An introduction to the research
Computational Meaning Dynamics (CMD) is the study of how meaning moves through AI systems — how semantic content shifts, drifts, and transforms as it is processed, transmitted, and re-processed by language models and other computational components.
The central insight is that AI systems are not neutral conduits for meaning. They alter it. And those alterations are not random — they are structured, measurable, and in many cases predictable. CMD treats semantic instability not as a nuisance to be minimized, but as a phenomenon to be understood and, where appropriate, deliberately exploited.
When a language model processes text, it operates on learned statistical representations of meaning — embedding vectors in high-dimensional space. These representations are approximate. They blur distinctions, collapse nuance, and introduce systematic biases that reflect both the training data and the architecture of the model itself.
Under recursive processing — when AI output is fed back in as input, as in multi-agent systems, retrieval-augmented generation, or iterative summarization — these approximations compound. Meaning drifts from its origin. The further downstream you go, the less the output reflects the input’s original semantic content.
This is not a failure of any particular system. It is a structural property of how neural language models represent and transform meaning.
The CMD Research Group applies tools from differential geometry to analyze semantic spaces. Specifically, we use Ollivier-Ricci curvature — a measure borrowed from the study of networks and manifolds — to characterize the local topology of embedding spaces.
Curvature tells us something important: regions of high curvature in semantic space are regions where small perturbations produce large meaning changes. These are the unstable zones. Knowing where they are, and how they form, gives us a principled basis for predicting where drift will occur and how severe it will be.
Diagnostic codes processed by clinical AI systems can drift in meaning, affecting patient outcomes and billing integrity.
Semantic instability is an attack surface. Understanding it enables both more robust defenses and more targeted adversarial techniques.
Multi-agent and retrieval-augmented systems that iterate over language are particularly susceptible to accumulated drift.
Computational — we work with formal, mathematical, and computational methods.
Meaning — our subject matter is semantics: what words, codes, and representations refer to.
Dynamics — we study change over time and across system states, not static snapshots.