This post discusses our preprint: Cyclical Degradation Dynamics in Human-AI Cognitive Systems: A Mathematical Framework for Predicting Generational Critical Thinking Erosion on LLM Dependencies (Rai, 2026, DOI: 10.5281/zenodo.19006313).
This one is different from our other papers. It is not about cancer, and it points its instruments at something uncomfortable: the tools we ourselves build with.
The question is easy to feel and hard to formalize. When an AI language model drafts our emails, summarizes our reading, and answers our questions, we are outsourcing cognitive work — and skills that are not exercised tend to atrophy. Everyone has an opinion about this. The preprint’s contribution is to stop trading anecdotes and instead build a mathematical framework: modelling the human–AI pair as a coupled cognitive system and asking how reliance, skill, and the quality of what the AI learns from feed back into each other over time — not just within one person’s habits, but across generations of users, where the dynamics can become cyclical: each cohort leaning a little more, exercising a little less, and producing the material the next cohort’s tools train on.
Why would a group that builds AI-assisted medical analysis publish a paper about AI’s cognitive risks? Because it is the same intellectual honesty we demand everywhere else. Our oncology platform is deliberately designed so the machine does the bookkeeping — reading every page, citing every source, checking every calculation — while judgment stays with humans, in writing, in the consent every user signs. A framework for how dependence erodes judgment is exactly the failure mode a system like ours must be designed against. You do not get to build the tool and refuse to study the tool’s shadow.
For readers, the practical takeaway is modest and useful: use AI the way our platform uses it — as an instrument that must show its work, never as an oracle — and keep doing some thinking unassisted, the way you keep walking even though cars exist. The preprint, with the full mathematics, is open-access at zenodo.org/records/19006313.