AI Alignment Risks Value Lock-In and Double Stagnation in Simulations

NTNenad Tomasev@weballergy12h ago · 76 Views

Joint work with @FranklinMatija and @sindero

NTNenad Tomasev@weballergy12h ago · 51 Views

1107TECH

Our analysis also highlights the potential for developing more adaptive personalisation and AI alignment solutions, to enable users to develop their own opinions and values, and find meaning without being overly restricted by the past, echoed via the AI influence.

NT

Nenad Tomasev@weballergy

Our models highlight the following systemic risks: Value lock-in: anchoring users to historical values Double stagnation: locked-in users decelerate institutional and societal progress Normative mode collapse: erosion of sub-cultural diversity, driving societies towards maladaptive global states

NTNenad Tomasev@weballergy12h ago · 61 Views

1107TECH

Our models highlight the following systemic risks: Value lock-in: anchoring users to historical values Double stagnation: locked-in users decelerate institutional and societal progress Normative mode collapse: erosion of sub-cultural diversity, driving societies towards maladaptive global states

NT

Nenad Tomasev@weballergy

We approach this problem by formulating a number of macro, population-level models of evolving norms and values in human populations paired with AI, under different initial assumptions and constraints. We then consider the consequences of these assumptions analytically and in simulations, effectively adopting a ‘social physics’ approach to early-stage predictive modeling. This is meant as a preliminary analysis, for rapid preliminary insights that may inform more involved agentic simulations.

NTNenad Tomasev@weballergy12h ago · 70 Views

The models we present are not meant to be in any way definitive - they are merely an informative illustration for how we should be thinking about these questions moving forward. Hopefully we can all do much more in this general space. The rapid pace of progress requires us to rapidly envision and adjust sociotechnical trajectories, and come up with appropriate policies that would lead to positive outcomes.

NTNenad Tomasev@weballergy12h ago · 21 Views

Finally, we hope to make a broader argument for adopting rapid formal experimentation as foresight in computational social science. The cost of doing so has dropped significantly nowadays with the availability of AI coding tools - enabling experts to quickly move beyond the qualitative considerations and make concrete, testable predictions.