AI is bloody useful. I use it every day. Research, tightening up phrasing, pressure-testing ideas, fact checking, checking angles I might have missed – adding balance to opinion pieces. It earns its keep.
But some of the most hyped AI features are the ones that fall apart the second you lean on them properly.
Image generation is a perfect example.
I write long articles. Proper ones. Words first. Thinking first. The image is just decoration. Eye candy so Google doesn’t sulk. Most of the time the image doesn’t even matter. It just needs to roughly represent the idea without looking like it was stolen from a 2006 stock library.
So I’ll open up Gemini and use Nano Banana (formerly Imagen). And to be fair, when it gets it right, it’s very good. Clean. Detailed. Strong lighting. No obvious weirdness. It follows initial instructions well.
Then it misunderstands any follow-up prompt.
Not slightly. Completely.
When the Penny Drops
So I tweak it. I’m specific. Clear. “No, this isn’t what I asked for. Change the angle. Remove that object. Make the tone darker.” You know, normal iterative creative process.
And it gives me… the same image again.
Sometimes almost pixel for pixel identical.
You try again. You rephrase. You simplify. You over-explain. You under-explain. And it stubbornly spits out a slightly rearranged version of the original mistake. It’s like arguing with a very confident intern who refuses to admit they misheard you.
Yet if you go on YouTube, you’d think this model had descended from the heavens holding tablets of stone.
Thousands of videos. Thumbnails with wide eyes and glowing text.
- “INSANE NEW IMAGE MODEL.”
- “THIS CHANGES EVERYTHING.”
- “BETTER THAN MIDJOURNEY?”
Does it though?
Or does it work brilliantly when you demo it with a carefully pre-tested prompt that you already know produces a strong result?
The Demo Is Not the Product
There’s a difference between a tool working well in a controlled demo, and a tool working reliably in messy, real-world creative flow.
Most of the hype lives in the first category. The friction lives in the second.
And it’s not just misunderstanding prompts. I’ve had Gemini regenerate an image in the same thread, show me the updated version on screen, then when I hit download, it sends me the previous iteration.
The wrong one.
The old one.
Think about that.
We’re talking about data centres burning through more electricity than small towns. Custom silicon. Billion-pound infrastructure. And it can’t reliably hand me the latest image in the thread.
That isn’t a philosophical limitation of AI. That’s plumbing.
And plumbing matters.
YouTubers don’t talk about that. Because “sometimes it downloads the wrong file” doesn’t get 300,000 views.
What gets views is confidence. Certainty. Big claims. Hype. BS.
There’s also a wider issue. A lot of AI demos are built around novelty, not repeatability. The first generation looks impressive because your brain fills in the gaps. You forgive the errors. You’re still in the wow phase.
Use it daily for three months and the shine wears off. You start noticing the patterns. The stubbornness. The times it ignores a constraint you explicitly gave it. The way it occasionally pretends to understand a follow-up instruction but clearly hasn’t.
There was a period, I think back in 2023 and 2024 when you could instantly spot images created with specific AI image generators; yes, you could name exactly which generator created the image based on a look.
And I’m not anti-AI. Far from it. I’ve built a chunk of my workflow around it. I pay for multiple tools. I defend it when people lazily dismiss it.
Defending it doesn’t mean pretending it’s flawless.
Why Iterative Prompting Is Still a Mess
If anything, the people who use it seriously are the ones most aware of its weak spots.
Image models struggle with iterative refinement. That’s not controversial. They don’t “remember” your intent the way a human designer would. Each prompt is a probabilistic nudge inside a latent space (probabilistic is typical of AI), and sometimes that space just keeps snapping back to the first solution it found.
That’s not magic. It’s maths. And maths doesn’t care about your blog header.
The problem isn’t that the tools have limitations. Everything does. The problem is the gap between what’s being marketed and what you actually experience.
If you’re new to AI and you’re watching those videos, you’d assume you’re the problem. You’d think, “Why can’t I get these results?”
Often it’s because the demo was cherry-picked. Or because the creator doesn’t use it in the same messy, iterative way you do. Or because YouTubers care more about watch time than nuance.
AI is brilliant. It’s also temperamental. Occasionally clumsy. Sometimes weirdly stubborn.
Both things can be true at the same time.
And if we’re going to use it properly, build businesses around it, rely on it for creative work, we need more honesty about that.
Not doom. Not blind worship. Just reality.
Because reality is far more useful than hype.
Still Got Questions? Here’s the Bit YouTube Won’t Tell You
Is AI image generation actually useful for everyday content creation?
It can be, yes, but it depends heavily on how you use it. For one-shot prompts where you have a clear, specific idea and you’re happy with the first result, it’s quick and often impressive. Where it struggles is when you need to refine and iterate, changing small details across multiple attempts. That’s where the wheels tend to come off.
Why do AI image tools keep ignoring my instructions?
It’s a known limitation. Image models don’t retain intent across prompts the way a human would. Each instruction is essentially a fresh attempt to nudge the model in a new direction, and sometimes the model just snaps back to its first interpretation. It’s not stubbornness exactly, it’s just how the underlying maths works.
Are AI image generation demos on YouTube misleading?
A lot of them are, yes, though not always deliberately. Most demos use pre-tested prompts that the creator already knows will produce a strong result. That’s very different from real-world use where you’re working with new ideas, awkward briefs, and iterative changes. The gap between a polished demo and daily use can be significant.
Does using AI tools daily make you more critical of them?
Generally, yes. The novelty fades after a few weeks and you start noticing the patterns, the quirks, and the consistent failure points. That’s not a bad thing. It means you’re using the tools properly rather than just being impressed by them. The people who know AI best tend to be the most clear-eyed about where it falls short.
Should I still use AI image tools despite the limitations?
If they fit your workflow, absolutely. The limitations are real but they’re not dealbreakers for everyone. If you need a decent header image quickly and you’re not precious about getting it pixel-perfect, they do the job. Just go in with realistic expectations rather than the ones YouTube has been selling you.