Don't get LLMs confused with specialized ML algorithms. Hallucination is an LLM problem, algorithms like gait recognition have been honing in accuracy since way before LLMs started development. Where LLMs come into the picture is that they can act as agents, processing queries and then selecting the best fit specialized algorithm to process the data and then cross reference results from different queries to compile a correlated multidomain dataset. Done properly, this will yield not just a single answer but a list of potential answers with their relative degree of certainty.

Look at the Harvard facial recognition glasses as a proof of concept of this kind of approach: https://specialconcentrations.fas.harvard.edu/news/heres-looking-you