Voice assistants launched with enormous ambition and settled into setting timers and playing music. The reason was simple: they matched keywords to commands, and anything outside the command list failed.
What changed
Language models understand what was meant rather than what was said. A request phrased naturally, with context, with a correction mid-sentence, gets interpreted rather than rejected.
That is a different category of capability, and the products built on it are only beginning to appear.
What still does not work
Public use. Nobody wants to talk to a device in a queue, an office, or a train. The interface is private by nature, which limits it to home, car and headphones.
Anything visual. A voice interface cannot show you options, a map, or a comparison. It works for tasks with one answer and fails for tasks that require choosing.
Latency. A conversation requires responses within a second or so, and a model that takes three seconds to think breaks the rhythm.
Where it wins
Hands-busy contexts – driving, cooking, working with tools. Accessibility, where it is not a convenience but the only option. And any task where describing what you want is faster than navigating to it.
The privacy question
A device that listens continuously in order to respond is a device that hears everything. Local processing addresses part of that and adds cost. The honest position is that convenience and surveillance are the same feature, and each product draws its own line.
A realistic forecast
Voice becomes one input among several rather than the primary one. Useful, integrated, and never the thing people reach for first when a screen is available.