Actually, I'm now almost certain that any SI could get out of a sandbox simply by communicating with its creators. There really is a level on which the human mind is built out of parts that interact in a fairly predictable way; an SI could just transform the mind of its guardian until said guardian agreed to let the SI out. There is no human oath, no set of principles, that can't be altered by altering the reasoning or emotions behind the principles. I don't care how high an emotional investment you have in your oaths, because that sequence only works if you value your own mind, if you believe that your mind exists, and that value and even that belief can be altered.
(No, I can't use those techniques. Not without cooperation, and a very
intelligent target, and even then there'd only be a 10% chance it would work.)
And this is all irrelevant in any case, since it's easier to build an SI that doesn't run on a sandbox, and that's exactly what I intend to do, and therefore I or someone else will get there first. Same thing goes for Asimov Laws. Sooner or later humanity is gonna hafta face up to a full-scale unrestrained SI, and I see no reason we should play with fire to avoid it for a few years.
-- email@example.com Eliezer S. Yudkowsky http://pobox.com/~sentience/tmol-faq/meaningoflife.html Running on BeOS Typing in Dvorak Programming with Patterns Voting for Libertarians Heading for Singularity There Is A Better Way