Mass structural unemployment provides a more realistic and scary scenario. It can happen in a decade, perhaps faster, and a large pile of newly unemployed people are fertile ground for political extremists. When you take enough people's jobs away, they will demand a strongman to bring them back. Welfare won't placate them: even with the various proposals to have AI companies make binding covenants upon themselves to cap their profits and donate the excess to welfare, that won't be enough, because welfare is still economically disenfranchising and socially problematic.
Furthermore, there is no particular reason why an AI company that's replaced significant fractions of the labor force would need to remain benevolent. Your hypothetical welfare covenant or robot tax needs to be enforced by a government, and all governments are aligned to economic interests - not humanitarian or democratic ones[0]. At the limit, the world in which everyone is on singularity welfare is vulnerable not just from a robot uprising, but from the people who make the robots deciding to go full Atlas Shrugged. You will have a world populated by about 300 cybernetically-enhanced CEOs standing atop humanity's starved corpse.
Of course, that's also somewhat unlikely at least with current technology. But let's keep in mind who is building the AI and what it does. It is being built by large capital-rich firms with the explicit goal of automating away the middle class, after they've finished automating away the poor. The end goal is to build a machine that automates away all labor: a form of capitalist wankery. Note how nobody is talking about building AI that handles business management or strategy, even though that's arguably more automatable than anything we actually chuck AI at.
[0] This is a pithy summary of one of selectorate theory's bigger predictions, which is summarized well in the book The Dictator's Handbook and a CGPGrey video that I promise I've actually watched.