AI Designed My Esperanto Banner—Then One Tiny Character Became the Biggest Problem

I thought designing a banner with AI would be the hard part.

It wasn't.

The hard part was getting AI to spell Esperanto correctly.

I recently helped create horizontal and vertical banners for the 58th Korean Esperanto Congress. I had last year's sample files and the required sizes, so I opened ChatGPT and started experimenting.

It quickly gave me several designs I liked.

I could try different layouts, adjust the colors, move elements around, and compare new versions without having to design everything from scratch. For someone who isn't a professional designer, that part felt surprisingly helpful.

🔗 I had already experimented with AI-generated graphics while creating a banner for my blog, but this time the design had to become a real banner for an actual event.

Then I looked more closely at the text.

That was where the real problem began.

The design looked good. The Esperanto didn't.

Esperanto uses letters such as ĉ, ĝ, ĥ, ĵ, ŝ, and ŭ.

Those small marks above the letters aren't decorative accents. They are part of the Esperanto alphabet, so getting them right matters.

As I refined the AI-generated banner, some of these characters kept appearing distorted or incorrect.

At first, I did what probably comes naturally when working with generative AI: I asked it to fix the problem.

Sometimes the result improved.

But regenerating the image wasn't a reliable solution. A character might look better in the next version, while another part of the design changed. Fixing one small detail could mean checking the whole banner again.

That became frustrating because I already liked the design.

I didn't want a completely new banner. I just wanted to correct the parts that were wrong.

And for an actual congress banner that would eventually be printed, “almost correct” wasn't good enough.

That's when something from my university classes suddenly became useful.

PNG, PPTX—and a lesson I finally understood

I'm studying AI at university in my 50s, and I had learned about file formats such as PNG, PDF, and PPTX in my computer graphics classes.

Until then, they mostly felt like technical terms I needed to remember.

Now I understood why they mattered.

The text inside my PNG design wasn't ordinary editable text. I couldn't simply click on an incorrect Esperanto character and replace it as I would in a document.

That meant endlessly regenerating the image was the wrong approach.

What I needed was a workflow where I could keep the design I liked but handle important text as something I could edit myself.

So I used ChatGPT Work to help me move from AI generation toward a more editable final design.

Instead of asking AI to recreate the entire banner every time one character was wrong, I could work with a PPTX version where the text could be corrected separately. I could type the Esperanto characters accurately, adjust their size and position, and keep refining the banner without throwing away the whole design.

The difference became very practical:

PNG (Image Format)	PPTX (Editable Format)
Good for keeping the finished visual appearance.	Better when text still needs correction.
Text is part of the image.	Text can be kept as editable elements.
A small text error may require image editing or regeneration.	Individual text can be corrected without redesigning everything.

The process wasn't perfectly automatic. Moving from an AI-generated image to an editable file also taught me that converting a finished image into separate editable elements has its own limitations.

But that was part of the lesson too.

For the first time, PNG and PPTX weren't just file extensions from a class.

They represented two very different ways of working with the same design.

And suddenly, something I had learned in class made sense in the real world.

AI got me started. Knowing what to do next finished the job.

This small project changed how I think about using generative AI.

AI helped me create design ideas much faster than I could have on my own. It gave me a starting point, let me experiment with layouts, and helped me get much closer to the banner I wanted.

But when the Esperanto characters went wrong, asking AI to “try again” wasn't enough.

The solution wasn't necessarily a better prompt.

I needed to understand what kind of problem I was dealing with and choose a better way to finish the work.

🔗 I learned something similar while building my Esperanto word game with Gemini Canvas: getting AI to create something is often only the beginning. Testing it, finding what doesn't work, and deciding what to change are part of the real work too.

That difference seems small, but it changed something for me.

A concept from my computer graphics class, an AI-generated image, an Esperanto character that wouldn't behave, and a real banner that needed to be printed suddenly connected.

Maybe that's what I'm really learning as an AI student.

Not how to make AI do everything for me—

but how to know what to do when it can't.