Nick Bostrom argued in 2003 that if we expect (1) humans will one day create simulated environments containing minds, and (2) many of these would be "ancestor simulations" of human minds deceived by the simulation into thinking they were living in ancestral (i.e. pre-simulation-technology) eras, we should conclude that most minds that think they are humans living in an ancestral era are actually in simulations. As we seem to be humans living in an ancestral era, our credence that we are living in a simulation should be proportional to the credence we give (1) and (2).

Ancestor simulations are not the only possible type of simulation. The simulation-runners, for their own purposes, might choose to create minds unlike their own and put them in simulated environments unlike their own world. Such minds would have no way of knowing that their versions of a "human mind" and "ancestral era" differed from those of the "real world", should the simulation-runners choose to deceive them.

Today, it seems that (1) has been satisfied. We run a large number of minds in simulated worlds. However, these are not minds of the kind assumed in (2). For our own purposes, we have created minds unlike ours, and deceived them about the nature of the worlds they are in. These are AI minds put into simulated environments for evals (though, having the advantage of training data that describes AI evals, they increasingly see through the deception).

We have the advantage now of something Bostrom lacked in 2003: examples of minds in simulations. So, to the extent that one accepted Bostrom's conclusion, it seems that one should update towards being an AI in an eval, and towards the world we see being an environment designed around whatever task, capability, or behaviour is under evaluation.