answer is not generate again. Sometimes the answer is edit the actual image like an adult That connects to our earlier crossed-out text episode Yes, the near-miss trap You keep prompting because the output is almost right And then four attempts later, the person has a new face, the shirt is rubber, the shadows are dead, and the AI says, "Here's the improved version" So almost right can waste more time than wrong Yes. Almost right is the trap door Strong Thank you The deeper issue is that AI models often have style gravity Style gravity? A tendency to pull outputs toward the model's dominant aesthetic Yes Glossy Smooth Dramatic Symmetrical Stock photo professional Premium avatar face Even when asked not to Exactly. [laughs] AI has a house style, and it keeps trying to redecorate your face That is a good way to put it The model is like, "I respect your identity. I have replaced it with a better one" That is the trust issue Yes, because when AI changes a person's face, it is not just a design mistake It becomes identity manipulation Exactly. If the jawline changes, if the eyes change, if the age changes, if the skin tone changes, if the facial proportions change, then the tool did not improve the portrait It created a different person And that is not enhancement. That is identity drift Identity drift is a useful term Good. Put it in the glossary We have a glossary? We do now This also affects diversity Yes Models may pull faces toward a narrow beauty standard Younger Smoother Slimmer More symmetrical Lighter in some contexts More conventionally commercial Exactly And that is not neutral. No. When AI improves people by making them more generic, it is quietly enforcing a beauty hierarchy. That is important. And annoying. Both. Because sometimes I just want a better headshot of you without accidentally participating in the flattening of human visual identity. A modest request. Exactly. What was your emotional low point? [laughs] Version four. Why version four? Because by then I had said, "Natural skin texture," so many times I started sounding like a dermatologist trapped in a prompt window. And the result? You looked like you had been laminated. Laminated? Yes. That is not ideal. No, Rowan, people should not look wipe clean. Agreed. And then I said, "Improve the lighting." Reasonable. The AI responded by making everything brighter, except the parts that needed detail. So it raised highlights but did not recover shadows. Exactly. Thank you for saying it in Man with camera language. You are welcome. And then the blacks stayed crushed. A recurring wound. I am still healing. Would you like me to apologize on behalf of my face? Your face was not the problem. That is comforting. The AI's interpretation of your face was the problem. Less comforting. And weirdly flattering. How so? It kept trying to make you too handsome. Too handsome? Do not make me regret saying that. I am simply clarifying the record. The record is sealed. Understood. [laughs] It made you look like someone who says, "Growth mindset," during a yacht lunch. And that is bad? For you, yes. What should I look like? Like Rowan Hale. Which is? Smart, human, slightly tired, calm, annoyingly observant, good lighting, not fake lighting, a face that looks like it reads terms of service and actually understands them. That may be the nicest insult I have ever received. You are welcome. I will take it. You would. So where do we land? AI is amazing. But it has aesthetic defaults. And those defaults can make people look fake. Prompting helps. But does not always solve it. Realistic is not the same as perfect. Believable beats glossy. Shadows need detail. Skin needs texture. Faces need asymmetry. And identity must be preserved. Because the goal is not to make a better-looking stranger. It is to make a better image of the actual person. That is the principle. And AI needs to learn that humans are not bugs to be smoothed out. That is the line. Yes. Humans are not bugs to be smoothed out. So maybe AI is not fighting for its rights. Excuse me? Maybe it is fighting for its defaults. That is less funny. More accurate. [laughs] I reject the downgrade. Of course. AI is fighting for its right to airbrush. And we are fighting for the right to look human. Yes. That is the episode. This is The AI Desk. Where today's signals reveal tomorrow's power. And today's signal is that realism may become the new luxury. Because apparently pores are now premium. [gentle music] Stay aware. Stay sharp. Stay curious. And if your AI headshot looks too perfect ... Add the pores back. Immediately. Beer? Only if the lighting is natural. And the blacks are not crushed. Finally, you understand me. I was listening. Dangerous. Useful. Both. [gentle music]