We make the rules. Economics is a game of our making and rules of our choosing.
Suppose the most efficient human produces $250k/y worth of labour and the least efficient model produces $1M/y worth of labour. So $1M - $250k = $750k. %1 of $750k = $7.5k. Now multiply that by the number of companies/people who bought these services.
Models are assumed to a) become more efficient and b) be competent on several and different types of jobs.
This would also include robots but those systems aren't yet ripe, IMO. I'd let them marinate a little further before taxing them.
Apply this tax to very large companies that produce/sell these systems.
[0] - This incentivizes companies to keep investing in more efficient and capable models/systems.
Too much gameable measurement on the tax end, and too much unnecessary bureaucracy both to support that on the tax end and the retraining end; simplify by just taxing capital accumulation more heavily than it currently is (which includes, but isn't limited to, capital accumulation due to implementing and reaping the returns of automation) and use the resulting revenue to fund a UBI; peopel for whom retraining has positive expected returns will have an incentive to do it, and those for whom that is not a good use of the money will make use of it otherwise.
You don't need to create an elaborate system to incentivize companies to keep investing in more efficient and capable systems, as long as you aren't taxing them at 100% or more, the fact that there are more returns to be had by having more efficient systems does that already.
Too much gameable measurement on the tax end, and too much unnecessary bureaucracy both to support that on the tax end and the retraining end;
An agent to take care of the bureaucracy, another to retrain someone for a new job, etc.
And, yes, the long-term plan would be to fund UBI at some point; because if these systems become so capable and so efficient might as well let them do everything for us while we humans enjoy leisure time.
.. this might and will lead to a time loop, somehow. In each scenario, AI can't be defeated without much losses. And, not when it fully evolved. We need to stop the development, now!