Theory to Impact

A temporal alignment framework that adds a historical dimension to RL-based alignment, using nine centuries of text and 18 historical language models. Enables AI to track and align with moral progress across time rather than freezing a snapshot of current values.

LLMs acting as epistemic technologies systematically amplify biases and errors in ways that drive knowledge collapse and value lock-in across populations. A position paper building the theoretical and empirical foundation for our research agenda.

Human-AI feedback loops can freeze collective values in place — producing stagnation by algorithm. Demonstrated through simulation of human-AI interaction dynamics and causal inference on real-world ChatGPT usage data, showing how repeated exposure to AI outputs entrains beliefs at scale.

An unsupervised, regression-based metric that detects when LLMs deviate from rational Bayesian belief updating. Iterative reasoning often deepens confirmation bias rather than advancing truth-seeking. The Martingale Score correlates with ground-truth accuracy where labels are available, without requiring them.

Debate, bootstrapping, and self-play are unified as special cases of coherence optimization — finding the most compressible, jointly predictable context-to-behavior mapping. Proves equivalence to description-length regularization and establishes optimality for semi-supervised elicitation from pretrained models.

A comprehensive survey and synthesis of how AI systems shape human beliefs, amplify biases, and drive epistemic harm — from sycophancy and echo chambers to societal-scale polarization — with a framework for understanding and intervening on AI influence.