Hear me out, so what we do is this:

• Humans are naturally fallible and even smart ones are prone to snap judgment on partial or biased information or emotional entanglement
• So let's create a decisionmaking AI - we can teach it how to govern fairly and it will make the best decisions
• But wait, any AI we create is naturally going to be programmed with our limited point of view in mind and may end up making a weird choice due to a programming flaw or edge case it wasn't trained to handle
• So what we do is create two more AIs, with slightly different parameterizations and randomized training scenarios
• The three AIs will act as checks and balances on one another
• We can even try to embody different aspects of how humans approach problems and make decisions, using something like Jungian archetypes to help choose among difficult tradeoffs
• Everything will be fine

spoiler

unless