# Flag the load-bearing premises

If some­one says, “I think your AI safety scheme is hor­ribly flawed be­cause X will go wrong,” and you re­ply “Nah, X will be fine be­cause of Y and Z”, then good prac­tice calls for high­light­ing that Y and Z are im­por­tant propo­si­tions. Y and Z may need to be de­bated in their own right, es­pe­cially if a lot of peo­ple con­sider Y and Z to be nonob­vi­ous or prob­a­bly false. Con­trast em­phe­meral premises. Needs to be paired with un­der­stand­ing of the Mul­ti­ple-Stage Fal­lacy so that list­ing load-bear­ing premises doesn’t make the propo­si­tion look less prob­a­ble - $$\neg X$$ will have load-bear­ing premises too.

Parents:

• AI safety mindset

Ask­ing how AI de­signs could go wrong, in­stead of imag­in­ing them go­ing right.

• This is a fair thing for more perfec­tion­is­tic re­searchers to ask from prag­ma­tists.

One thing that prag­ma­tists can fairly ask of perfec­tion­ists is “flag the ob­vi­ous, im­perfect solu­tion”. Often, if you don’t point out that some hacks are un­satis­fac­tory, peo­ple will think you haven’t no­ticed them at all, or will try to build new solu­tions from them, and be­sides, you could also be mis­taken about their satis­fac­tori­ness.