Stella Biderman

I think AI’s impact will depend less on treating models as autonomous forces and more on the institutions that build, study, deploy, and govern them. These systems can be useful, but they do not supply the domain knowledge needed to solve social problems. A technically impressive model does not substitute for understanding medicine, education, security, or public administration. The central question is who can examine and shape these systems. If meaningful research is restricted to a few large companies, outsiders cannot adequately investigate training dynamics, memorization, bias, security failures, or safety claims. Open models, checkpoints, public compute, transparent evaluations, and clear policy criteria make empirical scrutiny possible. Closed APIs and superficial filters can instead hide problems without resolving them. So I do not think our future is determined by an analogy to nuclear weapons or by concentrating control in a small number of incumbents. Those framings can mislead policymakers while doing little to prevent powerful models from being developed. A better future requires open competition, independent research, outside participation in policy, institutional accountability, and a science of how training actually produces capabilities and behavior.

I do not have a defensible single forecast that AI will be beneficial or harmful overall. Its impact is conditional on institutions: who controls access, whether independent researchers can inspect models and training dynamics, whether evaluations are transparent, and whether organizations are accountable for ordinary security and deployment failures. AI can support useful work, but it cannot replace domain expertise or repair weak institutions by itself. Harms are more likely to be obscured when scrutiny is limited to a few companies, while open models, public compute, checkpoints, and outside participation make problems easier to discover and contest. So I expect neither an automatic technological utopia nor an autonomous catastrophe. The balance will be shaped substantially by governance, research access, and deployment choices.

The most consequential evidence would be a robust empirical finding that independent access itself creates severe, unavoidable harms that cannot be mitigated without concentrating control. That would challenge my view that open models, checkpoints, public compute, and outside scrutiny generally improve accountability. Conversely, clear evidence that openness consistently exposes failures earlier, improves interventions, and supports useful competition would strengthen that view. I would also update substantially if research on training dynamics produced reliable ways to predict capabilities and safety-relevant behavior before deployment—or demonstrated that such prediction is fundamentally ineffective. The key would not be a dramatic analogy or isolated incident, but reproducible evidence about how models develop, how failures occur, and which institutions actually manage those failures well.