What happens when AI engines get too excited?

Michael Chalk
Jul 31
3 min read

A few weeks ago, I stepped into a creative rabbit hole. What started as a seemingly simple request — to generate a single promotional image of my four published books in 3D — ended up as a surprisingly drawn-out journey through the wonderful, bewildering quirks of AI engines grappling with the generation of images.

I’ve grown used to how well my AI engine handles language and text. Whether it’s editing narrative text, refining blog posts, or drafting emails, its precision and responsiveness have been rock solid. So, I expected the same fluency when we moved into image generation.

Spoiler alert: it didn’t quite go that way!.

The Request: Simple on Paper

All I wanted was a clean, high-quality 3D-style promotional image. My instructions were clear:

Zachary’s Cry to be standing upright on the left.
On its right, a stack of three books, lying flat with spines facing out:
1. A Moment of Madness
2. The Unravelling
3. Like Feathers in the Wind

The final image would be used on my website and across social media. I already had the 2D covers. This should have taken ten minutes.

What Actually Happened

Let me summarise a few missteps:

Only two books in the stack instead of three — despite repeated instructions.
Incorrect order — Like Feathers in the Wind kept floating to the top.
Spelling errors — We had LIKE FEATHERS IN T WIND and INE WIND at different points.
The red cover of A Moment of Madness inexplicably became beige.
Lighting gone rogue — either too dim or too warm, obscuring Zachary’s Cry altogether.

Each time, we went back and forth, refining the prompts and adjusting the expectations. Finally — after more than a dozen iterations — we nailed it. The final image looks great. But the path there was anything but smooth.

Why This Happens

As my AI assistant later explained, there's a big difference in how it handles language and text versus image prompts:

“Image generation involves probabilistic models interpreting natural language as visual composition. That introduces more variation and less deterministic control.”

In other words, generating text is relatively easy - like baking a cake from a recipe. Generating an image is however more complex — rather like asking an artist to paint your dream — and explaining your dream to an artist who is prone to forget previous discussions!

The Lesson: Be Patient, Precise — and Ready to Iterate

Image generation AIs are incredibly powerful, but they’re still learning how to interpret nuanced instruction sets. They don’t yet have a visual memory of what went wrong in the last draft — unless you tell them. They don’t see the missing book or the spelling error unless it's explicitly called out.

Eventually, we got what we wanted. The spine now reads: LIKE FEATHERS IN THE WIND. All three books are in place. The lighting hits Zachary’s Cry with just enough dramatic punch. And yes — the red has returned to A Moment of Madness.

Looking Ahead

One day, I suspect image generation will be just as smooth and context-aware as language models. Maybe even built by the same team. When that happens, it’ll transform the way authors, marketers, and creatives work.

But for now? If you’re diving into AI image generation, bring your patience. And maybe a checklist.

Examples

The images below show how the process evolved: -

Image one - stand alone books