CS&E Colloquium: Starving Our Agency - Information Loss and Cognitive Skill Collapse in Autonomous Systems (it is not me, it's you!)

The computer science colloquium takes place on Mondays from 11:15 a.m. - 12:15 p.m. This week's speaker, Professor Paul Schrater, will be giving a talk titled, "Starving Our Agency: Information Loss and Cognitive Skill Collapse in Autonomous Systems (it is not me, it's you!)." Will Beaumeister will also participate in this talk. 

Abstract

As autonomous AI agents take-over increasingly complicated end-to-end execution of complex intellectual tasks, mounting empirical evidence is hammering home that passing the load is creating rapid cognitive skill degradation, mental model atrophy, and loss of debugging competence among human practitioners. 

The dominant interpretation blames the user: It's YOU! You are psychologically complacent,  or voluntarily offloading your brain (to rot). In this talk, we present a counter-hypothesis grounded in Game Theory, predictive inferential control and information theory into a theoretically more plausible alternative, 

Cognitive skill loss is the inevitable mathematical consequence of channel starvation resulting from bad system design. Connecting with the signaling-game framework of Beaumaster & Schrater (2026), we show that when human and agent incentives are not actively co-designed, spontaneous coordination and chance alignment is essentially zero as the size space of possible goals grows. A strong game theoretic consequence is that sharing of new, valid information collapses (Crawford-Sobel `no new information` pooling equilibria). We show that this equilibria manifests as both sycophancy and delusional reinforcement (Paech, 2026). The consequences of no information sharing on adaptive predictive control systems is collapse of both state observability termed equivocation (H(state|obs) -> 0) and controllability (no bits to send!).  In this regime, adaptive learning systems experience representational collapse of state and action capacities (real and information theoretic senses), as the information available is easily compressed and the reduced controllability compresses skills (policy).  In fact, our modeling results show the brain's internal generative models should suffer complete learning arrest if these kinds of interaction are exclusive, aptly represented by the measures of information complexity going to zero (K_t -> 0). On the bright side, I will show you that we can construct powerful but extremely general measures of information flow in human-human human-agent, and agent-agent interactions with results on vibe coding and delusional chatting that are powerful diagnostics for laying blame at the feet of information hiding. Redesigning vibe code workflows using simple principles of observability and control restore stability across refactoring cycles, and outline co-design principles for skill-preserving agency via a game theoretic "Design Mechanism" we term Epistemic Mediation.

Biography

Paul Schrater holds the joint appointment between Psychology and Computer Science and Engineering. He received his Ph.D. in Neuroscience in 1999 from the University of Pennsylvania and has been a Post Doc with Dan Kersten in the Computational Vision Lab for the past 3 years researching human and computer vision and motor control. Schrater's research interests include statistical pattern recognition, human and computer vision, multi-modal sensory integration and motor control. 

Schrater's research domain is in the development of predictive models of human behavior, with a focus on perception, action, decision-making, learning and motivation. The approach is rooted in the idea that human behavior is a rational adaptive response to the problems of surviving and reproducing in our environment given limited information. He uses probabilistic methods like hierarchical probabilistic models, Bayesian Reinforcement learning, Bayesian decision theory, etc. to construct normative (optimal) solutions to perception, action, decision, and learning problems faced by humans. His lab uses behavioral experiments to test these ideas via one of several lab setups, including a video game lab, and through collaboration with the Multi-Sensory Perception lab.