“Edouard Harris's research found that in two-agent simulations, if an AI and a human have exactly the same goals, their assessments of power for any given state are identical.”