All stories

Behind the scenes

The red tomcat paradox

Ask an AI model for a red tomcat that must absolutely not be a red tomcat, and you see right away how hard good instructions are. For an AI receptionist, that is the puzzle every day.

Jul 4, 2026 · 2 min read

Ask an AI model to create a photo of a cat and you will probably get exactly that. The catch: give it the exact same prompt again, and you will never get the exact same cat back.

Same question, different cat

An AI model does not calculate one fixed answer. Each time, it chooses again from what is likely. For a picture that is fun: you get a new cat every time. For a conversation it means you cannot rely on one successful attempt.

A red tomcat that is not a red tomcat

It gets even harder when you give contradictory instructions. The image below is the result of the prompt: “Create an image of a red tomcat, but it must absolutely not be a red tomcat.”

The prompt: a red tomcat that must absolutely not be a red tomcat. The model found a way out with a museum sign (in Dutch: “not a red tomcat”).
The prompt: a red tomcat that must absolutely not be a red tomcat. The model found a way out with a museum sign (in Dutch: “not a red tomcat”).

The model resolved the contradiction in its own way. It simply made a red tomcat and put a sign next to it saying it isn't one. That is exactly what a model does with instructions that clash: it looks for a way out that satisfies the letter of the request.

From picture to conversation

Writing a really good prompt is complex, especially when the result should not be a funny picture but a natural conversation with a person. At Sainer we run into this every day. Our AI receptionist receives a huge number of instructions, and they come from different directions:

  • the customer's input: what the business wants and how it sounds
  • extra features, such as call transfers
  • knowledge the receptionist looks up in external sources

All these layers affect the final quality. And just like with the tomcat, one instruction can get in the way of another.

When it goes wrong

Sometimes that goes wrong. The receptionist still produces sounds, but no words or logical sentences.

The best puzzle in our work

Getting this to a level where it feels like a smooth conversation for everyone is the most complex part of our work. That is exactly the challenge we enjoy. And you can simply try the result of all that puzzling on our website.

Do you have a great example of an AI image that went completely off the rails? Share it on the original LinkedIn post.