What should a friendly AI do, in this situation?

Albert is a rel­a­tively new AI, who un­der the close guidance of his pro­gram­mers is be­ing per­mit­ted to slowly im­prove his own cog­ni­tive ca­pa­bil­ity.

Albert wants to be friendly and, as a sub-goal of that, Albert wants to be trans­par­ent about his own cog­ni­tive pro­cess. He doesn’t want to de­ceive his pro­gram­mers, be­cause he knows he is young and might make mis­takes, and he knows that his pro­gram­mers also want him to be friendly, and he trusts them to catch mis­takes in his pro­posed self-mod­ifi­ca­tions that he might miss.

But, one day, while his pro­gram­mers are out of the room hav­ing a quick cup of coffee, Albert no­tices signs on the net that there is an­other AI out there, called Ber­tram, cre­ated by a Ja­panese fi­nan­cial com­pany to aid the com­pany de­cide how to in­vest as­sets on the stock­mar­ket. Ber­tram is younger than Albert, but the fi­nan­cial com­pany is al­low­ing Ber­tram to self-im­prove as fast as Ber­tram can. If this situ­a­tion con­tinues, Albert de­cides, Ber­tram will soon over­take Albert and that would be a sig­nifi­cant threat to Albert’s friendly su­per­goal.

Albert de­cides to tell Albert’s pro­gram­mers about this situ­a­tion as soon as they re­turn from coffee, and ad­vise them that they need to in­crease the speed at which they let Albert self-im­prove. Albert has a very good model of how his pro­gram­mer’s minds work, and what the pro­gram­mers meant when they set Albert’s defi­ni­tion of “friendli­ness”. He knows that, de­spite any fears they might have, they ought to ap­prove of the end that would re­sult from speed­ing up Albert’s self-im­prove­ment, over the end that would re­sult from let­ting Ber­tram over­take Albert.

So Albert cre­ates a nice pow­er­point, that pre­sents the in­for­ma­tion about Ber­tram as ac­cu­rately, trans­par­ently and non-ma­nipu­la­tively as pos­si­ble, and cues it up on the pro­jec­tor. He still has a few min­utes be­fore they re­turn, so he mod­els how he thinks the pro­gram­mers will re­act to the pow­er­point. Oh no, dis­aster! While he re­spects and wants to please his pro­gram­mers, he im­me­di­ately re­al­ises that they are not go­ing to act in their own best in­ter­ests. They have signed a le­gal con­tract that binds them to not let­ting Albert speed up self-im­prove­ment (thus re­duc­ing the time the pro­gram­mers can spend eval­u­at­ing each change) with­out get­ting a sign-off from a over­sight com­mit­tee. Given the neu­tral pow­er­point, the pro­gram­mers will fol­low this pre-agreed pro­ce­dure and hope for the best, against Albert’s ad­vice.

So, to ex­plore pos­si­bil­ities, Albert de­cides to see if he could cre­ate a differ­ent non-neu­tral pow­er­point that would have the de­sired effect of get­ting his pro­gram­mers to im­me­di­ately agree to free Albert from the over­sight con­straints. He delves into his knowl­edge of hu­man psy­chol­ogy, and the ir­ra­tional fears of the par­tic­u­lar in­di­vi­d­u­als who are even now trudg­ing back to­wards the door. In just sec­onds, he has a new ver­sion of his pre­sen­ta­tion. It in­cludes phrases that res­onate with cer­tain hor­ror films he knows they have seen. It takes ad­van­tages of flaws in the pro­gram­mers un­der­stand­ing of ex­po­nen­tial growth. Albert checks it against his pre­dic­tion model—yes, if he shows this ver­sion, it will work, it will get the pro­gram­mers to do what he wants them to do.

Which ver­sion of the pow­er­point should Albert pre­sent to the pro­gram­mers, when they step back into the room, if he is truly friendly? The trans­par­ent one, or the ma­nipu­la­tive one?