Во-во.

LaMDA are transformer-based neural language models trained on both a text corpus and on conversations that have been manually annotated for sensibility, appeal, and safety.

Последнее особенно существенно, т. к. без ручного управления будет GIGO.