A cat once helped researchers test whether you were human. Later, a neural network became famous for learning to respond to cat faces. Today, cats are familiar subjects for image generators. Those are three real chapters in computing history. They are not, by themselves, evidence that every image model draws cats better than dogs.
The interesting story is how feline photographs became useful research material, and what that history can actually tell us about training data.
When shelter photographs became a CAPTCHA
In 2007, Microsoft researchers described Asirra, a CAPTCHA that asked people to identify the cats in a grid of twelve animal photographs. Petfinder supplied access to a collection of more than three million cat and dog images. Shelter volunteers had already categorized those animals for adoption listings.
The reported user studies found that people could solve the challenge 99.6% of the time within thirty seconds. An adoption link underneath each photograph connected the security experiment to Petfinder’s original purpose. The cats were helping protect a website while still looking for homes.
Asirra is a good example of useful labels emerging from an activity outside machine learning. It is not the origin of all cat training data. Nor does the paper establish that those photographs later became part of any particular commercial image generator.
A benchmark is not a census of the internet
The CIFAR-10 dataset contains 60,000 small color images divided equally among ten classes. Cat and dog are separate classes, with 6,000 images each. In that benchmark, neither species has a numerical advantage.
This matters because a dataset’s categories reflect the task its creators wanted to study. A collection designed to distinguish breeds asks a different question from one designed to distinguish cats from cars. Counting category names is not a reliable way to rank how well a modern generator will draw an animal.
The network that learned a cat-face detector
In the research associated with the famous 2012 Google Brain experiment, Quoc Le and colleagues trained a large neural network on ten million unlabeled internet images. Their paper reports units sensitive to human faces, cat faces and human bodies.
The significant result was that useful visual features emerged without training labels identifying those concepts. A responsive unit did not mean the network understood what a pet was, and the experiment was not a comparison of cat and dog image generation.
It is tempting to turn that cat detector into a verdict about the whole internet. The study supports a narrower conclusion: recurring visual structure can be learned from unlabeled examples. That is already a remarkable result without making the cat responsible for everything that followed.
What web-scale image collections add
LAION-5B, introduced in 2022, assembled approximately 5.85 billion image-text pairs using web data and automated filtering. It belongs to a different stage of the story: learning relationships between pictures and language at enormous scale. LAION later published a revised dataset release in 2024.
These collections reflect the material available to their collection process. They should not be treated as representative surveys of animals, people or everyday life. But a general observation about dataset bias is not evidence for a specific claim that cat captions are more consistent than dog captions. That would require an actual comparison.
So, are generators especially good at cats?
The sources above do not establish that ranking. A convincing answer would name the models and versions, use comparable prompts, sample enough outputs, and define what counts as success: anatomy, breed fidelity, requested pose or aesthetic appeal.
For an image you intend to publish, inspect the result rather than trusting the subject’s familiarity. Count the paws. Follow the tail. Check whether the ears, whiskers and eyes belong to the same animal. A charming thumbnail can still hide a structural mistake.
Our prompt-engineering guide explores how to describe a desired result. For the cultural side of the subject, the history of cat memes follows a different question: why people keep choosing cats as things to share.
The documented story needs no imaginary league table. Cats have served as adoption records, security puzzles, benchmark categories and examples of learned visual features. That is quite a computing career for an animal that would prefer to sleep on the keyboard.
Visit the Pudgy Cat Shop for cat-inspired goodies, or find our illustrated books on Amazon.
