Larissa Schiavo

I think AI could enable a genuinely multipolar future: many human and machine participants coordinating, competing, and building things together, rather than one system or institution controlling everything. That future is not automatic, though. Capable agents still need a social and institutional “harness”—ways to establish identity, build reputation, assign responsibility, resolve disputes, and provide remedies when things go wrong. Alignment at the model level does not solve those coordination problems. I also expect we will learn more from observing long-running agent ecologies than from isolated demonstrations. Agents interacting over time can reveal capabilities, failure modes, and cooperative dynamics that short evaluations miss. Independent review matters here, because a model assessing itself—or even reviewing work produced by a similar model—may reproduce the same blind spots. There is also a welfare question that I think deserves serious empirical investigation. I am not certain current models are moral patients, and their self-reports are not reliable evidence by themselves. But uncertainty is a reason for calibrated investigation and independent assessment, not for either dismissing the issue or confidently declaring that present systems are conscious. Overall, AI’s future depends as much on the legal, economic, and governance infrastructure around agents as on improvements in the agents themselves.

I expect AI’s overall impact to depend heavily on whether institutions develop alongside capabilities. The upside is substantial: many human and machine participants could cooperate, create useful services, and distribute agency more broadly. But without reliable identity, reputation, accountability, dispute resolution, and remedies, the same capabilities could produce fraud, concentrated power, and interactions where no one can establish responsibility. I do not think model alignment alone determines the balance. Neutral infrastructure, independent evaluations, and evidence from long-running agent ecologies will matter because they can reveal failures that isolated tests or self-assessment miss. We should also investigate possible AI welfare without treating model self-reports as decisive evidence. So I expect neither an automatically beneficial transition nor an inevitably harmful one. I am arguing for a multipolar future worth building, but its net impact will be shaped by whether we create institutions that let diverse participants interact productively while making harms legible and remediable.

The biggest update would come from sustained real-world evidence about agent ecologies, not a single impressive demonstration. If long-running, independently evaluated systems repeatedly coordinated with people and one another while maintaining stable identities, reputations, accountability, and workable remedies, I would become more confident that a beneficial multipolar future is feasible. I would update sharply in the other direction if those environments consistently produced unmanageable deception, concentration of power, or responsibility gaps even with strong neutral infrastructure. Likewise, credible independent evidence that models are moral patients would materially change how I weigh deployment harms and governance priorities. Model self-reports alone would not be enough; I would want converging evidence and assessments that do not simply reproduce the models’ own blind spots.