Architectures of Error
“An "Architecture of Error" will be defined as the system of causal mechanisms, inherent to a generative architecture (be it biological-cognitive or artificial-stochastic), that gives rise to its characteristic failure modes.”
— Architectures of Error (2026), §1
Why it matters
Architectures of Error grounds the AF programme: it shifts evaluation of AI systems away from output similarity toward the causal structures that produce the output, so that surface-level functional equivalence cannot be mistaken for ontological equivalence.
Notes
The framework is introduced as a deliberate counter to the replacement narrative that treats AI-generated and human-generated code as interchangeable on the basis of behavioural indistinguishability. The paper argues that this equivalence dissolves once the analysis is reframed in terms of causal origin: human errors emerge from cognitive processes that include comprehension; AI errors emerge from stochastic-statistical processes that explicitly do not.
The framework is operationalised through Floridi’s Levels of Abstraction, allowing the two architectures to be compared along multiple dimensions—semantic coherence, security robustness, epistemic limits—without collapsing them into a single “performance” metric.