If Tesla etc hires me I can think of ways to go past that approach. And I’ve thought about how to go past that approach years ago. So real pros of the pros have already surpassed it for sure. For instance look at this figure on how the semantics of a sentence can be broken down - https://media.geeksforgeeks.org/wp-content/uploads/20200329230855/Syntax1.png Beyond simple sentences like ‘the cat is red’ the possibilities for these break downs become very numerous very fast. Ambiguity happens extremely quick. Historically this is a path finding type solved problem not a machine learning one. Couple this with the standard 2-word or 3-word approach you’re alluding to and you can start making sure the bot is making grammatically correct sentences. Start doing AI on this syntax level as well, you might start getting some genius level sentence creation. Couple this with a syntax tree break down of sentence to sentence break down of topic progression and it might start getting really hard to tell it’s a bot anymore. Take this even further... you can start using “live learning” where the bot actually starts holding its own “opinions” and stances on topics etc. Like how they do with LSTM machine learning networks.