Writing toward more rational AI.

Abstract

Observation is essential to the production of knowledge, yet it is also limited by perception, instruments, and prior assumptions. This paper examines whether these limitations weaken knowledge, with reference to the natural sciences and history.

In the natural sciences, observational flaws are often treated as errors that can be identified and reduced. Replication, falsification, peer review, and improved technology allow scientific claims to be tested over time. The 1919 eclipse observations used to support Einstein’s theory of relativity show that observation does not simply produce knowledge. It also tests claims that already exist. The thalidomide case, however, shows that limited observation can produce harmful conclusions before those limits are recognised.

In history, observation works differently. Historians cannot observe past events directly. They rely on testimony, records, and surviving artefacts. These sources are incomplete and shaped by perspective. Historical knowledge therefore depends on interpretation. Colonial accounts of Indigenous peoples show how one perspective can dominate the record. The Vinland Map also shows how later methods can expose earlier errors, though they cannot remove the basic limits of historical evidence.

This paper argues that flawed observation does matter, but not in the same way across different fields. In the natural sciences, reliability depends on systems of testing and correction. In history, it depends on critical interpretation and the comparison of sources. Knowledge does not require perfect observation. It requires methods that recognise and respond to its limits.

Abstract

Large language models are increasingly used in sustained conversations rather than isolated prompts. Existing research has examined failures such as hallucination, sycophancy, and factual error, but these phenomena are typically analyzed at the level of individual outputs. Less attention has been given to how epistemic degradation develops over the course of interaction and how earlier outputs may reshape the conditions under which later knowledge claims are produced.

We argue that the epistemic risks of sustained interaction are best understood as a dynamic process. To capture this process, we introduce the Epistemic Conversational State (ECS) as a representation of the evolving epistemic organization of a conversation. We distinguish four idealized states—Stable, Uncertain, Contaminated, and Self-Reinforcing—and describe their transitions through AI Epistemic Dynamics. We then propose Recursive Epistemic Contamination (REC) as the mechanism through which low-quality claims (LQCs) become recursively reinforced by altering the epistemic conditions of later generation. At the conversational level, this process is realized through the Context–Contamination Reinforcement Loop (Micro-REC). At the level of the wider knowledge environment, we extend the analysis to Macro-REC and the Community–Corpus Contamination Loop (Macro-REC), showing how local contamination may circulate across communities, public corpora, retrieval systems, and successive model generations.

The framework recasts AI reliability as a dynamic epistemic property rather than a static measure of model performance. More broadly, it provides a unified conceptual account of how recursive epistemic contamination can develop within conversations and propagate across knowledge ecosystems, offering a foundation for future empirical research on epistemic stability, recursive degradation, and the long-term reliability of artificial cognitive systems.

Abstract

We argue that rational AI is not AI that removes uncertainty. It is AI that can handle uncertainty responsibly. This claim begins from a simple fact about human knowledge. Science and history do not work because they reach perfect certainty. They work because they expose doubt, test claims, give reasons, and correct mistakes. Error is not the enemy of reason. Unmanaged error is.

This point changes how we should understand large language models. The main problem is not simply that they hallucinate. The deeper problem is that their claims often sound authoritative even when their grounds are weak. These claims can enter classrooms, research, policy, search systems, and training data as if they were knowledge. Yet they often lack the responsibility structures that make knowledge claims trustworthy.

We develop this view through the H-W-R+G model. At the level of the individual claim, Humility requires AI to show its limits. Warrant requires it to give assessable grounds. Responsiveness requires it to correct errors and learn from counterevidence. Governance adds a wider level. Once AI-generated claims circulate through institutions, platforms, and data environments, responsibility cannot remain inside a single answer. We need systems that prevent weak claims from being copied, cited, and fed back into future models.

We also distinguish two forms of self-reinforcement: shifts in the Latent Epistemic Context State within dialogue, and Recursive Epistemic Contamination across the knowledge ecosystem. Rational AI, we conclude, should be understood not as an oracle, but as a corrigible cognitive partner.