What makes you think it hasn't long been using a large language model for it though? Maybe not Large Language Model, the term seems to have come with ChatGPT et al., but Google's been doing machine learning for a long long time; the transformer architecture paper (kind of the groundwork for the current wave) was published by a Google team in 2017.

A lot of new Google features are branded 'AI', because it's so hyped it has broad consumer awareness, but a lot of Google features for a long time have used AI and just been brandless or at least 'AI'-less features.