asendatavus

Asendatavus is a theoretical construct used in discussions of data persistence and pattern recurrence in computational systems. The term denotes a tendency for historical patterns, biases, or information to reappear in outputs after changes to data, models, or training processes, even when explicit efforts are made to remove them. It is used to frame questions about how legacy data, model priors, and feedback loops influence current behavior.

Etymology for the term is not standardized; it is a coined word intended to evoke ancestry (atavus) with connotations of return or ascent, reflecting the idea that past states can re-emerge in modern systems.

In practice, asendatavus is invoked to describe scenarios such as the recurrences of old biases after fine-tuning, leakage of training data into evaluation, or recurrent drift where updated models converge toward older states. Mechanisms proposed to underlie asendatavus include data leakage between iterations, overlapping or reused training sets across model updates, selection bias in data curation, convergent optimization pressures, and recursive self-improvement loops.

Applications and debates around the concept are primarily in the realm of thought experiments and theoretical discussions. Proponents use asendatavus to highlight risks in continual learning, data governance, and AI safety, while critics argue that the term is imprecise without formal definitions or measurable criteria and may conflate multiple distinct phenomena under a single label.

See also: concept drift, data drift, model drift, dataset decay, feedback loop.