"First, we collected an initial sample of human data (Acquire) and used it to train ‘virtual human players’ which were recurrent neural networks that learned to imitate human behaviour during the game and voted according to the same principles as human players"
The entire test was virtual including the "human-centred" component. They also mention it may have the same limitations as any democratic system in the real world of tyranny of the majority.
Bias and alignment related to AI will likely be limited to the same methods we have always had for society. It will potentially only be more effective at enforcing a particular type of existing flawed system - https://dakara.substack.com/p/ai-the-bias-paradox
The entire test was virtual including the "human-centred" component. They also mention it may have the same limitations as any democratic system in the real world of tyranny of the majority.
Bias and alignment related to AI will likely be limited to the same methods we have always had for society. It will potentially only be more effective at enforcing a particular type of existing flawed system - https://dakara.substack.com/p/ai-the-bias-paradox