I am thoroughly convinced that faux-sapient AIs will take the form of a layering of multiple basic AIs: a bit that processes spoken language into a form that can then be turned into instructions by a different one, a bit for visual processing, etc all strung together with a control framework loaded with filters, predetermined actions, etc. Basically getting the complex-task part with a facade of awareness.

Or else they're just going to keep making bigger and bigger neural networks until they've got something with comparable processing power to a dog, but focused entirely on things like human language and human-relevant data instead of chemical detection, keeping a body functioning, etc like an actual living creature needs, and the nature of a machine like that is unpredictable. There's no guarantee that something designed from the ground up to perform menial tasks without complaint wouldn't be built with reward mechanism controls that incentivize obedience and successful completion of ordered tasks, and planning controls that physically prevent any sort of personal agency or initiative, even if most of it is a black box neural network.

That is to say, the people designing AIs to be servants will be doing the thermian propaganda "but they actually like being slaves!" fantasy bit that reactionary authors do, but in real life as engineers with similar power over their creations.