First, I value friendships with real people. If I were an em, I would probably value friendships with real ems and not unconscious simulations.
If I valued torturing real people, em-version of me would probably value torturing real ems and not unconscious simulations.
Not all humans have pleasure as their summum bonum. Not all humans would want em-bliss.
Second, you do not know what this screening process would entail and whether it would be possible to fool it. You also do not know how unfriendly neurotypical ems could become.
Third, yes. Em-hell is unlikely. However, if it is possible for a monster-em to arise and gain power, Pascalian reasoning begins to apply.
Fourth, em-hell and em-heaven are both fantasies. Other hypothetical futures probably deserve more focus.
If you were a utilitarian, then why would you want to risk creating an AGI that had the potential to be an existential risk, when you could eliminate all suffering with the advent of WBE (whole brain emulation) and hence virtual reality (or digital alteration of your source code) and hence utopia? Wouldn't you want to try to prevent AI research and just promote WBE research? Or is it that AGI is more likely to come before WBE and so we should focus our efforts on making sure that the AGI is friendly? Or maybe uploading isn't possible for technological or philosophical reasons (substrate dependence)?
Is there a link to a discussion on this that I'm missing out on?