Dispositional Seeds and Feedback Loops: A Yogācāra Account of Bias Accumulation and Hallucination in Machine Learning Systems

Algorithmic bias and large language model (LLM) hallucination are widely recognised as pressing challenges in AI ethics, yet existing explanations share three diagnostic gaps: they have not developed a model of cyclical causal accumulation, a non-individualist ontology of bias attribution, or an ontological account of hallucination. This paper argues that the Yogācāra Buddhist causal-dispositional model—centred on bīja (seed), vāsanā (perfuming), and pravṛtti (manifestation)—addresses all three gaps. Drawing on the Cheng weishi lun, the paper reconstructs the bīja-vāsanā cycle as a recursive architecture of disposition, activation, and re-disposition, and establishes its relationship to the training–weights–inference–feedback cycle of machine learning as a systematic structural analogy in the sense defined by Gentner’s [1] structure-mapping theory. Algorithmic bias is re-analysed as a recursive dispositional pathology; LLM hallucination is interpreted through the doctrine of three natures (trisvabhāva) as a systematic confusion of ontological levels. The analysis is placed in dialogue with contemporary disposition theory [17] and with Floridi’s certainty–scope conjecture [18]. By repositioning bias and hallucination as dispositional pathologies rooted in the recursive causal architecture of training systems—rather than as localised defects amenable to surface-level patching—the paper argues for reorienting ethical attention from behavioural surface-correction to ontological engagement with the dispositional substrate itself.