Is decentralized AI the civic counterweight we need before centralized AI becomes uncontestable?

I’ve been thinking about AI governance through the lens of centralized power versus parallel democratic systems. The comparison that keeps coming to mind is Hong Kong. Hong Kong existed as a semi-autonomous, more liberal-democratic system nested inside China’s broader centralized state structure. When that parallel model became politically threatening, Beijing moved to absorb and neutralize it. I am not saying AI governance maps perfectly onto Hong Kong. It obviously does not. But the pattern seems important: A centralized system may tolerate a parallel system only while that parallel system is not powerful enough to challenge legitimacy. That raises a serious question about AI. If the future is dominated only by centralized AI systems — whether controlled by states, corporations, or state-corporate partnerships — then those systems may eventually control not just tools, but interpretation itself: what is visible, credible, safe, legal, suspicious, employable, insurable, or true. That seems dangerous. My view is that we do not need decentralized AI instead of centralized AI. We need decentralized AI running in parallel with centralized AI. Centralized AI may be necessary for scale, infrastructure, national security, medicine, logistics, and critical systems. But decentralized AI may be necessary for audit, transparency, contestability, civic resilience, independent verification, local autonomy, and anti-capture pressure. The danger is not intelligence. The danger is uncontestable intelligence. So my question is: Should decentralized, democratized AI be treated as one of the most urgent public-interest infrastructure projects of the next decade? And if so, what would a serious version look like that avoids both extremes: centralized control on one side, and unsafe open chaos on the other?