I saw one about their lighting one time. I love that stuff, even if I don't understand it all.
You'll enjoy this, then. They talk about what their initial goals were, how they didn't want to "cheat" too much with the AI, and why they ultimately chose for Ellie (and other buddies) to not alert enemies.
Then they go into how the AI actually chooses where it positions itself. The player character apparently projects out "rays" every frame and they use those rays to figure out where walls and obstacles are in the environment. Then they compare that against other factors, such as where enemies are and what they can see, and out of that (and other variables) Ellie chooses where she should be relative to Joel.
Despite how complex it is, though, I can't help but think they would have been better off just having the AI mirror Joel's movements or hang back until enemies were in an "alert / combat" state. Then pull out the fancy AI until they go back to searching for Joel. Rinse and repeat.
Having said all that, I'll love seeing what the AI is like for a potential TLOU 2. Having watched that presentation, I learned the AI was largely written in only 5 months, which really surprised me.