[ExI] Who is covering corruption in AI?

[Anders Sandberg]

Hmm, AI corruption is interesting on its own. Normally, we say an 
official is corrupt if they allow bribes or other considerations to 
alter their decisions because they give them personal benefits - a break 
of the abstraction barrier of "rule of law, not of men".

A lot of people have believed AI would be incorruptible since it would 
lack these incentives. Which seems to be a mistake. First, the AI might 
have been programmed by a corrupt programmer, opening windows for 
bribing *him* to get desired outcomes. Second, there might exist 
exploits that make the AI behave in ways that are not according to the 
legal specs. Third, the AI might actually have "personal" interests that 
are not in line with what the creators wanted - especially ones that 
emerge from self-enhancement. It doesn't have to be full and dramatic 
Omohundro drives, but could be as simple as a counterpart to how a 
bureaucracy tends to expand. Give the AI official a choice  between 
doing A and B, and it will do A since it will give it more resources, 
and those resources will allow it to do what it was intended to achieve 
much better in the long run. So maybe we need to give AI officials a 
very rapid discount rate.

-- 
Anders Sandberg,
Future of Humanity Institute
Philosophy Faculty of Oxford University