A pretrained base model is trained only to predict the most likely next words, so it continues the prompt's pattern instead of answering it — translating "The small dog" becomes "The small dog crossed the road." An aligned model, trained after pretraining to be helpful, answers the actual request. Same prompt, two objectives: likely versus good.

Likely is not the same as good

A base model only learned to predict the most likely next words. Nobody taught it to be helpful. Flip the switch and watch the same prompt produce two very different things.

your prompt

most likely continuation

base model

success = predicting the next word (likely)

aligned model

success = a good answer (helpful + honest + harmless)

The base model is not broken. It is doing its one job perfectly — guessing what text usually comes next. Being helpful is a separate skill, added after pretraining. That extra step is called post-training (alignment).