This morning I was trying to generate a simple accessibility modification for a backyard deck.
I wanted something like a small gentle ramp so an elderly person in a wheelchair could get onto the deck more easily.
Instead, I kept getting images like this:
The ramp is technically correct, but completely wrong for the homeowner!
My First Assumption: The Model Isn’t Good Enough
After about 10+ trials, I concluded that ChatGPT is just not good enough. Perhaps I should find a better model that's especially designed for image generation?
After sleeping on it for a day or two, I have decided to "think like AI". If it has been trained by ramp images, what would general ones look like?
Even though the incorrect images are all different, they all had the same underlying idea:
- long pathway
- extensive handrails
- institutional appearance
- accessibility-first design
The AI wasn’t being random. It was being remarkably consistent. That meant the problem might not be the model at all.
Investigating the Word “Ramp”
I searched Google Images. Most results looked surprisingly similar to the AI outputs.
The common theme was: long, extended slop, large handrails. It spells safety and accessibility, not residential elegance.
At that point I started wondering: if image models learn patterns from large collections of internet images, perhaps “wheelchair ramp” strongly correlates with these kinds of designs.
Changing the Prompt
So I made two changes.
- First, I removed the word: ramp and replaced it with: slope
The goal was to remove the strong association with commercial accessibility infrastructure.
- Second, I supplied a reference image showing the kind of small timber transition I actually wanted.
The generated design became much closer to my intent! Not because the model suddenly became smarter. Because the instructions became clearer. The AI no longer had to guess what kind of ramp I meant.
Final result:
And then I asked it to match the colour of the deck and put the handrail on the other side:
Now it’s perfect:)
What I Learned
The interesting lesson wasn’t about landscaping. It was about how AI systems interpret language.
When I used the word “ramp”, the model wasn’t hearing “elegant backyard modification”. It was hearing: “the average concept of a wheelchair ramp found across millions of images”. Once I understood that, the solution became obvious:
- choose words carefully
- look for patterns in failures
- provide examples when possible
Conclusion
I almost dismissed GPT and tried to signed up for a more expensive app. Instead, I solved it by understanding how the model was likely interpreting my prompt. Sometimes the next step isn’t a better model. Sometimes it’s spending ten minutes asking:
“What does this word mean to the AI?”
Top comments (2)
I completely agree — it's easy to blame AI when you can't even manage to provide a clear, well-formulated request… In the end, AI might just help us rediscover the richness of language 🙃
Yes I still have a lot to learn on how AI works:) I have to be a lot more precise now in my descriptions!