What Really Happens When AI "Learns" to Create Art
I thought it was pure magic until I dug deeper. Spoiler: it's way more complex and controversial than you'd think.
🎣 A few weeks ago I was scrolling through Reddit. Saw this crazy realistic image of an astronaut cat. Comments said "AI-generated". And that's when curiosity hit me. How the hell does a machine do that? 🤔
Turns out it's not just "click and done". There's a whole world behind it. And when I started digging, I found stuff that... well, let me show you what I discovered.
I always thought AI was like a fancy filter. But nope. It's more like teaching a kid who's never seen anything to draw. Except the "kid" processes millions of images in seconds.
That's how many images LAION-5B used to train some models. I can't even organize my phone's photo gallery 😅
This is roughly what the process looks like inside (obviously super simplified)
📚 The weirdest "school" in the world
Imagine you have to teach someone to draw. But this person has never seen anything in their life. Literally nothing. They don't even know what a circle is.
Well, that's how AI starts. And then comes the epic moment: you give it access to basically every image that exists on the internet. With descriptions. All at once.
The process (that blew my mind)
When I first tried it, I asked for "a dog wearing a hat". What happened next was incredible, but now I understand the process:
💡 The trick is in the critique: They use two AIs. One generates, another criticizes. "That doesn't look like a dog". "The hat is wrong". And the first one learns from the feedback. Just like us, really.
⚠️ Now comes the ugly part (but important)
Everything sounds great so far, right? But one day I stumbled upon a Twitter thread from a really pissed off artist. And that's when I understood there's a dark side to all this.
I think this is gonna explode soon. AI trains on millions of images that someone else created. Without permission. Imagine if someone took your work, "studied" it, and then competed against you. That's rough.
When I tried generating "CEO", I got mostly white men. When I asked for "nurse", mostly women. AI replicates the biases from the images it sees. And that's problematic as hell.
This kinda freaks me out. Even the engineers can't explain why AI sometimes generates weird stuff. It's like a black box that works... but we don't know exactly how.
Training a large model uses more energy than a small city in a year. Seriously. And then every time you generate an image, it also consumes energy. It's unsustainable long-term.
My takeaway after all this
Look, in my experience, AI art is incredible. But it's also creating a ton of problems that nobody wants to address. It's like having a super powerful car without brakes.
💡 The million-dollar question: How do we make AI a tool that adds value, instead of replacing or abusing other people's work?
I don't have all the answers. But I think understanding how it works is the first step. That's why I wrote this. So when you see an AI-generated image, you know everything that's happening behind the scenes.
💬 What do you think?
Did any of this surprise you? Have you used AI to create images? Share your experience in the comments. And if you think more people should know about this, share it around 👇
Explore the future impact of artificial intelligence on humanity.
🧠 This article was reviewed and edited to improve its clarity and presentation.