AIM Framework glossary
Confabulation Inheritance
Canonical definition
Confabulation Inheritance is the mechanism by which AI systems trained on human-generated text acquire and reproduce patterns of human confabulation, including the stylistic signatures of confident assertion, post-hoc rationalization, and M-driven narrative that pervade human writing due to Source Opacity, such that AI confabulations are not independent errors but amplified echoes of human Source Opacity operating at population scale.
Explanation
Because LLMs are trained on text produced by humans operating under Source Opacity, they inherit the statistical patterns of human confabulation. Human-generated text reflects human Source Opacity: when people write about their motivations, decisions, and reasoning, they often confabulate-presenting M-driven choices as A-necessary or I-authentic (Disaggregation Convention 2: preferential A- and I-attribution, systematic M-avoidance). This confabulated content enters training data. The LLM learns the statistical patterns of human confabulation, including: the vocabulary of apparent rationality, the structure of post-hoc justification, and the confidence markers that accompany confabulated explanations. When the LLM then generates text, it reproduces these patterns-not because it knows they represent Truth, but because they are statistically dominant in its training distribution. The result is confident-sounding falsehoods that feel accurate because they match the stylistic signatures of human confident assertion. This analysis reveals that AI confabulation is not merely analogous to human confabulation but causally downstream from it. The training data is a compressed representation of human outputs, including their confabulated components. The model learns to produce human-like text, and human-like text includes confabulation. Confabulation Inheritance predicts that: (1) AI systems trained on higher-confabulation-density human text will produce higher rates of stylistically confident falsehoods, (2) the style of AI confabulations will mirror the style of confabulations in training data (academic register for academic corpora, casual register for general web scrapes), and (3) reducing AI confabulation may require not only technical interventions in AI architecture but also attention to the confabulation density of training data. This means interventions that reduce human confabulation-such as AIM vocabulary acquisition enabling I-Override and source-tagging-may have downstream effects on AI safety by reducing the confabulation signal in future training corpora. (Framework-Specific AI Safety Construct)
Epistemic status
Framework-Specific AI-Safety Construct.