AI and ML models are trained on large datasets so they can spot patterns and make predictions on new, unseen data. Underfitting happens when a model fails to learn those patterns properly. It cannot capture the underlying complexity in the data and therefore performs poorly on both training and test sets, showing high bias and consistently incorrect or oversimplified predictions.

A model can underfit for several reasons. Its architecture may be too simple for the problem, it may not have enough features (variables), or it may have been trained for too few iterations. Excessive regularization, introduced to prevent overfitting, can also push a model into underfitting by forcing it to ignore important patterns.

Underfitting shows up as recommendation systems serving irrelevant suggestions, or chatbots responding with vague, generic answers. It can mislead users and erode trust in AI systems. Sometimes underfitting is confused with bad training data, and in some cases developers even tolerate mild underfitting to avoid overfitting. Striking the right balance between simplicity and accuracy remains central to building reliable AI models.