Is overfitting always a bad thing?
The answer is a resounding yes, every time. The reason being that overfitting is the name we use to refer to a situation where your model did very well on the training data but when you showed it the dataset that really matter(i.e the test data or put it into production), it performed very bad. This can never be good, because we write models so that they can predict with high accuracy on the datasets it has not seen before. No one cares about a model that got 100% accuracy on the training set and gets 0% on the test set. This is a useless model. We train models so that they can predict with high accuracy on data it has never seen before.
Now for the long answer:
I will start by giving a quick explanation of what overfitting is. Then how we can detect overfitting, this will help me explain why overfitting is allways a bad.
What is overfitting:
In machine learning problem there is signal then there is noise. Signal is the true underlying pattern that you wish to learn from the data whist noise is the irrelevant information or randomness in a dataset.
The reason that leads to overfitting is that your model is making use of "Noise" to make decisions instead of using the signal. Overfitting means that your model performs very well on the training data and performs very poorly on unseen dataset(s) (notice that we pretty much don't care how well a model performs on the training data, we care about how it will perform on data it hasn't seen before). A model that has overfitted is similar to a student who got 100% on the quiz he/she did in preparation for the exam and will get a very poor mark on the exam because he/she because he chose to cram/memorize the quiz's answers and questions instead of understanding how to get to those answers. Another way of thinking about this is that your model has memorized/crammed instead of learning the learning the general relationship contained in the training dataset.
How do we detect Overfitting:
The only way to know if your model is overfitting is to test its performance on a dataset it has never seen before. As such, overfitting is something you know with the benefit of hindsight following testing a model on data it has never seen before.
For this usage, where the answer is much simpler than recognizing a cat, isn't overfitting a good thing?
No, because if your model is overfitting it will perform bad on the test set(in your case, it will perform very bad when you deploy it in a real world ACAS and start showing it data it hasnt seen before). see How do we detect Overfitting above.
If it turns out that your model performs very well when you show it data it hasn't seen before, then that means it didn't overfit (see What is overfitting above).
What would overfitting "look like" in this case and why would it be bad ?
Your model will get very high accuracy on the training set and get very poor accuracy when you start showing it examples it hasn't seen before. So in short, you will get your hopes high, jump up and down in excitement thinking that your model is great only for it no get very poor results in a production ACAS system.