Picture data shaped like a smile: high on the left, low in the middle, high on the right. Fit a straight line and the best it can do is run fairly flat through the middle. It predicts too low at both ends and too high in the middle, and no choice of slope or intercept fixes that, because a line cannot bend. The model is too simple for the pattern, and that is underfitting.
The scores give it away. An underfit model does poorly on its training data, say 55%, and about equally poorly on new data, say 54%. There is no gap, because the model memorized nothing; it never captured the pattern in the first place. Its residuals have a shape too: for the smile they are positive at both ends and negative in the middle. Switch to a quadratic, which can bend, and both scores should rise together.
Low scores on new data alone do not mean underfitting. A model that scores 99% on training but 61% on new data learned too much, not too little: that is overfitting, and its cure is the opposite, a simpler model. Underfitting is also different from noisy data. If the data itself is very random, even the right model leaves large residuals, but those residuals scatter without any pattern.
Think of it like this. An underfit model is like describing a mountain range as 'about 2,000 metres high'. The number is not crazy, but it misses every peak and valley, and repeating it more confidently does not help. You need more detail, such as a contour map. The comparison has a limit built in: too much detail, mapping every pebble, becomes the opposite problem, overfitting.
Core idea. Underfitting is when a model is so basic it misses the real structure and does poorly everywhere.